While 3DGS has emerged as a high-fidelity scene representation, encoding rich, general-purpose features directly from its primitives remains under-explored. We address this gap by introducing Chorus, a multi-teacher pretraining framework that learns a holistic feed-forward 3D Gaussian Splatting (3DGS) scene encoder by distilling complementary signals from 2D foundation models. Chorus employs a shared 3D encoder and teacher-specific projectors to learn from language-aligned, generalist, and object-aware teachers, encouraging a shared embedding space that captures signals from high-level semantics to fine-grained structure.We evaluate Chorus on a wide range of tasks: open-vocabulary semantic and instance segmentation, linear and decoder probing, as well as data-efficient supervision. Besides 3DGS, we also test Chorus on several benchmarks that only support point clouds by pretraining a variant using only Gaussians’ centers, colors, estimated normals as inputs. Interestingly, this encoder shows strong transfer and outperforms the point clouds baseline while using $39.9\times$ fewer training scenes. Finally, we propose a render-and-distill adaptation that facilitates out-of-domain finetuning. Our code and model will be released upon publication.

One of the most exciting applications of vision models involve pixel-level reasoning.Despite the abundance of vision foundation models, we still lack representations that effectively embed spatio-temporal properties of visual scenes at the pixel level.Existing frameworks either train on image-based pretext tasks, which do not account for dynamic elements, or on video sequences for action-level reasoning, which does not scale to dense pixel-level prediction.We present a framework that learns pixel-accurate feature descriptors from videos, LILA.The core element of our training framework is linear in-context learning.LILA leverages spatio-temporal cue maps -- depth and motion -- estimated with off-the-shelf networks.Despite the noisy nature of those cues, LILA trains effectively on uncurated video datasets, embedding semantic and geometric properties in a temporally consistent manner.We demonstrate compelling empirical benefits of the learned representation across a diverse suite of vision tasks: video object segmentation, surface normal estimation and semantic segmentation.

Keypoint-based matching is a fundamental component of modern 3D vision systems, such as Structure-from-Motion (SfM) and SLAM. Most existing learning-based methods are trained on image pairs, a paradigm that fails to explicitly optimize for the long-term trackability of keypoints across sequences under challenging viewpoint and illumination changes. In this paper, we reframe keypoint detection as a sequential decision-making problem. We introduce TraqPoint, a novel, end-to-end Reinforcement Learning (RL) framework designed to optimize the Track-quality (Traq) of keypoints directly on image sequences. Our core innovation is a track-aware reward mechanism that jointly encourages the consistency and distinctiveness of keypoints across multiple views, guided by a policy gradient method. Extensive evaluations on sparse matching benchmarks, including relative pose estimation and 3D reconstruction, demonstrate that TraqPoint significantly outperforms some state-of-the-art keypoint detection and description methods.

We revisit the problem of estimating the fundamental matrix of a pair of perspective cameras, a cornerstone of geometric computer vision.As is well-known, linear solvers require at least 8 point correspondences, whereas nonlinear minimal solvers require just 7 in the uncalibrated case or 5 in the calibrated case.In this paper, we consider a special case of the 7-point problem where 5 of the points are configured to lie on two lines, which has previously been shown to have a unique solution.As a theoretical contribution, we offer an analysis of how this uniqueness manifests in the standard 7-point algorithm. On a practical level, we provide the first practical linear solver for the minimal problem associated to this special configuration.Additionally, we evaluate a heuristic 5-point fundamental matrix solver based on the construction of virtual midpoints.When combined with early non-minimal fitting, the runtime and accuracy of our solver is competitive with the state-of-the-art (SoTA) on multiple benchmarks.

Semantic Scene Completion (SSC) is crucial for 3D perception in mobile robotics, as it enables holistic scene understanding by jointly estimating dense volumetric occupancy and per-voxel semantics. Although SSC has been widely studied in terrestrial domains such as autonomous driving, aerial scenarios like autonomous flying remain largely unexplored, thereby limiting progress on downstream applications. Furthermore, LiDAR sensors represent the primary modality for SSC data generation, which poses challenges for most uncrewed aerial vehicles (UAVs) due to flight regulations, mass and energy constraints, and the sparsity of LiDAR-based point clouds from elevated viewpoints. To address these limitations, we introduce OccuFly, the first real-world, camera-based aerial SSC benchmark, captured at altitudes of 50m, 40m, and 30m during spring, summer, fall, and winter. OccuFly covers urban, industrial, and rural scenarios, provides 22 semantic classes, and the data format adheres to established conventions to facilitate seamless integration with existing research. Crucially, we propose a LiDAR-free data generation framework that is based on camera modality, which is ubiquitous on modern UAVs. By utilizing traditional 3D reconstruction, our framework automates label transfer by projecting annotated 2D masks into the reconstructed 3D point cloud, thereby minimizing manual 3D annotation effort. Finally, we benchmark several state-of-the-art SSC methods on OccuFly using standard metrics, and highlight challenges specific to aerial viewpoints, yielding a comprehensive aerial vision benchmark that fosters holistic aerial 3D scene understanding.

We present VGGT-Ω, a feed-forward model for 3D reconstruction that substantially advances the state of the art in accuracy, efficiency, and capability for both static and dynamic scenes. Prior models such as VGGT have shown that feed-forward 3D reconstruction can already be competitive with traditional optimization-based methods. Here, we further demonstrate that the accuracy and robustness of these models scale predictably with model capacity and data size. To enable training 3D reconstruction models at an unprecedented scale, we introduce a high-quality data annotation pipeline that handles dynamic scenes, a self-supervised learning protocol, and architectural changes that greatly reduce memory requirements. We significantly simplify VGGT’s architecture by replacing multiple dense prediction heads with loss-driven multitask learning, removing unstable DPT blocks, and introducing more efficient global attention via scene tokens. These changes allow us to efficiently train VGGT-Ω with 20$\times$ more supervised data and 100$\times$ more unsupervised data than prior work, while requiring only 30% of VGGT’s memory and running 1.6$\times$ faster at inference. As a result, VGGT-Ω establishes a new state of the art for 3D reconstruction on both static and dynamic scenes across a wide range of benchmarks, e.g., improving the camera estimation accuracy by 77% on the Sintel dataset. Models and code will be publicly released.

Agentic vision–language models are increasingly trained to “think with images” by calling image operations. However, we show that high final-answer accuracy often hides unfaithful visual reasoning: models may invoke tools on irrelevant regions or ignore tool outputs entirely, yet still guess the correct answer. In this work, we first propose a faithfulness evaluation protocol that measures whether intermediate visual tool outputs (e.g., crops) actually contain the queried evidence. This reveals that recent visual agents achieve high final-answer accuracy but exhibit low rates of faithful tool-use on visual search benchmarks. We then introduce CodeV, a code-based visual agent trained with Tool-Aware Policy Optimization (TAPO). TAPO is a process-level RL framework that augments GRPO with dense rewards defined directly on visual tool inputs and outputs, rather than on chain-of-thought tokens, making supervision easier to verify and less susceptible to reward hacking. CodeV represents visual tools as executable Python code, and TAPO assigns step-wise rewards based solely on the question and tool output, encouraging both necessary and evidence-consistent tool use. In a two-stage SFT+RL pipeline, CodeV achieves competitive or superior accuracy while substantially increasing faithful tool-use rates on related visual search benchmarks. Beyond visual search, CodeV attains strong performance on a range of multimodal reasoning and math benchmarks, suggesting that explicitly supervising intermediate tool behavior is crucial for building trustworthy, agentic visual reasoning systems.

We introduce NitroGen, a video-action foundation model for generalist gaming agents, trained on 40,000 hours of gameplay videos across more than 1000 games. We incorporate three key ingredients: 1) an internet-scale video-action dataset constructed by automatically extracting player actions from publicly available gameplay videos, 2) a multi-game benchmark environment that can measure cross-game generalization, and 3) a unified vision-action model trained with large-scale behavior cloning. NitroGen exhibits strong competence across diverse domains, including combat encounters in 3D action games, high-precision control in 2D platformers, and exploration in procedurally generated worlds. It transfers effectively to unseen games, achieving up to 52% relative improvement in success rates over models trained from scratch. We release the dataset, benchmark, and model weights to advance research on generalist embodied agents.

Physical AI aims to develop models that can perceive and predict real-world dynamics; yet, the extent to which current multi-modal large language models and video generative models support these abilities is insufficiently understood. We introduce Physical AI Bench (PAI-Bench), a unified and comprehensive benchmark that evaluates perception and prediction capabilities across video generation, conditional video generation, and video understanding, comprising 2,808 real-world cases with task-aligned metrics designed to capture physical plausibility and domain-specific reasoning. Our study provides a systematic assessment of recent models and shows that video generative models, despite strong visual fidelity, often struggle to maintain physically coherent dynamics, while multi-modal large language models exhibit limited performance in forecasting and causal interpretation. These observations suggest that current systems are still at an early stage in handling the perceptual and predictive demands of Physical AI. In summary, PAI-Bench establishes a realistic foundation for evaluating Physical AI and highlights key gaps that future systems must address.

Autonomous Vehicles (AVs) collect and pseudo-label terabytes of multi-modal data localized to HD maps during normal fleet testing. However, identifying interesting and safety-critical scenarios from uncurated driving logs remains a significant challenge. Traditional scenario mining techniques are error-prone and prohibitively time-consuming, often relying on hand-crafted structured queries. In this work, we revisit spatio-temporal scenario mining through the lens of recent vision-language models (VLMs) to detect whether a described scenario occurs in a driving log and, if so, precisely localize it in both time and space. To address this problem, we introduce RefAV, a large-scale dataset of $10,000$ diverse natural language queries that describe complex multi-agent interactions relevant to motion planning derived from $1000$ driving logs in the Argoverse 2 Sensor dataset. We evaluate several referential multi-object trackers and present an empirical analysis of our baselines. Notably, we find that naively repurposing off-the-shelf VLMs yields poor performance, suggesting that scenario mining presents unique challenges. Lastly, we discuss our recently held competition and share insights from the community.

Soccer understanding has recently garnered growing research interest due to its domain-specific complexity and unique challenges.However, prior works typically rely on task-specific expert models, which are resource-intensive and hinder a holistic view of the game.This paper aims to propose a unified framework that enables a single model to handle diverse soccer visual understanding tasks, spanning both fine-grained perception (e.g., athlete detection) and semantic reasoning (e.g., event classification).Concretely, we make the following contributions in this paper:(i) we present SoccerMaster, the first soccer-specific vision foundation model that unifies comprehensive understanding tasks within a single framework via supervised multi-task pretraining;(ii) we consolidate multiple existing soccer video datasets and develop an automated data curation pipeline, termed as SoccerFactory, to produce scalable multi-task training annotations;and (iii) we conduct extensive experiments demonstrating that SoccerMaster consistently outperforms task-specific expert models across diverse downstream tasks, underscoring its breadth and superiority.The data, code, and model will be publicly available to the research community.

Recent advancements in Vision Language Models (VLMs) have expanded their capabilities to interactive agent tasks, yet existing benchmarks remain limited to single-agent or text-only environments. In contrast, real-world scenarios often involve multiple agents interacting within rich visual and textual contexts, posing challenges with both multimodal observations and strategic interactions. To bridge this gap, we introduce Visual Strategic Bench (VS-Bench), a multimodal benchmark that evaluates VLMs for strategic abilities in multi-agent environments. VS-Bench comprises ten vision-grounded environments that cover cooperative, competitive, and mixed-motive interactions. The performance of VLM agents is evaluated across three dimensions: perception measured by element recognition accuracy; strategic reasoning measured by next-action prediction accuracy; and decision-making measured by normalized episode return. Extensive experiments on fifteen leading VLMs show that, although current models exhibit strong perception abilities, there remains a significant gap to optimal performance in reasoning and decision-making, with the best-performing model attaining 46.6% prediction accuracy and 31.4% normalized return. We further analyze the key factors influencing performance, conduct human studies, and examine failure modes to provide a deeper understanding of VLMs' strategic abilities. By standardizing the evaluation and highlighting the limitations of existing models, we envision VS-Bench as a foundation for future research on strategic multimodal agents.

Conventional multi-projector calibration requires projecting and capturing structured light patterns for each projector sequentially, causing calibration time and effort to increase linearly with the number of projectors. This scalability bottleneck has long limited the deployment of large-scale projection mapping systems. We present a new calibration framework that breaks this limitation by embedding cameras into the surface of the calibration target. The embedded cameras directly capture the incoming projection light, enabling the separation of simultaneously projected structured light patterns from multiple projectors according to their incident directions. Our method establishes correspondences between the optical centers of the embedded cameras and the projector pixels, allowing the intrinsic and extrinsic parameters of all projectors to be simultaneously estimated. We further introduce a correction technique for small misalignments between the calibration board and camera optical centers. As a result, our system achieves calibration accuracy comparable to conventional methods while reducing the required number of projection-capture cycles from linear to nearly constant with respect to the number of projectors, dramatically improving scalability for large multi-projector environments.

3D Gaussian Splatting (3DGS) has emerged as a prominent 3D representation for high-fidelity and real-time rendering. Prior work has coupled physics simulation with Gaussians, but predominantly targets soft, deformable materials, leaving brittle fracture largely unresolved. This stems from two key obstacles: the lack of volumetric interiors with coherent textures in GS representation, and the absence of fracture-aware simulation methods for Gaussians. To address these challenges, we introduce GaussianFluent, a unified framework for realistic simulation and rendering of dynamic object states. First, it synthesizes photorealistic interiors by densifying internal Gaussians guided by generative models. Second, it integrates an optimized Continuum Damage Material Point Method (CD-MPM) to enable brittle fracture simulation at remarkably high speed. Our approach handles complex scenarios including mixed-material objects and multi-stage fracture propagation, achieving results infeasible with previous methods. Experiments clearly demonstrate GaussianFluent's capability for photo-realistic, real-time rendering with structurally consistent interiors, highlighting its potential for downstream application, such as VR and Robotics.

Modern vision-language models (VLMs) are expected to have abilities of spatial reasoning with diverse scene complexities, but evaluating such abilities is difficult due to the lack of benchmarks that are not only diverse and scalable but also fully customizable. Existing benchmarks offer limited customizability over the scene complexity and are incapable of isolating and analyzing specific VLM failure modes under distinct spatial conditions. To address this gap, instead of individually presenting benchmarks for different scene complexities, in this paper we present InfiniBench, a fully automated, customizable and user-friendly benchmark generator that can synthesize a theoretically infinite variety of 3D scenes with parameterized control on scene complexity. InfiniBench uniquely translates scene descriptions in natural language into photo-realistic videos with complex and physically plausible 3D layouts. This is achieved through three key innovations: 1) a LLM-based agentic framework that iteratively refines procedural scene constraints from scene descriptions; 2) a flexible cluster-based layout optimizer that generates dense and cluttered scenes previously intractable for procedural methods; and 3) a task-aware camera trajectory optimization method that renders scenes into videos with full object coverage as VLM input. Experiments demonstrate that InfiniBench outperforms state-of-the-art procedural and LLM-based 3D generation methods in prompt fidelity and physical plausibility, especially in high-complexity scenarios. We further showcased the usefulness of InfiniBench, by generating benchmarks for representative spatial reasoning tasks including measurement, perspective-taking and spatiotemporal tracking.

Active mapping aims to determine how an agent should move to efficiently reconstruct an unknown environment. Most existing approaches rely on greedy next-best-view prediction, resulting in inefficient exploration and incomplete scene reconstruction.To address this limitation, we introduce MAGICIAN a novel long-term planning framework that maximizes accumulated surface coverage gain through Imagined Gaussians, a predicted scene representation derived from a pre-trained occupancy network with strong structural priors. This representation enables efficient computation of overage gain for any novel viewpoint via fast volumetric rendering.The resulting speedup allows the integration of the gain metric into a tree-search algorithm for planning long-horizon paths.We update Imagined Gaussians and refine the planned trajectory in a closed-loop manner.Our method achieves state-of-the-art performance across indoor and outdoor benchmarks with varying action spaces, demonstrating the critical advantage of long-term planning in active mapping.

Aerial object-goal navigation (Aerial ObjectNav) requires an Unmanned Aerial Vehicle (UAV) to navigate to target objects in large-scale outdoor environments using only visual observations and high-level object descriptions, without detailed step-by-step instructions. Existing approaches rely on local observations or short-term history, lacking comprehensive scene understanding and efficient spatial exploration strategies, which constrains their navigation capability in complex aerial scenarios. To address these challenges, we propose OctMem-Agent, an octree memory-augmented framework for aerial object-goal navigation. Specifically, we introduce an Adaptive Octree Memory that incrementally aggregates RGB-D observations into a hierarchical 3D representation, capturing both explored regions and unexplored frontiers across large-scale aerial environments. We further propose a Instruction-Guided Memory Query module that extracts task-relevant scene and exploration tokens through instruction-modulated queries. By integrating these tokens with visual observations and language instructions, OctoMem-Agent achieves comprehensive scene understanding and effective spatial exploration for target localization. Extensive experiments on the Aerial ObjectNav benchmark UAV-ON demonstrate that our method achieves a significant 7.5\% improvement in success rate over existing methods, validating the effectiveness of our design.

Open-vocabulary 3D occupancy is vital for embodied agents, which need to understand complex indoor environments where semantic categories are abundant and evolve beyond fixed taxonomies. While recent work has explored open-vocabulary occupancy in outdoor driving scenarios, such methods transfer poorly indoors, where geometry is denser, layouts are more intricate, and semantics are far more fine-grained. To address these challenges, we adopt a geometry-only supervision paradigm that uses only binary occupancy labels (occupied vs. free). Our framework builds upon 3D Language-Embedded Gaussians, which serve as a unified intermediate representation coupling fine-grained 3D geometry with a language-aligned semantic embedding. On the geometry side, we find that existing Gaussian-to-Occupancy operators fail to converge under such weak supervision, and we introduce an opacity-aware, Poisson-based approach that stabilizes volumetric aggregation. On the semantic side, direct alignment between rendered features and open-vocabulary segmentation features suffers from feature mixing; we therefore propose a Progressive Temperature Decay schedule that gradually sharpens opacities during splatting, strengthening Gaussian–language alignment. On Occ-ScanNet, our framework achieves 59.50 IoU and 21.05 mIoU in the open-vocabulary setting, surpassing all existing occupancy methods in IoU and outperforming prior open-vocabulary approaches by a large margin in mIoU. Code will be released.

In-context segmentation (ICS) aims to segment arbitrary concepts, objects, parts, or personalized instances given a few annotated visual examples. Existing work relies on (i) fine-tuning vision foundation models (VFMs), which improves in-domain results but limits generalization, or (ii) combines multiple frozen VFMs, which preserves generalization but yields architectural complexity and fixed segmentation granularities. We revisit ICS from a minimalist perspective and ask: Can a single self-supervised backbone support both semantic matching and segmentation, without any supervision or auxiliary models? We show that scaled-up dense self-supervised features from DINOv3 exhibit strong spatial structure and semantic correspondence. We introduce INSID3, a training-free approach that segments concept at varying granularities only from frozen DINOv3 features, given an in-context example. INSID3 achieves state-of-the-art results across one-shot semantic, part, and personalized segmentation, outperforming previous work by +6.1 % mIoU, while using 3x fewer parameters and without any mask or category-level supervision.

Recent advances in semantic correspondence rely on dual-encoder architectures, combining DINOv2 with diffusion backbones. While accurate, these billion-parameter models generalize poorly beyond training keypoints, revealing a gap between benchmark performance and real-world usability, where queried points rarely match those seen during training.Building upon DINOv2, we introduce MARCO, a unified model for generalizable correspondence driven by a novel training framework that enhances both fine-grained localization and semantic generalization. By coupling a coarse-to-fine objective that refines spatial precision with a self-distillation framework, which extends sparse supervision beyond annotated regions, our approach transforms a handful of keypoints into dense, semantically coherent correspondences.MARCO sets a new state of the art on SPair-71k, AP-10K and PF-PASCAL, with gains that amplify at fine-grained localization thresholds (+10.3 PCK@0.01), strongest generalization to unseen keypoints (+3.8, SPair-U) and categories (+5.6, MP-100), while remaining 3× smaller and 10× faster than diffusion-based approaches.

Visual Foundation Models (VFMs) such as the Segment Anything Model (SAM) have significantly advanced broad use of image segmentation. However, SAM and its variants necessitate substantial manual effort for prompt generation and additional training for specific applications. Recent approaches address these limitations by integrating SAM into in-context (one/few shot) segmentation, enabling auto-prompting through semantic alignment between query and support images. Despite these efforts, they still generate sub-optimal prompts that degrade segmentation quality due to visual inconsistencies between support and query images. To tackle this limitation, we introduce PR-MaGIC (Prompt Refinement via Mask Decoder Gradient Flow for In-Context Segmentation), a training-free test-time framework that refines prompts via gradient flow derived from SAM’s mask decoder. PR-MaGIC seamlessly integrates into in-context segmentation frameworks, being theoretically grounded yet practically stabilized through a simple top-1 selection strategy that ensures robust performance across samples.Extensive evaluations demonstrate that PR-MaGIC consistently improves segmentation quality across various benchmarks, effectively mitigating inadequate prompts without requiring additional training or architectural modifications.

Foundation models for medical image segmentation struggle under out-of-distribution (OOD) shifts, often producing fragmented false positives on OOD tumors. We introduce **R$^2$-Seg**, a **training-free** framework for robust OOD tumor segmentation that operates via a two-stage **Reason-and-Reject** process. First, the **Reason** step employs an LLM-guided anatomical reasoning planner to localize organ anchors and generate multi-scale ROIs. Second, the **Reject** step applies two-sample statistical testing to candidates generated by a frozen foundation model (BiomedParse) within these ROIs. This statistical rejection filter retains only candidates significantly different from normal tissue, effectively suppressing false positives. Our framework requires no parameter updates, making it compatible with zero-update test-time augmentation and avoiding catastrophic forgetting. On multi-center and multi-modal tumor segmentation benchmarks, **R$^2$-Seg** substantially improves Dice, specificity, and sensitivity over strong baselines and the original foundation models.

Automated video analysis is critical for wildlife conservation. A foundational task in this domain is multi-animal tracking (MAT), which underpins applications such as individual re-identification and behavior recognition. However, existing datasets are limited in scale, constrained to a few species, or lack sufficient temporal and geographical diversity -- leaving no suitable benchmark for training general-purpose MAT models applicable across wild animal populations. To address this, we introduce SA-FARI, the largest open-source MAT dataset for wild animals. It comprises 11,609 camera trap videos collected over approximately 10 years (2014-2024) from 741 locations across 4 continents, spanning 99 species categories. Each video is exhaustively annotated culminating in $\sim$46 hours of densely annotated footage containing 16,224 masklet identities and 942,702 individual bounding boxes, segmentation masks, and species labels. Alongside the task-specific annotations, we publish anonymized camera trap locations for each video. Finally, we present comprehensive benchmarks on SA-FARI using state-of-the-art vision-language models for detection and tracking, including SAM 3, evaluated with both species-specific and generic animal prompts. We also compare against vision-only methods developed specifically for wildlife analysis. SA-FARI is the first large-scale dataset to combine high species diversity, multi-region coverage, and high-quality spatio-temporal annotations, offering a new foundation for advancing generalizable multi-animal tracking in the wild. The dataset is available at [ANONYMIZED]

Instance-level object segmentation across disparate egocentric and exocentric views is a fundamental challenge in visual understanding, critical for applications in embodied AI and remote collaboration. This task is exceptionally difficult due to severe changes in scale, perspective, and occlusion, which destabilize direct pixel-level matching. While recent geometry-aware models like VGGT provide a strong foundation for feature alignment, we find they often fail at dense prediction tasks due to significant pixel-level projection drift, even when their internal object-level attention remains consistent. To bridge this gap, we introduce VGGT-Segmentor (VGGT-S), a framework that unifies robust geometric modeling with pixel-accurate semantic segmentation. VGGT-S leverages VGGT's powerful cross-view feature representation and introduces a novel Union Segmentation Head. This head operates in three stages: mask prompt fusion, coarse point-guided prediction, and iterative mask refinement, effectively translating high-level feature alignment into a precise segmentation mask. Furthermore, we propose a single-image self-supervised training strategy that eliminates the need for paired annotations and enables strong zero-shot generalization. On the challenging Ego–Exo4D benchmark, VGGT-S sets a new state-of-the-art, achieving 67.7% and 68.0% average IoU for Ego→Exo and Exo→Ego tasks, respectively, significantly outperforming prior methods. Notably, our zero-shot model surpasses most fully-supervised baselines, demonstrating the effectiveness and scalability of our approach.

Estimating accurate, view-consistent geometry and camera poses from uncalibrated multi-view/video inputs remains challenging—especially at high spatial resolutions and over long sequences. We present DAGE, a dual-stream transformer whose main novelty is to disentangle global coherence from fine detail. A low-resolution stream operates on aggressively downsampled frames with alternating frame/global attention to build a view-consistent representation and estimate cameras efficiently, while a high-resolution stream processes the original images per-frame to preserve sharp boundaries and small structures. A lightweight adapter fuses these streams via cross-attention, injecting global context without disturbing the pretrained single-frame pathway. This design scales resolution and clip length independently, supports inputs up to 2K, and maintains practical inference cost. DAGE delivers sharp depth/pointmaps, strong cross-view consistency, and accurate poses, establishing new state-of-the-art results for video geometry estimation and multi-view reconstruction.

We present Wave-Former, a novel method capable of high-accuracy 3D shape reconstruction for completely occluded, diverse, everyday objects. This capability can open new applications spanning robotics, augmented reality, and logistics. Our approach leverages millimeter-wave (mmWave) wireless signals, which can penetrate common occlusions and reflect off hidden objects. In contrast to past mmWave reconstruction methods, which suffer from limited coverage and high noise, Wave-Former introduces a physics-aware shape completion model capable of inferring full 3D geometry. At the heart of Wave-Former's design is a novel three-stage pipeline which bridges raw wireless signals with recent advancements in vision-based shape completion by incorporating physical properties of mmWave signals. The pipeline proposes candidate geometric surfaces, employs a transformer-based shape completion model designed specifically for mmWave signals, and finally performs entropy-guided surface selection. This enables Wave-Former to be trained using entirely synthetic point-clouds, while demonstrating impressive generalization to real-world data. In head-to-head comparisons with state-of-the-art baselines, Wave-Former raises recall from 54% to 72% while maintaining a high precision of 85%.

Recent advances in stereo matching have focused on accuracy, often at the cost of significantly increased model size. Traditionally, the community has regarded efficient models as incapable of zero-shot ability due to their limited capacity. In this paper, we introduce Lite Any Stereo, a stereo depth estimation framework that achieves strong zero-shot generalization while remaining highly efficient. To this end, we design a compact yet expressive backbone to ensure scalability, along with a carefully crafted hybrid cost aggregation module. We further propose a three-stage training strategy on million-scale data to effectively bridge the sim-to-real gap. Together, these components demonstrate that an ultra-light model can deliver strong zero-shot generalization, ranking 1st across four widely used real-world benchmarks. Remarkably, our model attains accuracy comparable to or exceeding state-of-the-art non-prior-based accurate methods while requiring less than 1% of their computational cost, setting a new standard for efficient stereo matching. Code will be made publicly available.

Self-supervised learning on images seeks to extract meaningful visual representations from unlabeled data. When scaled to large datasets, this paradigm has achieved state-of-the-art performance and the resulting trained models such as DINOv3 have seen widespread adoption. However, most prior efforts are optimized for semantic understanding rather than geometric reasoning. One important exception is Cross-View Completion, CroCo, which is a form of masked autoencoding (MAE) tailored for 3D understanding. In this work, we continue on the path proposed by CroCo and focus on learning features tailored for 3D vision. In a nutshell, we extend MAE to arbitrarily many views of the same scene. By uniformly masking all views and employing a lightweight decoder with inter-frame attention, our approach is inherently simpler and more scalable than CroCo. We evaluate the resulting model, MuM, extensively on downstream tasks including feedforward reconstruction, dense image matching and relative pose estimation, finding that it outperforms the state-of-the-art visual encoders DINOv3 and CroCo v2.

Feed-forward transformer models such as VGGT and $\pi^3$ are highly accurate, but their computational cost grows quadratically with the number of input images, making them slow to evaluate on large collections. More efficient approaches ameliorate this cost at the expense of reconstruction quality. We introduce Fast Reconstruction Model, a stateful feed-forward reconstruction model that uses a bidirectional architecture that scales linearly in the number of input views, while matching or surpassing the reconstruction quality of quadratic-time methods. FRM employs test-time training layers to compress images into a compact hidden scene state during a single forward pass, enabling our model to reconstruct 3D scenes at speeds up to 75 FPS on a single H100 GPU---over 20 times faster than SOTA methods such as VGGT. This hidden state also serves as an implicit scene representation which can be queried at real-time rates to produce colored point maps from novel views.

This paper addresses the task of large-scale 3D scene reconstruction from long video sequences. Recent feed-forward reconstruction models have shown promising results by directly regressing 3D geometry from RGB images without explicit 3D priors or geometric constraints. However, these methods often struggle to maintain reconstruction accuracy and consistency over long sequences due to limited memory capacity and the inability to effectively capture global contextual cues. In contrast, humans can naturally exploit the global understanding of the scene to inform local perception. Motivated by this, we propose a novel neural global context representation that efficiently compresses and retains long-range scene information, enabling the model to leverage extensive contextual cues for enhanced reconstruction accuracy and consistency. The context representation is realized through a set of lightweight neural sub-networks that are rapidly adapted during test time via self-supervised objectives, which substantially increases memory capacity without incurring significant computational overhead. The experiments on multiple large-scale benchmarks, including the KITTI Odometry~\cite{Schops2019CVPR} and Oxford Spires~\cite{tao2025spires} datasets, demonstrate the effectiveness of our approach in handling ultra-large scenes, achieving state-of-the-art performance in both pose estimation and 3D reconstruction accuracy while maintaining efficiency.

The task of multi-view inpainting necessitates 3D consistency in the inpainted images. Most prior methods first employ single-view 2D inpainting and then enforce multi-view consistency in a post-hoc 3D optimization stage, which leads to undesirable artifacts and lengthy optimization times. The existing single-stage method, MVInpainter, uses video priors and is pose-free, making it less suitable for inputs beyond video sequences. In this paper, we propose a framework that trains an inpainting model to condition on the explicit and reliable multi-view correspondences from a 3D foundation model. Central to our framework is a cross-view conditioning architecture, LaRP, carefully designed to utilize both the generative prior of a pretrained diffusion inpainting model and the reprojected cross-view appearance latents. We additionally propose a scalable data pipeline for stable training of LaRP. Extensive experiments demonstrate that LaRP outperforms prior methods in 3D consistency and novel view synthesis quality competitive with the state-of-the-art, while being ∼50x faster.

Multi-agent 3D reconstruction, as a key technology for large-scale VR/AR, robot swarms, and digital twins, has attracted growing attention. Recent end-to-end 3D reconstruction methods achieve strong performance in single-agent scenarios, but they are difficult to directly extend to multi-agent collaborative settings, where they often suffer from unstable tracking, excessive memory consumption, and frequent loop-closure failures, thus failing to meet real-time and large-scale deployment requirements. To address these issues, we propose TOPOMA, a real-time end-to-end 3D reconstruction framework tailored for multi-agent collaboration. TOPOMA explicitly models the spatial topological structure of the scene and tightly couples it with end-to-end representation learning, thereby jointly solving core challenges such as inter-agent spatial alignment and submap fusion. Concretely, we introduce topology skeleton modeling and optimization, decentralized loop closure, and topology-guided residual transport, and build upon them a fully distributed inference architecture in which each agent can independently store, reconstruct, and incrementally optimize its map while collaborating through lightweight topological information. Extensive experiments demonstrate that, compared with existing methods, TOPOMA achieves consistently higher trajectory accuracy, reconstruction quality, robustness, and topological consistency, showing superior adaptability and scalability.

We present ON3R, an online-trained neural regressor addressing sparse-view structureless localization, where database images have limited visual overlap and no prebuilt 3D map. Given any sparse matches between a query and a $K$-tuple of posed database views, ON3R predicts 3D coordinates for matched query keypoints, supervised by database reprojection residuals and a monocular depth prior. Afterwards, the absolute pose of the query is estimated via P3P-RANSAC and refined with lightweight bundle adjustment. Across MegaDepth, Cambridge Landmarks, and a sparsified version of Aachen Day-Night, ON3R outperforms existing methods. ON3R is particularly effective when the data is extremely sparse -- we focus on $K\leq10$ database images. The code, data splits, and SfM models will be made available for full reproducibility.

Reliable incremental estimation of camera poses and 3D reconstruction is key to enable various applications including robotics, interactive visualization, and augmented reality. However, this task is particularly challenging in dynamic natural environments, where scene dynamics can severely deteriorate camera pose estimation accuracy. In this work, we propose a novel monocular visual SLAM system that can robustly estimate camera poses in dynamic scenes. To this end, we leverage the complementary strengths of geometric patch-based online bundle adjustment and recent feed-forward reconstruction models. Specifically, we propose a feed-forward reconstruction model to precisely filter out dynamic regions, while also utilizing its depth prediction to enhance the robustness of the patch-based visual SLAM. By aligning depth prediction with estimated patches from bundle adjustment, we robustly handle the inherent scale ambiguities of the batch-wise application of the feed-forward reconstruction model. Extensive experiments on multiple tasks show the superior performance of our proposed method compared to state-of-the-art approaches.

Learning-based structure-from-motion methods such as ACE-Zero have demonstrated strong performance in estimating camera poses and scene coordinates from unordered image collections without requiring ground truth supervision. However, the lack of global and multi-view consistency constraints in ACE-Zero can lead to pose drift and misalignment, particularly in complex or ambiguous scenes. In this work, we propose a hybrid framework that integrates pose graph optimization (PGO) into ACE-Zero to refine camera poses and suppress incorrect refinements. We construct pose graphs directly from ACE-Zero outputs by extracting relative pose constraints from predicted scene coordinates. Furthermore, we introduce an uncertainty-aware optimization strategy by estimating confidence scores using geometric priors, including epipolar and optical flow consistencies across views. Our approach improves the robustness and accuracy of pose estimation, demonstrating that global geometric reasoning can effectively complement learning-based inference in structure-from-motion.

3D reconstruction from multi-view images is a core challenge in computer vision. Recently, feed-forward methods have emerged as efficient and robust alternatives to traditional per-scene optimization techniques. Among them, state-of-the-art models like the Visual Geometry Grounding Transformer (VGGT) leverage full self-attention over all image tokens to capture global relationships. However, this approach suffers from poor scalability due to the quadratic complexity of self-attention and the large number of tokens generated in long image sequences.In this work, we introduce FlashVGGT, an efficient alternative that addresses this bottleneck through a descriptor-based attention mechanism. Instead of applying dense global attention across all tokens, FlashVGGT compresses spatial information from each frame into a compact set of descriptor tokens. Global attention is then computed as cross-attention between the full set of image tokens and this smaller descriptor set, significantly reducing computational overhead.Moreover, the compactness of the descriptors enables online inference over long sequences via a chunk-recursive mechanism that reuses cached descriptors from previous chunks. Experimental results show that FlashVGGT achieves reconstruction accuracy competitive with VGGT while reducing inference time to just 9.3% for 1,000 images, and scaling efficiently to sequences exceeding 3,000 images.

We present the first study of cross-sensor view synthesis across different modalities. We examine a practical, fundamental, yet widely overlooked problem: getting aligned RGB-X data, where most RGB-X prior work assumes such pairs exist and focuses on modality fusion, but it empirically requires huge engineering effort in calibration. We propose a match-densify-consolidate method. First, we perform RGB-X image matching followed by guided point densification. Using the proposed confidence-aware densification and self-matching filtering, we attain better view synthesis and later consolidate them in 3D Gaussian Splatting (3DGS). Our method uses no 3D priors for X-sensor and only assumes nearly no-cost COLMAP for RGB. We aim to remove the cumbersome calibration for various RGB-X sensors and advance the popularity of cross-sensor learning by a scalable solution that breaks through the bottleneck in large-scale real-world RGB-X data collection. Code will be released.

Dynamic driving scene reconstruction is critical for autonomous driving simulation and closed-loop learning. While recent feed-forward methods have shown promise for 3D reconstruction, they struggle with long-range driving sequences due to quadratic complexity in sequence length and challenges in modeling dynamic objects over extended durations. We propose UFO, a novel recurrent paradigm that combines the benefits of optimization-based and feed-forward methods for efficient long-range 4D reconstruction. Our approach maintains a 4D scene representation that is iteratively refined as new observations arrive, using a visibility-based filtering mechanism to select informative scene tokens and enable efficient processing of long sequences. For dynamic objects, we introduce an object pose-guided modeling approach that supports accurate long-range motion capture. Experiments on the Waymo Open Dataset demonstrate that our method significantly outperforms both per-scene optimization and existing feed-forward methods across various sequence lengths. Notably, our approach can reconstruct 16-second driving logs within 0.5 second while maintaining superior visual quality and geometric accuracy.

With recent advances, Feed-forward Reconstruction Models (FFRMs) have demonstrated great potential in reconstruction quality and adaptiveness to multiple downstream tasks. However, the excessive reliance on multi-view geometric annotations, e.g. 3D point maps and camera poses, makes the fully-supervised training scheme of FFRMs difficult to scale up.In this paper, we propose Reliev3R, a weakly-supervised paradigm for training FFRMs from scratch without cost-prohibitive multi-view geometric annotations. Relieving the reliance on geometric sensory data and compute-exhaustive structure-from-motion preprocessing, our method draws 3D knowledge directly from monocular relative depths and image sparse correspondences given by zero-shot predictions of pretrained models.At the core of Reliev3R, we design an ambiguity-aware relative depth loss and a trigonometry-based reprojection loss to facilitate supervision for multi-view geometric consistency.Training from scratch with the less data, Reliev3R catches up with its fully-supervised sibling models, taking a step towards low-cost 3D reconstruction supervisions and scalable FFRMs.

3D vision foundation models have shown strong generalization in reconstructing key 3D attributes from uncalibrated images through a single feed-forward pass. However, when deployed in online settings such as driving scenarios, predictions are made over temporal windows, making it non-trivial to maintain consistency across time. Recent strategies align consecutive predictions by solving global transformation, yet our analysis reveals their fundamental limitations in assumption validity, local alignment scope, and robustness under noisy geometry. In this work, we propose a higher-DOF and long-term alignment framework based on Thin Plate Spline, leveraging globally propagated control points to correct spatially varying inconsistencies. In addition, we adopt a point-agnostic submap registration design that is inherently robust to noisy geometry predictions. The proposed framework is fully plug-and-play, compatible with diverse 3D foundation models and camera configurations (e.g., monocular or surround-view). Extensive experiments demonstrate that our method consistently yields more coherent geometry and lower trajectory errors across multiple datasets, backbone models, and camera setups, highlighting its robustness and generality. Codes are provided in the supplementary material and will be publicly released.

Structure-from-Motion -- the process of simultaneously estimating camera poses and 3D scene structure from a collection of images -- remains a central challenge in computer vision, with many open problems yet to be solved.Recent advances in feedforward 3D reconstruction have made significant strides in overcoming persistent failure cases of classical SfM methods, particularly in scenarios characterized by low texture, limited image overlap, and symmetries.However, while feedforward approaches excel in these challenging conditions, they often face limitations regarding scalability, accuracy, and robustness, and typically fall short of classical methods in standard reconstruction settings.In this work, we systematically analyze these limitations and propose a new state-of-the-art Structure-from-Motion pipeline by combining the respective strengths of classical and feedforward methods.Extensive experiments over a wide range of reconstruction scenarios demonstrate the benefits of our approach by achieving state-of-the-art results across the board.The implementation of our pipeline will be shared as open source software.

Current visual text generation models struggle with the trade-off between text accuracy and overall image coherence like aesthetic appeal during training. We find that to achieve a high text accuracy could reduce aesthetic score and instruction following capability. Although reinforcement learning approaches can alleviate the problem through aligning with multiple rewards, they are often unstable for text generation, as existing approaches normally optimize multiple rewards in a weighted-sum way. In addition, it is difficult to balance the weight of each reward. Moreover, reinforcement learning requires a set of training instructions. A large number of prompts require more training time and computing resources, while a small set leads to poor performance. Hence, how to select the prompts for efficient and effective training is an unsolved problem.In this study, we propose Pareto-Optimal Curriculum Alignment (POCA), a framework that addresses this issue as a multi-objective problem by: 1) identifying the Pareto-optimal set to avoid simple scalarization and 2) designing an adaptive curriculum alignment strategy to manage a learning sequence of a multi-reward dataset using automatic difficulty assessment, which is crucial for optimal convergence as RL methods explore in a limited data environment. In synergy, POCA finds the Pareto-optimal set in a unified reward space, which eliminates inconsistent signals to find the best trade-off solution from different rewards under an easy-to-hard optimization landscape. The experimental results show that POCA significantly improves all metrics such as CLIP, HPS scores and sentence accuracy.

Unified multimodal generation aims to jointly model image-to-text and text-to-image tasks within a single architecture. However, current approaches struggle to produce coherent, interleaved sequences of text and images. This limitation hinders applications that rely on tightly integrated multimodal outputs—such as step-by-step instructional guides, visual planning tools, and interactive content editing—where textual explanations and visual elements must be generated in a coordinated manner. We introduce DuetGen, a general-purpose interleaved multimodal generation model and investigate data curation, architecture design, and evaluation. In terms of data, we construct a large-scale high-quality instruction-tuning corpus combining curated web content, rewritten multimodal conversations, and diverse synthetic examples covering everyday scenarios. Architecturally, DuetGen builds upon a pretrained MLLM and diffusion transformer (DiT) pretrained on video generation, avoiding costly unimodal pretraining while remaining scalable. A two-stage decoupled training strategy first instruct-tunes the MLLM and then aligns it with the DiT using large-scale curated interleaved image–text sequences. Experiments on public and newly constructed benchmarks show that DuetGen substantially outperforms prior open-source systems across text quality, image fidelity, and image–context alignment, achieving substantial gains on text-to-image and image-editing benchmarks. Code and data will be released.

Creating new visual concepts often requires connecting distinct ideas through their most relevant shared attributes—their vibe. We introduce Vibe Blending, a novel task for generating coherent and meaningful hybrids that reveals these shared attributes between images. Achieving such blends is challenging for current methods, which struggle to identify and traverse nonlinear paths linking distant concepts in latent space. We propose Vibe Space, a hierarchical graph manifold that learns low-dimensional geodesics in feature spaces like CLIP, enabling smooth and semantically consistent transitions between concepts. To evaluate creative quality, we design a cognitively inspired framework combining human judgments, LLM reasoning, and a geometric path-based difficulty score. We find that Vibe Space produces blends that humans consistently rate as more creative and coherent than current methods.

Generating multi-frame, action-rich visual narratives without fine-tuning faces a threefold tension: action text faithfulness, subject identity fidelity, and cross frame background continuity. We propose StoryTailor, a zero-shot pipeline that runs on a single RTX 4090 (24 GB) and produces temporally coherent, identity-preserving image sequences from a long narrative prompt, per subject references, and grounding boxes. Three synergistic modules drive the system: Gaussian-Centered Attention (GCA) to dynamically focus on each subject core and ease grounding-box overlaps; Action-Boost Singular Value Reweighting (AB-SVR) to amplify action-related directions in the text embedding space; and Selective Forgetting Cache (SFC) that retains transferable background cues, forgets nonessential history, and selectively surfaces the retained cues to build cross scene semantic ties. Compared with baseline methods, the experiments show that CLIP-T improves by up to 10–15\%, with DreamSim lower than strong baselines, while CLIP-I stays in a visually acceptable, competitive range. With a matched resolution and steps on a 24 GB GPU, inference is faster than FluxKontext. Qualitatively, StoryTailor delivers expressive interactions and evolving yet stable scenes.

Creativity is a complex phenomenon. When it comes to representing and assessing creativity, treating it as a single undifferentiated quantity would appear naive and under-whelming. In this work, we learn the first type-specific creativity reward model, coined CREward, which spans three creativity “axes,” geometry, material, and texture, to allow us to view creativity through the lens of the image formation pipeline. To build our reward model, we first conduct a human benchmark evaluation to capture human perception of creativity for each type across various creative images. We then analyze the correlation between human judgments and predictions by large vision-language models (LVLMs), confirming that LVLMs exhibit strong alignment with human perception. Building on this observation, we collect LVLM-generated labels to train our CREward model that is applicable to both evaluation and generation of creative images. We explore three applications of CREward: creativity assessment, explainable creativity, and creative sample acquisition for both human design inspiration and guiding creative generation through low-rank adaptation.

We present LumiX, a structured diffusion framework for coherent text-to-intrinsic generation. Conditioned on text prompts, LumiX jointly generates a comprehensive set of intrinsic maps (e.g., albedo, irradiance, normal, depth, and final color), providing a structured and physically consistent description of an underlying scene. This is enabled by two key contributions: 1) Query-Broadcast Attention, a mechanism that ensures structural consistency by sharing queries across all maps in each self-attention block. 2) Tensor LoRA, a tensor-based adaptation that parameter-efficiently models cross-map relations for efficient joint training. Together, these designs enable stable joint diffusion training and unified generation of multiple intrinsic properties. Experiments show that LumiX produces coherent and physically meaningful results, achieving 23% higher alignment and a better preference score (0.19 vs. -0.41) compared to the state of the art, and it can also perform image-conditioned intrinsic decomposition within the same framework.

Text-to-Image (T2I) generation has long been an open problem, with compositional synthesis remaining particularly challenging. This task requires accurate rendering of complex scenes containing multiple objects that exhibit diverse attributes as well as intricate spatial and semantic relationships, demanding both precise object placement and coherent inter-object interactions. In this paper, we propose a novel compositional curriculum reinforcement learning framework named CompGen that addresses compositional weakness in existing T2I models. Specifically, we leverage scene graphs to establish a novel difficulty criterion for compositional ability and develop a corresponding adaptive Markov Chain Monte Carlo graph sampling algorithm. This difficulty-aware approach enables the synthesis of training curriculum data that progressively optimize T2I models through reinforcement learning. We integrate our curriculum learning approach into Group Relative Policy Optimization (GRPO) and investigate different curriculum scheduling strategies. Our experiments reveal that CompGen exhibits distinct scaling curves under different curriculum scheduling strategies, with easy-to-hard and Gaussian sampling strategies yielding superior scaling performance compared to random sampling. Extensive experiments demonstrate that CompGen significantly enhances compositional generation capabilities for both diffusion-based and auto-regressive T2I models, highlighting its effectiveness in improving the compositional T2I generation systems.

Multimodal generative models can process instructions in various modalities and demonstrate outstanding performance across a wide range of image generation tasks. However, their robustness in complex real-world scenarios remains limited due to insufficient generalized instruction alignment. We introduces \textbf{OmniGen2}, a unified multimodal generator designed to follow complex, fine-grained instructions. Our core contribution is a two-stage design that first builds a strong, world-knowledge-grounded foundation model and then aligns it using a progressive, multi-task instruction tuning strategy. The foundation model features a streamlined architecture with decoupled decoding for versatile multimodal generation and a novel positional encoding scheme to improve learning efficiency. We ground this model in real-world knowledge using large-scale data construction pipelines. Building on this foundation, we propose a progressive, reinforcement-based alignment process. This phase carefully schedules training tasks and reward signals to foster cross-task knowledge transfer, significantly improving the model's instruction-following capabilities. Our models demonstrate competitive performance on standard benchmarks and our dedicated in-context generation benchmark, \textbf{OmniContext}. We will release our models, code, benchmark, and training datasets to catalyze future research in building more capable and instruction aligned generative models.

Text-to-image models are increasingly utilized in design workflows, but articulating nuanced design intentions through text remains a challenge. This work proposes a method that extracts a visual attribute from a reference image and injects it directly into the generation pipeline. The method optimizes a text token to exclusively represent the target attribute using a custom training prompt and two novel embeddings: distilled embedding and residual embedding. Through this approach, a wide range of attributes can be extracted, including the shape, material, or color of an object, as well as the camera angle of the image. The method is validated on various target attributes and text prompts drawn from a newly constructed dataset. The results show that it outperforms existing approaches in selectively extracting and applying target attributes across diverse contexts. Ultimately, the proposed method enables intuitive and controllable text-to-image generation, streamlining the design process.

Personalizing visual generative models to meet specific user needs has gained increasing attention, yet current methods like Low-Rank Adaptation (LoRA) remain impractical due to their demand for task-specific data and lengthy optimization. While a few hypernetwork-based approaches attempt to predict adaptation weights directly, they struggle to map fine-grained user prompts to complex LoRA distributions, limiting their practical applicability.To bridge this gap, we propose LoFA, a general framework that efficiently predicts personalized priors for fast model adaptation. We first identify a key property of LoRA: structured distribution patterns emerge in the relative changes between LoRA and base model parameters. Building on this, we design a two-stage hypernetwork: first predicting sparse response maps that capture key adaptation regions, then using these to guide final LoRA weight prediction. Extensive experiments demonstrate that our method consistently predicts high-quality personalized priors within seconds, across multiple tasks and user prompts, even outperforming conventional LoRA that requires hours of processing.

Current text-to-image (T2I) generation models often struggle with prompts that require complex reasoning or specialized knowledge, failing to accurately interpret implicit user intent. To bridge this gap, we introduce \textbf{T2I-Reason}, a large-scale dataset designed to empower text-to-image generation in unified multimodal models (UMMs) with reasoning and knowledge. The dataset contains 120k pairs of text triplet and image. The text triplet consists of (1) an implicit prompt, which requires reasoning or knowledge to decipher its underlying meaning; (2) a reasoning chain, which provides a step-by-step analysis to resolve the implicit prompt's meaning; and (3) an explicit prompt, a clear and straightforward visual description prepared for T2I generation. T2I-Reason is meticulously constructed: 65k samples are dedicated to reasoning, specifically targeting arithmetic reasoning, spatial-attribute relationship reasoning, deductive reasoning (cause to effect), and abductive reasoning (effect to cause). While 55k samples necessitate specialized knowledge, which covers multiple disciplines, spatial-temporal concepts, and entity knowledge. To validate the effectiveness of our dataset, we train a unified multimodal model, Bagel, on our dataset. Results across multiple benchmarks that evaluate the reasoning capabilities of T2I generation demonstrate that our model achieves significant and consistent improvements on both composition and reasoning, confirming that explicit training on intermediate reasoning chains is a pivotal step towards more intelligent unified generative models.

Personalized visual understanding has advanced significantly, yet existing approaches struggle to localize and extract specific concepts when input images contain multiple objects. Many prior methods rely heavily on segmentation-based supervision or exhibit poor compositional generalization, limiting their ability to accurately disentangle and manipulate individual concepts. In this work, we propose UniVerse, a Unified Modulation Framework for segmentation-free, disentangled multi-concept personalization in diffusion transformers. Our method allows for composable and decomposable concept extraction, enabling fine-grained localization and representation of target objects without explicit segmentation masks. UniVerse learns to decompose complex scenes into concept-specific representations and then compose them in a unified manner, enabling robust personalization across diverse visual contexts. Through extensive experiments on multiple benchmarks, we demonstrate that UniVerse significantly outperforms state-of-the-art baselines in both localization accuracy and visual fidelity. Qualitative and quantitative results show that our approach can precisely extract target concepts in cluttered scenes, paving the way for more flexible, interpretable, and personalized visual generation and understanding.

Visual Autoregressive (AR) models generate images by predicting discrete tokens that are decoded by a visual tokenizer.Despite demonstrating strong overall image generation ability, they still underperform on text rendering with blur strokes and disrupt letter shapes. In this work, we trace this limitation to the visual tokenizer, which struggles to reconstruct fine-grained detail.Improving the tokenizer is straightforward but expensive, as it necessitates retraining both the tokenizer and the AR model.Can we improve text rendering performance of AR models without retraining the existing tokenizer and AR model? To achieve this, we propose the Residual Decoder Adapter(\method) that upgrades an existing tokenizer post-hoc without changing its token space. Specifically, it refines the decoder output of the visual tokenizer by introducing two novel components: (i) a paired codebook that shares the token distribution with the original one; (ii) a parallel branch to learn the tiny differences (residual) between the reconstructed image and the ground-truth images in the pixel space. This residual design allows us to enhance the tokenizer non-invasively while preserving compatibility with prior AR models. \method substantially improves text rendering significantly by a large margin. For instance, we boost finetuned Janus-Pro OCR accuracy rises from 24.52\% to 58.26\% (TextVisionBlend), from 12.75\% to 36.81\% (StyledTextSynth) on competitive TextAtlas benchmark. Codes will be released.

Text-to-video generation has advanced rapidly in visual fidelity, whereas standard methods still have limited ability to control the subject composition of generated scenes. Prior work shows that adding localized text control signals, such as bounding boxes or segmentation masks, can help. However, these methods struggle in complex scenarios and degrade in multi-object settings, offering limited precision and lacking a clear correspondence between individual trajectories and visual entities as the number of controllable objects increases. We introduce Text-Grounded Trajectories (TGT), a framework that conditions video generation on trajectories paired with localized text descriptions. We propose $\textit{Location-Aware Cross-Attention}$ (LACA) to integrate these signals and adopt a dual-CFG scheme to separately modulate local and global text guidance. In addition, we develop a data processing pipeline that produces trajectories with localized descriptions of tracked entities, and we annotate two million high quality video clips to train TGT. Together, these components enable TGT to use point trajectories as intuitive motion handles, pairing each trajectory with text to control both appearance and motion. Extensive experiments show that TGT achieves higher visual quality, more accurate text alignment, and improved motion controllability compared with prior approaches.

Recent text-to-image (T2I) diffusion models achieve remarkable realism, yet faithful prompt–image alignment remains challenging, particularly for complex prompts with multiple objects, relations, and fine-grained attributes. Existing training-free inference-time scaling methods rely on fixed iteration budgets that cannot adapt to prompt difficulty, while reflection-tuned models require carefully curated reflection datasets and extensive joint fine-tuning of diffusion and vision–language models, often overfitting to reflection paths data and lacking transferability across models. We introduce RAISE (Requirement-Adaptive Self-Improving Evolution), a training-free, requirement-driven evolutionary framework for adaptive T2I generation. RAISE formulates image generation as a requirement-driven adaptive scaling process, evolving a population of candidates at inference time through a diverse set of refinement actions—including prompt rewriting, noise resampling, and instructional editing. Each generation is verified against a structured checklist of requirements, enabling the system to dynamically identify unsatisfied items and allocate further computation only where needed. This achieves adaptive test-time scaling that aligns computational effort with semantic query complexity. On GenEval and DrawBench, RAISE attains state-of-the-art alignment (0.94 overall GenEval) while incurring fewer generated samples (reduced by 30-40%) and VLM calls (reduced by 80%) than prior scaling and reflection-tuned baselines, demonstrating efficient, generalizable, and model-agnostic multi-round self-improvement.

In this paper, we present FlowFixer, a refinement framework for subject-driven generation (SDG) that restores fine details lost during scene generation caused by changes in scale and perspective of a subject. FlowFixer proposes direct image-to-image translation from visual references, avoiding ambiguities in language prompts. To enable image-to-image training, we introduce a one-step denoising scheme to generate self-supervised training data, which automatically removes high-frequency details while preserving global structure, effectively simulating real-world SDG errors. We further propose a keypoint matching-based metric to properly assess fidelity in details beyond semantic similarities usually measured by CLIP or DINO. Experimental results demonstrate that FlowFixer outperforms state-of-the-art SDG methods in both qualitative and quantitative evaluations, setting a new benchmark for high-fidelity subject-driven generation.

Visual Text Rendering (VTR) remains a critical challenge in text‑to‑image generation, where even advanced models frequently produce text with structural anomalies such as distortion, blurriness, and misalignment.% We identify a key bottleneck across both VTR evaluation and Reinforcement Learning (RL) processes: current evaluators and reward models lack the ability for fine-grained structural perception. However, we find that leading MLLMs and specialist OCR models largely fail to perceive these structural anomalies, creating a critical bottleneck for both VTR evaluation and RL‑based optimization. As a result, even state‑of‑the‑art generators (e.g., SeedDream4.0, Qwen‑Image) still struggle to render structurally faithful text.To address this, we propose TextPecker,a plug-and-play structural anomaly perceptive RL strategy that mitigates noisy reward signals and works with any text-to-image generator. To enable this capability, we construct a recognition dataset with character‑level structural‑anomaly annotations and develop a stroke‑editing synthesis engine to expand structural‑error coverage. Experiments show that TextPecker consistently improves diverse text‑to‑image models; even on the well‑optimized Qwen‑Image, it significantly yields average gains of 4\% in structural fidelity and 8.7\% in semantic alignment for Chinese text rendering, establishing a new state-of-the-art in high-fidelity VTR.Our work fills a gap in VTR optimization, providing a foundational step towards reliable and structural faithful visual text generation.

Diffusion transformers have recently delivered strong text-to-image generation around 1K resolution, but we show that extending them to native 4K across diverse aspect ratios exposes a tightly coupled failure mode spanning positional encoding, VAE compression, and optimization. Tackling any of these factors in isolation leaves substantial quality on the table. We therefore take a data--model co-design view and introduce UltraFlux, a Flux-based DiT trained natively at 4K on MultiAspect-4K-1M, a 1M-image 4K corpus with controlled multi-AR coverage, bilingual captions, and rich VLM/IQA metadata for resolution- and AR-aware sampling. On the model side, UltraFlux couples (i) Resonance 2D RoPE with YaRN for training-window-, frequency-, and AR-aware positional encoding at 4K; (ii) a simple, non-adversarial VAE post-training scheme that improves 4K reconstruction fidelity; (iii) an SNR-Aware Huber Wavelet objective that rebalances gradients across timesteps and frequency bands; and (iv) a Stage-wise Aesthetic Curriculum Learning strategy that concentrates high-aesthetic supervision on high-noise steps governed by the model prior. Together, these components yield a stable, detail-preserving 4K DiT that generalizes across wide, square, and tall ARs. On the Aesthetic-Eval@4096 benchmark and multi-AR 4K settings, UltraFlux consistently outperforms strong open-source baselines across fidelity, aesthetic, and alignment metrics, and—with a LLM prompt refiner—matches or surpasses the proprietary Seedream 4.0.

Fashion design aims to express a designer’s creative intent and to depict how garments interact with the human body. Recent generative approaches condition on multimodal inputs to support garment editing and enable virtual try-on. However, existing methods still (i) confine design to garment-related images, excluding creative design sources such as artwork, abstract imagery, and natural photographs, and (ii) cannot support complete outfits, including accessories. We present FEAT Fashion Editing and Try-On from Any Design, a method that enables editing and try-on across both garments and accessories using diverse design sources. To achieve this, we introduce Disentangled Dual Injection (DDI). It takes both apparel and non-apparel design sources and selectively injects design cues via content and style disentanglement. Furthermore, we propose Orthogonal-Guided Noise Fusion (OGNF), a training-free mechanism that removes residual garments via orthogonal projection and applies region-specific noise strategies to enable virtual try-on for both garments and accessories. Extensive experiments demonstrate that FEAT achieves state-of-the-art performance in design flexibility, prompt consistency, and visual realism.

Achieving precise alignment between user intent and generated visuals remains a central challenge in text-to-visual generation, as a single attempt often fails to produce the desired output. To handle this, prior approaches mainly scale the visual generation process (e.g., increasing sampling steps or seeds), but this quickly leads to a quality plateau. This limitation arises because the prompt, crucial for guiding generation, is kept fixed. To address this, we propose Prompt Redesign for Inference-time Scaling, coined PRIS, a framework that adaptively revises the prompt during inference in response to the scaled visual generations. The core idea of PRIS is to review the generated visuals, identify recurring failure patterns across visuals, and redesign the prompt accordingly before regenerating the visuals with the revised prompt.To provide precise alignment feedback for prompt revision, we introduce a new verifier, element-level factual correction, which evaluates the alignment between prompt attributes and generated visuals at a fine-grained level, achieving more accurate and interpretable assessments than holistic measures.Extensive experiments on both text-to-image and text-to-video benchmarks demonstrate the effectiveness of our approach, including a 15\% gain on VBench 2.0. These results highlight that jointly scaling prompts and visuals is key to fully leveraging scaling laws at inference-time.

The development of 3D Vision-Language Models (VLMs), crucial for applications in robotics, autonomous driving, and augmented reality, is severely constrained by the scarcity of paired 3D-text data. Existing methods rely solely on next-token prediction loss, using only language tokens for supervision. This results in inefficient utilization of limited 3D data and leads to a significant degradation and loss of valuable geometric information in intermediate representations.To address these limitations, we propose {\mname}, a novel feature-level alignment regularization method. {\mname} explicitly supervises intermediate point cloud representations to preserve fine-grained 3D geometric-semantic information throughout the language modeling process. Specifically, we constrain the intermediate point cloud tokens within the LLM to align with visual input tokens via a consistency loss. By training only a lightweight alignment projector and LoRA adapters, {\mname} achieves explicit feature-level supervision with minimal computational overhead, effectively preventing geometric degradation.Extensive experiments on ModelNet40 and Objaverse datasets demonstrate that our method achieves \textbf{2.08} pp improvement on average for classification tasks, with a substantial \textbf{7.50} pp gain on the challenging open-vocabulary Objaverse classification task and \textbf{4.88} pp improvement on 3D object captioning evaluated by Qwen2-72B-Instruct, validating the effectiveness of {\mname}.

Contrastive pre-training frameworks such as CLIP have demonstrated impressive zero-shot performance across a range of vision-language tasks. Recent studies have shown that aligning individual text tokens with specific image patches or regions enhances fine-grained compositional understanding. However, it remains challenging to capture compositional semantics spanning multiple image regions.To address this limitation, we propose PowerCLIP, a novelcontrastive pre-training framework enhanced by powerset alignment, which exhaustively optimizes region-to-phrase alignments by minimizing the loss defined between powersets of image regions and textual parse trees.As this approach increases computational complexity exponentially due to the combinatorial explosion in the number of region subsets, we introduce efficient non-linear aggregators (NLAs) that reduce complexity from $\mathit{\mathcal{O}(2^{M})}$ to $\mathit{\mathcal{O}(M)}$ with respect to the number of regions $M$, provably approximating the exact loss value with arbitrary precision.Our extensive experiments demonstrate that PowerCLIPoutperforms state-of-the-art methodsin zero-shot classification and retrieval tasks, underscoring compositionality and robustness of our approach. Our code will be made publicly available.

We aim to learn a joint representation between inertial measurement unit (IMU) signals and 2D pose sequences extracted from video, enabling accurate cross-modal retrieval, temporal synchronization, subject and body-part localization, and action recognition. To this end, we introduce MoBind, a hierarchical contrastive learning framework designed to address three challenges: (1) filtering out irrelevant visual background, (2) modeling structured multi-sensor IMU configurations, and (3) achieving fine-grained, sub-second temporal alignment. To isolate motion-relevant cues, MoBind aligns IMU signals with skeletal motion sequences rather than raw pixels. We further decompose full-body motion into local body-part trajectories, pairing each with its corresponding IMU to enable semantically grounded multi-sensor alignment. To capture detailed temporal correspondence, MoBind employs a hierarchical contrastive strategy that first aligns token-level temporal segments, then fuses local (body-part) alignment with global (body-wide) motion aggregation. Evaluated on mRi, TotalCapture, and EgoHumans, MoBind consistently outperforms strong baselines across all four tasks, demonstrating robust fine-grained temporal alignment while preserving coarse semantic consistency across modalities.

Fine-tuning approaches for Vision-Language Models (VLMs) face a critical three-way trade-off between In-Distribution (ID) accuracy, Out-of-Distribution (OOD) generalization, and adversarial robustness. Existing robust fine-tuning strategies resolve at most two axes of this trade-off. Generalization-preserving methods retain ID/OOD performance but leave models vulnerable to adversarial attacks, while adversarial training improves robustness to targeted attacks but degrades ID/OOD accuracy. Our key insight is that the robustness trade-off stems from two geometric failures: sharp, anisotropic minima in parameter space and unstable feature representations that deform under perturbation. To address this, we propose GRACE (Gram-aligned Robustness via Adaptive Curvature Estimation), a unified fine-tuning framework that jointly regularizes the parameter-space curvature and feature-space invariance for VLMs. Grounded in Robust PAC-Bayes theory, GRACE employs adaptive weight perturbations scaled by local curvature to promote flatter minima, combined with a feature alignment loss that maintains representation consistency across clean, adversarial, and OOD inputs. On ImageNet fine-tuning of CLIP models, GRACE simultaneously improves ID accuracy by 10.8%, and adversarial accuracy by 8.9% while maintaining 57.0% OOD accuracy (vs. 57.4% zero-shot baseline). Geometric analysis confirms that GRACE converges to flatter minima without feature distortion across distribution shifts, providing a principled step toward generalized robustness in foundation VLMs.

Hateful meme classification aims to identify memes containing hateful content and has become increasingly important in the era of social media dominance. Large multimodal models (LMMs) have significantly enhanced the understanding of multimodal content, advancing this field. However, cognitive biases in LMMs can impede effective collaboration among models. To address this issue, we introduce \textbf{GECO}, a \textbf{G}ame-theoretic multi-ag\textbf{E}nt \textbf{C}ollaboration framew\textbf{O}rk that organizes multiple LMMs into interacting agents and employs game-theoretic principles to guide them toward an optimal cooperative equilibrium. GECO integrates a mixed bonus scheme, incorporating both individual accuracy and cross-model agreement to promote convergence toward a consistent cooperative solution. In addition, we implement efficient policy learning and introduce a penalty coefficient to optimize the framework effectively and ensure training stability. Extensive experiments on five publicly available benchmarks demonstrate that our framework achieves new state-of-the-art performance.

Image captioning remains a fundamental task for vision–language understanding, yet ground-truth supervision still relies predominantly on human-annotated references.Because human annotations reflect subjective preferences and expertise, ground-truth captions are often incomplete or even incorrect, which in turn limits caption models.We argue that caption quality should be assessed by two objective aspects: completeness (does the caption cover all salient visual facts?) and correctness (are the descriptions true with respect to the image?).To this end, we introduce CCCaption: a dual-reward reinforcement learning framework with a dedicated fine-tuning corpus that explicitly optimizes these properties to generate \textbf{C}omplete and \textbf{C}orrect \textbf{Captions}.For completeness, we use diverse LVLMs to disentangle the image into a set of visual queries, and reward captions that answer more of these queries, with a dynamic query sampling strategy to improve training efficiency.For correctness, we penalize captions that contain hallucinations by validating the authenticity of sub-caption queries, which are derived from the caption decomposition.Our symmetric dual-reward optimization jointly maximizes completeness and correctness, guiding models toward captions that better satisfy these objective criteria. Extensive experiments across standard captioning benchmarks show consistent improvements, offering a principled path to training caption models beyond human-annotation imitation.

Multimodal Large Language Models (MLLMs) have recently demonstrated promising capabilities in multimodal coding tasks such as chart-to-code generation. However, existing methods primarily rely on supervised fine-tuning (SFT), which requires the model to learn code patterns through chart-code pairs but does not reveal the model in a code execution environment. Moreover, while self-correction through execution feedback offers a potential route to improve coding quality, even state-of-the-art MLLMs have been shown to struggle with effective self-correction. In this work, we introduce MM-ReCoder, a chart-to-code generation model trained with reinforcement learning (RL) and equipped with self-correction ability. We propose a two-stage multi-turn self-correction RL strategy based on Group Relative Policy Optimization (GRPO) . The first stage enhances the model's self-correction ability via rolling out a shared first turn, while the second stage improves the coding capability with full-trajectory optimization. MM-ReCoder learns to produce more accurate and executable code through the interaction with the environment and by iteratively correcting its own outputs. Our results on three chart-to-code benchmarks demonstrate the state-of-the-art performance of MM-ReCoder.

Recently, unified multimodal models (UMMs) have made remarkable progress in integrating visual understanding and generation, demonstrating strong potential for complex text-to-image (T2I) tasks. Despite their theoretical promise, a persistent capability gap exists: UMMs typically exhibit superior visual understanding but comparatively weaker generative capabilities.This discrepancy arises largely from the intrinsic decoupling between the understanding and generation processes.While a UMM can accurately interpret fine-grained visual details, it often struggles to produce semantically coherent images from complex textual prompts.To address this challenge, we explore UMMs' internal understanding capability to enhance generation quality.We propose a token-level intrinsic text-image alignment reward mechanism,GvU, enabling the UMM to act simultaneously as teacher and student: it evaluates its own outputs using the understanding branch to guide the generations accordingly. Building upon this, we design a self-supervised reinforcement learning framework, allowing UMMs to iteratively improve their generation quality through understanding-based intrinsic reward signals—without reliance on external supervision.Experimental results show that our method substantially boosts UMMs' generation, which in turn strengthens their fine-grained visual understanding, narrowing the capability gap between UMMs' visual understanding and generation.

Symbolic computer vision represents diagrams through explicit logical rules and structured representations, enabling interpretable understanding in machine vision. This requires fundamentally different learning paradigms from pixel-based visual models. Symbolic visual learners parse diagrams into geometric primitives—points, lines, and shapes—whereas pixel-based learners operate on textures and colors. We propose a novel self-supervised symbolic auto-encoder that encodes diagrams into structured primitives and their interrelationships within the latent space, and decodes them through our executable engine to reconstruct the input diagrams. Central to this architecture is SymHPR (Symbolic Hierarchical Process Reward Modeling), which applies hierarchical step-level parsing rewards to enforce point-on-line, line-on-shape, and shape-on-relation consistency. Since vanilla reinforcement learning exhibits poor exploration in the policy space during diagram reconstruction; we thus introduce stabilization mechanisms to balance exploration and exploitation. We fine-tune our symbolic encoder on downstream tasks, developing a neuro-symbolic system that integrates the reasoning capabilities of neural networks with the interpretability of symbolic models through reasoning-grounded visual rewards. Evaluations across reconstruction, perception, and reasoning tasks demonstrate the effectiveness of our approach: achieving a 98.2\% reduction in MSE for geometric diagram reconstruction, surpassing GPT-4o by 0.6\% with a 7B model on chart reconstruction, and improving by +13\% on the MathGlance perception benchmark, and by +3\% on MathVerse and GeoQA reasoning benchmarks.

Reward models are critical for aligning vision-language systems with human preferences, yet current approaches suffer from hallucination, weak visual grounding, and an inability to use tools for verification, limiting their reliability on complex multimodal reasoning tasks.We present ARM-Thinker, an Agentic multimodal Reward Model that autonomously invokes external tools (e.g., image cropping, doc page retrieval) to ground judgments in verifiable evidence, replacing static, non-interactive reward scoring.This enables the model to verify fine-grained visual details, cross-reference multi-page evidence, and validate reasoning claims, which are capabilities absent in existing reward models.We train ARM-Thinker with multi-stage reinforcement learning, jointly optimizing tool-calling decisions and judgment accuracy.To evaluate agentic reward modeling, we introduce ARMBench-VL, comprising three benchmarks that assess fine-grained visual grounding (image-level tools), multi-page document understanding (retrieval tools), and instruction following (text-level verification).ARM-Thinker achieves +16.2\% average improvement on reward modeling benchmarks, +9.6\% on tool-use tasks, and outperforms baselines on multimodal math and logical reasoning benchmarks.Our results demonstrate that agentic capabilities significantly enhance both accuracy and interpretability of reward models.

Generating new Low-Rank Adaptation (LoRA) weights from pre-trained LoRAs has demonstrated strong generalization capabilities across various tasks, enabling the efficient transfer of AI models, particularly on resource-constrained edges. However, previous studies either merge base LoRAs via weighting coefficients or train a generative model under the closed-world assumption, limiting their efficiency and flexibility in complex edge user cases. This challenge may further increase when there are significant domain shifts between training and deployment. To this end, we propose $S$emantic-$G$uided $LoRA$ Parameter Generation ($SG$-$LoRA$), a tuning-free generative framework to efficiently produce task-specific parameters for unseen tasks in a semantic-to-LoRA pipeline. Concretely, SG-LoRA uses task descriptions as the semantic bridge, measuring their proximity to a set of known expert tasks in a shared embedding space. Based on this semantic guidance, it models the target task's LoRA parameter distribution to generate high-performing parameters for novel tasks. SG-LoRA enables the real-time construction of LoRA models aligned with individual intents by distilling knowledge from prominent LoRA experts, while also offering a privacy-preserving solution for personalized model adaptation in a novel zero-shot open-world setting proposed in this work. Extensive experiments on multiple challenging tasks confirm the superior performance and remarkable adaptability of SG-LoRA. The code is attached in the supplementary material.

Test-time adaptation (TTA) aims to mitigate performance degradation under distribution shifts by updating model parameters during inference. Existing approaches have primarily framed adaptation around affine modulation, focusing on recalibrating normalization layers. This perspective, while effective, overlooks another influential component in representation dynamics: the activation function. We revisit this overlooked space and propose AcTTA, an activation-aware framework that reinterprets conventional activation functions from a learnable perspective and updates them adaptively at test time. AcTTA reformulates conventional activation functions (e.g., ReLU, GELU) into parameterized forms that shift their response threshold and modulate gradient sensitivity, enabling the network to adjust activation behavior under domain shifts. This functional reparameterization enables continuous adjustment of activation behavior without modifying network weights or requiring source data. Despite its simplicity, AcTTA achieves robust and stable adaptation across diverse corruptions. Across CIFAR10-C, CIFAR100-C, and ImageNet-C, AcTTA consistently surpasses normalization-based TTA methods. Our findings highlight activation adaptation as a compact and effective route toward domain-shift–robust test-time learning, broadening the prevailing affine-centric view of adaptation.

Balancing performance trade-off on long-tail data distributions remains a long-standing challenge. In this paper, we posit that this dilemma stems from a phenomenon called "catastrophic forgetting'' in continual learning (the model tends to severely overfit on head classes while quickly forgetting tail classes) and pose a solution from a loss landscape perspective. We observe that different classes possess divergent convergence points in the loss landscape. Besides, this divergence is aggravated when the model settles into sharp and non-robust minima, rather than a shared and flat solution that beneficial for all classes. In light of this, we propose a continual learning inspired framework to prevent "catastrophic forgetting''. To avoid inefficient per-class parameter preservation, a Grouped Knowledge Preservation module is proposed to memorize group-specific convergence parameters, promoting convergence towards a shared solution. Concurrently, our framework integrates a Grouped Sharpness Aware module to seek flatter minima by explicitly addressing the geometry of the loss landscape. Notably, our framework requires neither external training samples nor pre-trained models, facilitating the broad applicability. Extensive experiments on four benchmarks demonstrate significant performance gains over state-of-the-art methods.

Existing active learning (AL) strategies capture fundamentally different notions of data value, e.g., uncertainty or representativeness. Consequently, the effectiveness of strategies can vary substantially across datasets, models, and even AL cycles. Committing to a single strategy risks suboptimal performance, as no single strategy dominates throughout the entire AL process. We introduce REFINE, an ensemble AL method that combines multiple strategies without knowing in advance which will perform best. In each AL cycle, REFINE operates in two stages: (1) Progressive filtering iteratively refines the unlabeled pool by considering an ensemble of AL strategies, retaining promising candidates capturing different notions of value. (2) Coverage-based selection then chooses a final batch from this refined pool, ensuring all previously identified notions of value are accounted for. Extensive experiments across 6 classification datasets and 3 foundation models show that REFINE consistently outperforms individual strategies and existing ensemble methods. Notably, progressive filtering serves as a powerful preprocessing step that improves the performance of any individual AL strategy applied to the refined pool, which we demonstrate on an audio spectrogram classification use case. Finally, the ensemble of REFINE can be easily extended with upcoming state-of-the-art AL strategies.

Model merging aims to integrate multiple task-adapted models into a unified model that preserves the knowledge of each task.In this paper, we identify that the key to this knowledge retention lies in maintaining the directional consistency of singular spaces between merged multi-task vector and individual task vectors. However, this consistency is frequently compromised by two issues: i) an imbalanced energy distribution within task vectors, where a small fraction of singular values dominate the total energy, leading to the neglect of semantically important but weaker components upon merging, and ii) the geometric inconsistency of task vectors in parameter space, which causes direct merging to distort their underlying directional geometry. To address these challenges, we propose DC-Merge, a method for directional-consistent model merging. It first balances the energy distribution of each task vector by smoothing its singular values, ensuring all knowledge components are adequately represented. These energy-balanced vectors are then projected onto a shared orthogonal subspace to align their directional geometries with minimal reconstruction error. Finally, the aligned vectors are aggregated in the shared orthogonal subspace and projected back to the original parameter space. Extensive experiments on vision and vision-language benchmarks show that DC-Merge consistently achieves state-of-the-art performance in both full fine-tuning and LoRA settings.

On-the-fly category discovery (OCD) aims to recognize known categories while simultaneously discovering novel ones from an unlabeled online stream, using a model trained only on labeled data. Existing approaches freeze the feature extractor trained offline and employ a hash-based framework that quantizes features into binary codes as class prototypes. However, discovering novel categories with a fixed knowledge base is counterintuitive, as the learning potential of incoming data is entirely neglected. In addition, feature quantization introduces information loss, diminishes representational expressiveness, and amplifies intra-class variance. It often results in category explosion, where a single class is fragmented into multiple pseudo-classes. To overcome these limitations, we propose a test-time adaptation framework that enables learning through discovery. It incorporates two complementary strategies: a semantic-aware prototype update and a stable test-time encoder update. The former dynamically refines class prototypes to enhance classification, whereas the latter integrates new information directly into the parameter space. Together, these components allow the model to continuously expand its knowledge base with newly encountered samples. Furthermore, we introduce a margin-aware logit calibration in the offline stage to enlarge inter-class margins and improve intra-class compactness, thereby reserving embedding space for future class discovery. Experiments on standard OCD benchmarks demonstrate that our method substantially outperforms existing hash-based state-of-the-art approaches, yielding notable improvements in novel-class accuracy and effectively mitigating category explosion.

Event-based low-light image enhancement (LIE) methods mainly focus on incorporating high dynamic range (HDR) information from events while overlooking the essential global illumination in images and the inherent noise sensitivity of event signals in real-world scenarios. To address these issues, we propose EIC-LIE, an event-illumination collaborative LIE framework. Concretely, we first design an Event-Illumination Collaborative Interaction (EICI) module, which contains two key processes: forward gathering, which gathers HDR features across varying lighting conditions, and backward injection, which provides complementary content for illumination and event representations.Next, we introduce an Illumination-aware Event Filter (IAEF) that dynamically reduces event noise based on brightness statistics derived from images. Additionally, we build a beam-splitter-based hybrid imaging system to collect high-quality event-image pairs with temporal synchronization from dynamic scenes, providing the first high-resolution, real-world event-based LIE dataset.Extensive experiments show that our EIC-LIE outperforms state-of-the-art methods on five real-world and synthetic datasets, significantly surpassing previous methods with improvements of up to 1.24dB in PSNR and 0.069 in SSIM. The code and dataset will be released.

High-quality imaging of dynamic scenes in extremely low-light conditions is highly challenging. Photon scarcity induces severe noise and texture loss, causing significant image degradation. Event cameras, featuring a high dynamic range (120 dB) and high sensitivity to motion, serve as powerful complements to conventional cameras by offering crucial cues for preserving subtle textures. However, most existing approaches emphasize texture recovery from events, while paying little attention to image noise or the intrinsic noise of events themselves, which ultimately hinders accurate pixel reconstruction under photon-starved conditions. In this work, we propose NEC-Diff, a novel diffusion-based event–RAW hybrid imaging framework that extracts reliable information from heavily noisy signals to reconstruct fine scene structures. The framework is driven by two key insights: (1) combining the linear light-response property of RAW images with the brightness-change nature of events to establish a physics-driven constraint for robust dual-modal denoising; and (2) dynamically estimating the SNR of both modalities based on denoising results to guide adaptive feature fusion, thereby injecting reliable cues into the diffusion process for high-fidelity visual reconstruction. Furthermore, we construct the REAL (Raw and Event Acquired in Low-light) dataset which provides 47,800 pixel-aligned low-light RAW images, events, and high-quality references under 0.001–0.8 lux illumination. Extensive experiments demonstrate the superiority of NEC-Diff under extreme darkness.

Small object detection (SOD) plays a vital role in applications such as anti-UAV tasks, yet conventional image-based methods struggle in high-speed scenarios due to the limited frame rate. Event cameras offer a promising alternative by capturing spatiotemporal event streams with microsecond-level temporal resolution. To address the inherent sparsity of small objects in event data, existing methods typically formulate the detection task as semantic segmentation on spatiotemporal point clouds to leverage long-term contextual information. However, these methods often fail to enforce effective spatiotemporal consistency constraints, resulting in fragmented object trajectories. To mitigate these problems, we propose a topology-constrained sparse convolutional network (SpTopoNet), which models the topological structure of moving object trajectories in event point clouds. Our network comprises two key components: a Topology Learning Module (TLM) that discriminates local structures to separate genuine targets from noise, and a Spatial Consistency Module (SCM) that captures long-range spatiotemporal dependencies to enhance trajectory continuity. Additionally, we introduce an event topology-aware loss function that leverages topological correlations to guide the network to maintain structural integrity of target event patterns.Experiments on the benchmark dataset demonstrate the superiority of our method in both detection performance and trajectory completeness.

Event cameras offer a high temporal resolution over traditional frame-based cameras, which makes them suitable for motion and structure estimation. However, it has been unclear how event-based 3D Gaussian Splatting (3DGS) approaches could leverage fine-grained temporal information of sparse events. This work proposes a framework to address the trade-off between accuracy and temporal resolution in the event-based 3DGS. Our key idea is to decouple the rendering into two branches: event-by-event geometry (depth) rendering and snapshot-based radiance (intensity) rendering, by using ray-tracing and the image of warped events. The extensive evaluation shows that our method achieves the state-of-the-art performance on the real-world datasets and competitive performance on the synthetic datasets. Also, the proposed method works without prior information (e.g., pretrained image reconstruction models) or COLMAP-based initialization, is more flexible in the event accumulation size, and achieves sharp reconstruction on scene edges. We hope that this work deepens our understanding of the sparse nature of events for 3D reconstruction. We will release the code upon acceptance.

Event cameras sense the world through asynchronous brightness changes with microsecond latency and high dynamic range, offering motion fidelity far beyond frame-based sensors and capturing temporal structure that conventional exposures often miss. These properties make events a powerful complement to RGB in autonomous driving, especially under blur, glare, and rapid motion where frame-based perception can become unreliable. However, existing event-aware vision–language models remain limited to generic perception and do not reveal how event sensing contributes to reasoning and decision-making across the full driving loop. We present EventDrive, a large-scale benchmark and model suite that unifies event streams, RGB frames, and language supervision across four core dimensions: Perception, Understanding, Prediction, and Planning, covering captions, structured QA, grounding, motion-state recognition, trajectory forecasting, and planning tasks. Building on this foundation, EventDrive-VLM introduces a multi-horizon event pyramid and a modality-routing mixture-of-experts to adaptively encode and fuse asynchronous and frame-based information for downstream reasoning. Comprehensive evaluation across diverse tasks shows that event streams provide substantial gains in temporal precision, motion awareness, and robustness, bringing event sensing into the center of driving intelligence.

Gait recognition enables non-intrusive, privacy-preserving identification but suffers in uncontrolled environments due to illumination and motion sensitivity in conventional cameras. In this work, we explore gait recognition using event cameras, which offer microsecond temporal resolution and high dynamic range, naturally capturing robust dynamic cues and suppressing static noise. Existing event-based approaches typically aggregate event streams into event images over long time windows, thereby discarding fine-grained motion dynamics critical for gait recognition.Therefore, we propose \textbf{EventGait}, an end-to-end dual-stream framework that separately models motion and shape while preserving the advantages of events. Our dynamic stream leverages a Mixture of Spiking Experts (MoSE) with diverse neuron constants for robust dynamic perception across complex motion and illumination scenes, while the static stream learns dense shape representations via Cross-modal Structural Alignment (CroSA) with large vision foundation models. To address the absence of large-scale event-based gait datasets, we introduce a synthesis pipeline and release two new benchmarks: SUSTech1K-E and CCGR-Mini-E. Extensive experiments have shown that event-based gait recognition not only achieves results comparable to camera-based gait recognition under normal conditions but also significantly outperforms it in low-light scenarios. Our approach sets a new state of the art on both synthesized and real-world event-based gait benchmarks, highlighting the robustness and potential of event-driven gait analysis. The code and datasets will be released.

Recent Vision-Language-Action (VLA) models reformulate vision-language models by tuning them with millions of robotic demonstrations. While they perform well when fine-tuned for a single embodiment or task family, extending them to multi-skill settings remains challenging: directly merging VLA experts trained on different tasks results in near-zero success rates. This raises a fundamental question: what prevents VLAs from mastering multiple skills within one model? With an empirical decomposition of learnable parameters during VLA fine-tuning, we identify two key sources of non-mergeability:(1) Finetuning drives LoRA adapters in the VLM backbone toward divergent, task-specific directions beyond the capacity of existing merging methods to unify.(2) Action experts develop inter-block dependencies through self-attention feedback, causing task information to spread across layers and preventing modular recombination.To address these challenges, we present MergeVLA, a merging-oriented VLA architecture that preserves mergeability by design.MergeVLA introduces sparsely activated LoRA adapters via task masks to retain consistent parameters and reduce irreconcilable conflicts in the VLM.Its action expert replaces self-attention with cross-attention-only blocks to keep specialization localized and composable.When the task is unknown, it uses a test-time task router to adaptively select the appropriate task mask and expert head from the initial observation, enabling unsupervised task inference.Across LIBERO, LIBERO-Plus, RoboTwin, and multi-task experiments on the real SO101 robotic arm, MergeVLA achieves performance comparable to or even exceeding individually finetuned experts, demonstrating robust generalization across tasks, embodiments, and environments.The source code can be found in the supplementary material for reference.

This paper proposes the Continual Dual-Critic with Cross-Attention (CD-CCA) framework for visual reinforcement learning to address the plasticity-stability conflict. Our method introduces continual learning techniques into the visual RL architecture, constructing two complementary critics using Continual Backpropagation (CBP) and Elastic Weight Consolidation (EWC) -- one for maintaining representational plasticity for rapid environmental adaptation, and the other for preserving knowledge stability to prevent catastrophic forgetting. Furthermore, we design a cross-attention based fusion mechanism that balances the value estimates from the dual critics according to observation characteristics. Experimental results on DeepMind Control and CARLA benchmarks show that CD-CCA effective mitigates issues of representation drift and policy degradation. Compared to existing visual RL methods, our approach exhibits enhanced robustness and adaptability in non-stationary environments and long-horizon decision-making tasks, providing a new architectural paradigm for the advancement of continual reinforcement learning.

Real-world data collection for embodied agents remains costly and unsafe, calling for scalable, realistic, and simulator-ready 3D environments. However, existing scene-generation systems often rely on rule-based or task-specific pipelines, yielding artifacts and physically invalid scenes. We present SAGE, an agentic framework that, given a user-specified embodied task (e.g., “pick up a bowl and place it on the table”), understands the intent and automatically generates simulation-ready environments at scale. The agent couples multiple generators for layout and object composition with critics that evaluate semantic plausibility, visual realism, and physical stability. Through iterative reasoning and adaptive tool selection, it self-refines the scenes until meeting user intent and physical validity. The resulting environments are realistic, diverse, and directly deployable in modern simulators for policy training. Policies trained purely on this data exhibit clear scaling trends and generalize to unseen objects and layouts, demonstrating the promise of simulation-driven scaling for Embodied AI. We will release both 3D scene and action generation code to foster further research.

Audio-visual navigation enables embodied agents to navigate toward sound-emitting targets by leveraging both auditory and visual cues. However, most existing approaches depend on precomputed room impulse responses (RIRs) for binaural audio rendering, restricting agents to discrete grid positions and leading to spatially discontinuous observations. To establish a more realistic setting, we introduce Semantic Audio-Visual Navigation in Continuous Environments (SAVN-CE), where agents can move freely in 3D spaces and perceive temporally and spatially coherent audio-visual streams. In this setting, targets may intermittently become silent or stop emitting sound entirely, causing agents to lose goal information. To tackle this challenge, we propose MAGNet, a multimodal transformer-based model that jointly encodes spatial and semantic goal representations and integrates historical context with self-motion cues to enable memory-augmented goal reasoning. Comprehensive experiments demonstrate that MAGNet significantly outperforms state-of-the-art methods, achieving up to a 12.1\% absolute improvement in success rate. These results also highlight its robustness to short-duration sounds and long-distance navigation scenarios.

Vision-Language-Action (VLA) models built upon Chain-of-Thought (CoT) have achieved remarkable success in advancing general-purpose robotic agents, owing to its significant perceptual comprehension. Recently, since text-only CoT struggles to adequately capture scene details in complex spatial environments, a highly promising strategy involves leveraging visual priors to guide robotic action generation. Nevertheless, these strategies face two inherent challenges: (i) a modality gap between visual observations and low-level actions, and (ii) unstable training due to competing objectives between visual prediction and action generation. To address these challenges, we propose a Vision-Integrated Trajectory Alignment (VITA) framework that learns a shared discrete latent space for vision and action, enabling joint modeling of perception and motor control. VITA introduces a implicit visual CoT: autoregressively generated tokens is simultaneously decoded into future frames predictions and robot actions, thereby internalizing visual dynamics as an inductive bias for motion planning. Extensive experiments on simulated and real-world environments demonstrate state-of-the-art performance. VITA improves 14.5\%, 9.6\% and 12.1\% over existing baselines on CALVIN, LIBERO and SimplerEnv. Furthermore, VITA attains an average success rate of 80.5\% across six real-world tasks, demonstrating its potential as a generalist robotic manipulation model.

Vision-Language-Action Models~(VLAs) have demonstrated significant promise in generalizing to complex, long-horizon robotic manipulation tasks. However, their performance remains brittle, as they are typically trained on trajectory-monotonic, failure-free demonstrations. This reliance on "perfect" data leaves them unable to recover from common execution errors, such as a missed grasp, a dropped object, or an unexpected collision. In this paper, we propose FLARE, a novel framework that endows VLAs with robust error recovery capabilities through a "Retry" and "Reset" paradigm. First, we introduce a "Retry" mechanism by injecting perturbation and bridging segments that decouple robot pose from environment state into demonstrations, enabling the policy to autonomously handle execution deviations. Second, to address critical, state-breaking (OOD) failures, we introduce a "Reset" pipeline. We leverage an MLLM for offline failure analysis to automatically identify OOD states from execution videos. This analysis enables the efficient, targeted collection of a small library of object-centric "Reset" skills, which are trained to restore the environment to a task-valid state. Our full framework integrates these learned policies. At inference, an online MLLM monitor arbitrates between task execution and "Reset" skills. Experiments on challenging, contact-rich manipulation tasks show our approach significantly improves task success and robustness.

Recent advances in Multimodal Large Language Models (MLLMs) have led to promising progress in web agents. However, existing web agents often rely on handcrafted execution pipelines or expensive expert trajectories, limiting their adaptability to complex, dynamic environments. To address these challenges, we propose SCALE (Self-Cognitive-Aware Learning and Exploration), which leverages three advertise roles——selector, predictor, and judger to autonomously discover their limitations and expand its cognitive boundaries through the environment exploration. Moreover, we propose SCALE-Hop, a graph exploration strategy that facilitates global planning and helps agents avoid local exploration traps. To further support learning, we construct SCALE-20k, a large-scale dataset collected from 19 real-world websites, containing diverse task types and structured demonstrations generated from SCALE’s exploration traces. Experimental results show that our approach significantly improves the performance and generalization of multiple MLLMs in various web environments. Our framework offers a scalable and generalizable solution for building truly autonomous and adaptive web agents.

The primary obstacle for applying reinforcement learning (RL) to real-world robotics is the design of effective reward functions. While recently learning-based Process Reward Models (PRMs) are a promising direction, they are often hindered by two fundamental limitations: their reward models lack step-aware understanding and rely on single-view perception, leading to unreliable assessments of fine-grained manipulation progress; and their reward shaping procedures are theoretically unsound, often inducing a semantic trap that misguides policy optimization.To address these, we introduce Robo-Dopamine, a novel reward modeling method for learning a general-purpose, step-aware process reward model from multi-view inputs. At its core is our General Reward Model (GRM), trained on a vast 3,400+ hour dataset, which leverages Step-wise Reward Discretization for structural understanding and Multi-Perspective Reward Fusion to overcome perceptual limitations. Building upon Robo-Dopamine, we propose Dopamine-RL, a robust policy learning framework that employs a theoretically-sound Policy-Invariant Reward Shaping method, which enables the agent to leverage dense rewards for efficient self-improvement without altering the optimal policy, thereby fundamentally avoiding the semantic trap.Extensive experiments across 10 simulated and 8 real-world tasks validate our approach. {GRM achieves state-of-the-art accuracy in reward assessment}, and {Dopamine-RL built on GRM significantly improves policy learning efficiency}.For instance, after {GRM} is adapted to a new task in a one-shot manner from a single expert trajectory, the resulting reward model enables Dopamine-RL to improve the policy from near-zero to 95\% success with only 150 online rollouts (approximately 1 hour of real robot interaction), while retaining strong generalization across tasks.

Recent generative models allow robots to generate future visual outcomes for action guidance, yet most still address imagination and control independently, resulting in visually coherent rollouts but physically inconsistent behaviors. While structural priors enhance spatial grounding, these methods remain visually correlation-driven rather than causally informed, overlooking the bidirectional coupling between robot actions and the evolving environment. We formalize the coupling as interaction dynamics, which specify where environmental changes occur and how actions cause them. Based on this formulation, we introduce DynBridge, an end-to-end framework that unifies imagination and control through the shared dynamics representation. Specifically, DynBridge realizes this via three components: (1) an Interaction Dynamics Generator that forecasts interaction dynamics via joint trajectory generation and action prediction; (2) an Action-Conditioned Dynamics Aggregator that integrates dynamics under control signals; and (3) a Dynamics-Guided Action Predictor that leverages the aggregated dynamics to produce executable, context-aware actions. Results demonstrate that DynBridge consistently outperforms prior methods on simulated and real-world benchmarks without external pretraining.

Long-horizon, open-world robotic manipulation is increasingly important for real-world deployment, requiring spatial disambiguation in complex layouts and temporal resilience under dynamic interaction. However, existing end-to-end and hierarchical Vision–Language–Action (VLA) policies often rely on text-only cues while keeping plan intent latent, which undermines \textit{referential grounding} in cluttered or underspecified scenes, impedes effective \textit{task decomposition} of long-horizon goals with close-loop interaction, and limits \textit{causal explanation} by obscuring the rationale behind action choices. To address these issues, we first introduce \textbf{Visual Sketch}, a lightweight visual intermediate that renders points, boxes, arrows, and typed relations on the robot’s current views to externalize spatial intent, bind language to scene geometry, and provide a human-verifiable bridge between high-level reasoning and low-level control. Building on \textit{Visual Sketch}, we present \textbf{Action-Sketcher}, a VLA framework that operates in a cyclic \textit{See $\rightarrow$ Think $\rightarrow$ Sketch $\rightarrow$ Act} workflow coordinated by adaptive token-gated strategy for reasoning triggers, sketch revision, and action issuance, thereby supporting reactive corrections and human interaction while preserving real-time action prediction. To enable scalable training and evaluation, we curate a 2.3M-sample corpus with interleaved images, text, \textit{Visual Sketch} supervision, and action sequences, and train \textit{Action-Sketcher} with a multi-stage curriculum recipe that combines interleaved sequence alignment for modality unification, language-to-sketch consistency for precise linguistic grounding, and imitation learning augmented with sketch-to-action reinforcement for robustness. Experiments on cluttered tabletops and multi-object tasks, in simulation and on real robots, show improved long-horizon success, stronger robustness to dynamic scene changes, and enhanced interpretability via editable sketches and step-wise plans.

Humans rely on the synergistic control of head (cephalomotor) and eye (oculomotor) to efficiently search for visual information in 360°. However, prior approaches to visual search are limited to a static image, neglecting the physical embodiment and its interaction with the 3D world. How can we develop embodied visual search agents as efficient as humans while bypassing the constraints imposed by real-world hardware? To this end, we propose humanoid visual search where a humanoid agent actively rotates its head to search for objects or paths in an immersive world represented by a 360° panoramic image. To study visual search in visually-crowded real-world scenarios, we build H* Bench, a new benchmark that moves beyond household scenes to challenging in-the-wild scenes that necessitate advanced visual-spatial reasoning capabilities, such as transportation hubs, large-scale retail spaces, urban streets, and public institutions. Our experiments first reveal that even top-tier proprietary models falter, achieving only ~30% success in object and path search. We then use post-training techniques to enhance the open-source Qwen2.5-VL, increasing its success rate by over threefold for both object search (14.83% → 47.38%) and path search (6.44% → 24.94%). Notably, the lower ceiling of path search reveals its inherent difficulty, which we attribute to the demand for sophisticated spatial commonsense. Our results not only show a promising path forward but also quantify the immense challenge that remains in building MLLM agents that can be seamlessly integrated into everyday human life.

Discriminative and generative vision models excel in their respective domains but remain semantically misaligned, hindering progress toward unified visual learning. We introduce LEASE (LEArning from SEmantic Dictionaries), a self-supervised framework that bridges this gap using a paired generative–discriminative codebook design. LEASE operates entirely in a discrete token space produced through a one-time precomputation step, enabling efficient training without data augmentations, teacher models, or online tokenizers.LEASE integrates two complementary objectives: a masked token reconstruction loss that captures fine-grained generative detail, and a codebook contrast loss that aligns encoder features with discriminative semantics via adaptive centroid weighting. This dual supervision yields a unified latent space that supports both high-quality generation and strong representation learning.On ImageNet-1K, LEASE achieves state-of-the-art unified performance, outperforming prior VQGAN-based methods such as MAGE and Sorcen across linear probing (up to +1.7%), unconditional generation (-1.26 FID and +10.19 IS w.r.t MAGE), few-shot learning (+0.56% on average against Sorcen), transfer (+0.75% average improvement against MAGE and Sorcen), and robustness benchmarks (+5.86% and +4.25% average improvement against MAGE and Sorcen, respectively). It also competes favorably with domain-specialized contrastive and generative models while surpassing previous MIM methods. The unsupervised LEASE model can also be extended to conditional generation by building upon its learned representations, proving competitive with specialized baselines.Overall, LEASE provides an efficient and effective step toward general-purpose vision models that jointly understand and generate visual content. Code will be released upon acceptance.

We propose MagicQuill V2, a novel framework that introduces a layered composition paradigm to generative image editing, bridging the gap between the semantic power of modern diffusion models and the granular control of traditional graphics software. While state-of-the-art diffusion transformers excel at holistic generation, their use of singular, monolithic prompts fails to disentangle distinct user intentions for content, position, and style. To overcome this limitation, our method deconstructs creative intent into a stack of independently controllable visual cues: a content layer for what to create, a spatial layer for where to place it, a structural layer for how it is shaped, and a color layer for its palette. Our technical contributions include a specialized data generation pipeline for context-aware content integration, a unified control module to process all visual cues, and a fine-tuned spatial branch for precise local editing, including object removal. Extensive experiments validate that this layered approach effectively resolves the user intention gap, granting creators direct, intuitive, and powerful control over the generative process.

Disentangling visual layers in real-world images is a persistent challenge in vision and graphics, as such layers often involve non-linear and globally coupled interactions, including shading, reflection, and perspective distortion. In this work, we present an in-context image decomposition framework that leverages large diffusion foundation models for layered separation. We focus on the challenging case of logo-object decomposition, where the goal is to disentangle a logo from the surface on which it appears while faithfully preserving both layers. Our method fine-tunes a pretrained diffusion model via lightweight LoRA adaptation and introduces a cycle-consistent tuning strategy that jointly trains decomposition and composition models, enforcing reconstruction consistency between decomposed and recomposed images. This bidirectional supervision substantially enhances robustness in cases where the layers exhibit complex interactions. Furthermore, we introduce a progressive self-improving process, which iteratively augments the training set with high-quality model-generated examples to refine performance. Extensive experiments demonstrate that our approach achieves accurate and coherent decompositions and also generalizes effectively across other decomposition types, suggesting its potential as a unified framework for layered image decomposition.

The integration of visual understanding and generation into unified multimodal models represents a significant stride toward general-purpose AI. However, a fundamental question remains unanswered by existing benchmarks: does this architectural unification actually enable synergetic interaction between the constituent capabilities? Existing evaluation paradigms, which primarily assess understanding and generation in isolation, are insufficient for determining whether a unified model can leverage its understanding to enhance its generation, or use generative simulation to facilitate deeper comprehension. To address this critical gap, we introduce RealUnify, a benchmark specifically designed to evaluate bidirectional capability synergy. RealUnify comprises 1,000 meticulously human-annotated instances spanning 10 categories and 32 subtasks. It is structured around two core axes: 1) Understanding Enhances Generation, which requires reasoning (e.g., commonsense, logic) to guide image generation, and 2) Generation Enhances Understanding, which necessitates mental simulation or reconstruction (e.g., of transformed or disordered visual inputs) to solve reasoning tasks. A key contribution is our dual-evaluation protocol, which combines direct end-to-end assessment with a diagnostic stepwise evaluation that decomposes tasks into distinct understanding and generation phases. This protocol allows us to precisely discern whether performance bottlenecks stem from deficiencies in core abilities or from a failure to integrate them. Through large-scale evaluations of 12 leading unified models and 6 specialized baselines, we find that current unified models still struggle to achieve effective synergy, indicating that architectural unification alone is insufficient. These results highlight the need for new training strategies and inductive biases to fully unlock the potential of unified modeling.

Text-to-image (T2I) models are increasingly used for synthetic dataset generation, but generating synthetic training data to improve fine-grained classification performance remains challenging. Fine-tuning the T2I model with a few real examples can help generate more appropriate synthetic training data; however, this fine-tuning may also introduce overfitting and reduce diversity in the generated samples. We propose a fine-tuning strategy BOB (Beyond OBjects) for mitigating these concerns. Given a small set of real examples, we first describe them using class-agnostic attributes such as scene background and object pose. We then explicitly condition on these attributes during fine-tuning of the T2I model and marginalize them out during generation. This design mitigates overfitting, thus preserving the T2I model’s generative prior and reducing estimation errors, and further minimizes unintended inter-class associations. Extensive experiments across multiple T2I models, backbones, and datasets demonstrate state-of-the-art performance in low-shot fine-grained classification when augmented with synthetic data. Concretely, BOB outperforms DataDream by 7.4% on the Aircraft dataset (from 50.0% to 57.4% when fine-tuning a CLIP classifier with 5 real images augmented with 100 synthetic images). Additionally, in three of the four datasets, the fine-tuning downstream models with synthetic data generated from BOB and five real images achieves better performance than fine-tuning with 10 real images. Collectively, BOB outperforms prior art in 18 of 24 experimental settings, with over 2% accuracy improvements in 14 of these settings.

Diffusion-based text-to-image (T2I) models have enabled remarkable generative capabilities, yet precise text-based image editing that preserves the original’s structural and perceptual fidelity remains non-trivial. Existing approaches either rely on retraining with large bespoke datasets, incurring significant computational and curation costs, or adopt lightweight fine-tuning strategies that still require optimization and often fail in fine-grained or semantically complex edits.We propose NEAF (Natural image Editing with Attention Fusion), a novel zero-shot, universal tuning-free framework for arbitrary T2I models, obviating the need for dataset curation or retraining. NEAF introduces a lightweight, learnable XA-Conductor module that dynamically identifies salient cross-attention contributions pertinent to the edit. This module optimizes a weight vector to orchestrate an adaptive fusion of cross-attention maps derived from the source, edited, and reconstruction branches. This triadic-feedback optimization strategy ensures the precise instantiation of user directives while rigorously preserving the fidelity of quiescent regions.Extensive experiments validate NEAF as a flexible and general framework that consistently surpasses existing methods across diverse editing tasks, demonstrating particular dominance in complex, non-rigid editing scenarios where other approaches falter.

Generative data augmentation techniques open new avenues for improving image recognition models. The core of image recognition lies in accurately capturing the ontological features of the subject. However, existing methods often treat the image as a whole during augmentation, ignoring the uneven semantic distribution between foreground and background. This can lead to semantic shifts in generated samples, weakening the model’s ability to represent the subject’s ontology. In human perception, category recognition typically relies on the stable essence of the subject while tolerating variations in background and environment. Inspired by this human perceptual mechanism of “stable subjects, diverse backgrounds, and overall coherence,” we propose OntoAug, a data augmentation framework based on the distinction between ontology and environment that redefines the boundary of ontology-oriented enhancement. OntoAug explicitly separates the foreground subject and background context, guiding diffusion models through structured layout control to generate samples with consistent subjects and diverse backgrounds. Experiments show that OntoAug significantly improves performance in image classification, few-shot learning, weakly supervised object localization (WSOL), and large vision-language model (LVLM) reasoning, demonstrating its advantages in semantic fidelity and sample diversity. It offers a new direction for building visual systems more aligned with human perception. Code will be available.

Radiance field methods (e.g.~3D Gaussian Splatting) have emerged as a powerful paradigm for novel view synthesis, yet their appearance modeling often relies on Spherical Harmonics (SH), which impose fundamental limitations.SH struggle with high-frequency signals, exhibit Gibbs ringing artifacts, and critically fail to capture specular reflections -- a key component of realistic rendering. While alternatives like Spherical Gaussians offer improvements, they introduce significant optimization complexity.We propose Spherical Voronoi (SV) as a unified framework for appearance representation in 3D Gaussian Splatting.SV partitions the directional domain into learnable regions with smooth boundaries, providing an intuitive and stable parameterization for view-dependent effects. For diffuse appearance, SV achieves competitive results while maintaining simpler optimization compared to existing alternatives. For reflections -- where SH fundamentally fail -- we leverage SV as learnable reflection probes, taking reflected directions as input following principles from traditional graphics. This formulation achieves state-of-the-art results across both synthetic and real-world datasets, demonstrating that SV offers a principled, efficient, and general solution for appearance modeling in explicit 3D representations.

This paper addresses the problem of dynamic scene surface reconstruction using Gaussian Splatting (GS), aiming to recover temporally consistent geometry. While existing GS-based dynamic surface reconstruction methods can yield superior reconstruction, they are typically limited to either a single object or objects with only small deformations, struggling to maintain temporally consistent surface reconstruction of large deformations over time. We propose ``4DSurf'', a novel and unified framework for generic dynamic surface reconstruction that does not require specifying the number or types of objects in the scene, can handle large surface deformations and temporal inconsistency in reconstruction. The key innovation of our framework is the introduction of Gaussian deformations induced Signed Distance Function Flow Regularization that constrains the motion of Gaussians to align with the evolving surface. To handle large deformations, we introduce an Overlapping Segment Partitioning strategy that divides the sequence into overlapping segments with small deformations and incrementally passes geometric information across segments through the shared overlapping timestep. Experiments on two challenging dynamic scene datasets, Hi4D and CMU Panoptic, demonstrate that our method outperforms state-of-the-art surface reconstruction methods by 49\% and 19\% in Chamfer distance, respectively, and achieves superior temporal consistency under sparse-view settings.

Robust 3D representation learning forms the perceptual foundation of spatial intelligence, enabling downstream tasks in scene understanding and embodied AI. However, learning such representations directly from unposed multi-view images remains challenging. Recent self-supervised methods attempt to unify geometry, appearance, and semantics in a feed-forward manner, but they often suffer from weak geometry induction, limited appearance detail, and inconsistencies between geometry and semantics.We introduce $\textbf{\textit{UniSplat}}$, a feed-forward framework designed to address these limitations through three complementary components. First, we propose a $\textit{dual-masking strategy}$ that strengthens geometry induction in the encoder. By masking both encoder and decoder tokens, and targeting decoder masks toward geometry-rich regions, the model is forced to infer structural information from incomplete visual cues, yielding geometry-aware representations even under unposed inputs.Second, we develop a $\textit{coarse-to-fine Gaussian splatting strategy}$ that reduces appearance-semantics inconsistencies by progressively refining the radiance field.Finally, to enforce geometric–semantic consistency, we introduce a $\textit{pose-conditioned recalibration mechanism}$ that interrelates the outputs of multiple heads by reprojecting predicted 3D point and semantic maps into the image plane using estimated camera parameters, and aligning them with corresponding RGB and semantic predictions to ensure cross-task consistency and resolving geometry–semantic mismatches. Together, these components yield unified 3D representations that are robust to unposed, sparse-view inputs and generalize across diverse tasks, laying a perceptual foundation for spatial intelligence.

Novel view synthesis has been significantly advanced by NeRFs and 3D Gaussian Splatting (3DGS), which require ordering volumetric samples or primitives for correct color blending. While the recent Gaussian-Enhanced Surfels (GES) enable high-performance, sort-free rendering, they suffer from aliasing artifacts and suboptimal reconstruction. To address these limitations, we propose DP-GES, a novel representation that augments opaque surfels with semi-transparent boundaries and leverages Depth Peeling to establish accurate per-pixel ordering. This design enables sort-free Gaussian splatting with correct transmittance modulation, effectively eliminating aliasing and popping artifacts while facilitating a fully differentiable joint optimization. Extensive experiments demonstrate that our method achieves superior reconstruction quality and compares favorably against state-of-the-art techniques across a wide range of scenes.

We introduce Intrinsic Image Fusion, a method that reconstructs high-quality physically based materials from multi-view images.Material reconstruction is highly underconstrained and typically relies on analysis-by-synthesis, which requires expensive and noisy path tracing. To better constrain the optimization, we incorporate single-view priors into the reconstruction process. We leverage a diffusion-based material estimator that produces multiple, but often inconsistent, candidate decompositions per view.To reduce the inconsistency, we fit an explicit low-dimensional parametric function to the predictions.We then propose a robust optimization framework using soft per-view prediction selection together with confidence-based soft multi-view inlier set to fuse the most consistent predictions of the most confident views into a consistent parametric material space. Finally, we use inverse path tracing to optimize for the low-dimensional parameters. Our results outperform state-of-the-art methods in material disentanglement on both synthetic and real scenes, producing sharp and clean reconstructions suitable for high-quality relighting.

Volumetric videos offer immersive 4D experiences, but remain difficult to reconstruct, store, and stream at scale. Existing Gaussian Splatting based methods achieve high-quality reconstruction but break down on long sequences, temporal inconsistency, and fail under large motions and disocclusions. Moreover, their outputs are typically incompatible with conventional video coding pipelines, preventing practical applications. We introduce PackUV, a novel 4D Gaussian representation that maps all Gaussian attributes into a sequence of structured, multi-scale UV atlas, enabling compact, image-native storage. To fit this representation from multi-view videos, we propose PackUV-GS, a temporally consistent fitting method that directly optimizes Gaussian parameters in the UV domain. A flow-guided Gaussian labeling and video keyframing module identifies dynamic Gaussians, stabilizes static regions, and preserves temporal coherence even under large motions and disocclusions. The resulting UV atlas format is the first unified volumetric video representation compatible with standard video codecs (e.g., HEVC, FFV1) without quality loss, enabling efficient streaming within existing multimedia infrastructure. To evaluate long-duration volumetric capture, we present PackUV-2B, the largest multi-view 4D dataset to date, featuring more than 50 synchronized 360 cameras, substantial motion, and frequent disocclusions across 100 sequences and 2B (billion) frames. Extensive experiments demonstrate that our method surpasses existing baselines in rendering fidelity while scaling to sequences up to 30 minutes with consistent quality.

3D reconstruction of transparent objects from multiple views has been a long-standing challenge. In contrast to opaque objects, transparent objects exhibit complex refraction that causes serious image distortion, resulting in a highly ill-posed problem. Existing reconstruction methods commonly depend on special capture devices or controlled environments, which provide more priors and simplify the modeling of refraction. More importantly, these methods lack the capability for reconstruction of mixed transparent and opaque objects, being confined to transparent or opaque materials. To address these challenges, we propose Opti-NeuS, a novel method for reconstructing transparent and opaque objects without controlled environments or additional input. Opti-NeuS incorporates a novel IoRNetwork to obtain spatially-varying IoR for tracing the refractive ray paths, which can finally model refractive visual distortion. To deal with dual-layered transparent and opaque objects, we devise a two-stage hierarchical reconstruction strategy that decouples outer and inner geometry, combined with alpha-blending for transparency-aware surface separation. Experiments show that Opti-NeuS achieves practical utility and effectiveness and outperforms prior works.

We introduce PhysGaia, a novel physics-aware dataset specifically designed for Dynamic Novel View Synthesis (DyNVS), encompassing both structured objects and unstructured physical phenomena.Unlike existing datasets that primarily focus on photorealistic reconstruction, PhysGaia is created to actively support physics-aware dynamic scene modeling.Our dataset provides complex dynamic scenarios with rich interactions among multiple objects, where they realistically collide with each other and exchange forces.Furthermore, it contains a diverse range of physical materials, such as liquid, gas, textile, and rheological substances, which moves beyond the rigid bodies prevalent in existing datasets.All scenes in PhysGaia are faithfully generated to strictly adhere to physical laws, leveraging carefully selected material-specific physics solvers. To enable quantitative evaluation of physical modeling, our dataset provides essential ground-truth information, including 3D particle trajectories and physics parameters, e.g., viscosity.To facilitate research adoption, we also provide essential integration pipelines for using recent 4D Gaussian Splatting models with our dataset and report their results.By addressing the critical lack of datasets for physics-aware modeling, PhysGaia will significantly advance research in dynamic view synthesis, physics-based scene understanding, and deep learning models integrated with physical simulation-ultimately enabling more faithful reconstruction and interpretation of complex dynamic scenes.

Manual modeling of material parameters and 3D geometry is a time consuming yet essential task in the gaming and film industries. While recent advances in 3D reconstruction have enabled accurate approximations of scene geometry and appearance, these methods often fall short in relighting scenarios due to the lack of precise, spatially varying material parameters. At the same time, diffusion models operating on 2D images have shown strong performance in predicting physically based rendering (PBR) properties such as albedo, roughness, and metallicity. However, transferring these 2D material maps onto reconstructed 3D geometry remains a significant challenge. We propose a framework for fusing 2D material data into 3D geometry using a combination of novel learning-based and projection-based approaches. We begin by reconstructing scene geometry via Gaussian Splatting. From the input images, a diffusion model generates 2D maps for albedo, roughness, and metallic parameters. Any existing diffusion model that can convert images or videos to PBR materials can be applied. The predictions are further integrated into the 3D representation either by optimizing an image-based loss or by directly projecting the material parameters onto the Gaussians using Gaussian ray tracing. To enhance fine-scale accuracy and multi-view consistency, we further introduce a light-weight neural refinement step (Neural Merger), which takes ray-traced material features as input and produces detailed adjustments. Our results demonstrate that the proposed methods outperform existing techniques in both quantitative metrics and perceived visual realism. This enables more accurate, relightable, and photorealistic renderings from reconstructed scenes, significantly improving the realism and efficiency of asset creation workflows in content production pipelines.

Object motion blur in static scenes is spatially heterogeneous, differing from conventional deblurring problems yet frequently occurring in real handheld capture scenarios. Existing datasets either rely on costly beam-splitting capture with residual misalignment or employ synthetic blur that fails to model the continuous photon-integration process during exposure. To overcome these limitations, we introduce OMoBlur, a physically grounded dataset that emulates realistic exposure integration via programmable sensor control, ensuring close alignment between synthetic and real blur distributions. OMoBlur provides 20,000 blur–sharp–mask pairs covering diverse object motion types. Leveraging this dataset, we further propose OMDNet, an object-motion-aware deblurring network that integrates a Motion–Appearance Extract Block, a Flow-Guided Gate Predictor, and an Adaptive Gated Fusion mechanism. This design enables the network to selectively restore blurred regions while preserving static backgrounds, without requiring pixel-accurate mask annotations. Extensive experiments demonstrate that OMoBlur’s physically faithful data collection and large-scale diversity significantly enhance the network’s generalization to real-world motion blur, establishing OMoBlur and OMDNet as a robust benchmark and practical solution for local motion deblurring. The dataset and code will be publicly released.

Existing Image Restoration (IR) studies typically focus on task-specific or universal modes individually, relying on the mode selection of users and lacking cooperation between multiple task-specific/universal restoration modes. This leads to insufficient interaction for unprofessional users and limits their restoration capability for complicated real-world applications. In this work, we present HybridAgent, intending to incorporate multiple restoration modes into a unified image restoration model and achieve intelligent and efficient user interaction through our proposed hybrid agents. Concretely, we propose the hybrid rule of fast, slow, and feedback restoration agents. Here, the slow restoration agent optimizes the powerful multimodal large language model (MLLM) with our proposed instruction-tuning dataset to identify degradations within images with ambiguous user prompts and invokes proper restoration tools accordingly. The fast restoration agent is designed based on a lightweight large language model (LLM) via in-context learning to understand the user prompts with simple and clear requirements, which can obviate the unnecessary time/resource costs of MLLM. Moreover, we introduce the mixed distortion removal mode for our HybridAgents, which is crucial but not concerned in previous agent-based works. It can effectively prevent the error propagation of step-by-step image restoration and largely improve the efficiency of the agent system. We validate the effectiveness of HybridAgent with both synthetic and real-world IR tasks.

Zero-shot image denoising has gained prominence in recent years, as it inherently relies on the intrinsic priors of images rather than learning from external data. Nevertheless, most existing methods either fail to fully exploit global priors, or do not properly preserve the fine-grained details governed by local priors. In this work, we propose a novel framework of pseudo sample generation for zero-shot denoising guided by local and global image priors. Specifically, we propose a well-crafted down-sampler based on gradient merging and grouping within a small window to generate down-sampled samples by exploiting spatial locality. Meanwhile, a global random sampler conditioned on a Gaussian distribution is devised to incorporate the nonlocal self-similarity of natural images. These two samplers build a new paradigm of pseudo sample generation powered by both local and global priors, which is termed as Zero-Shot Hybrid Prior-guided Denoising (ZS-HPD). Considering that noise is more likely to affect high-frequency details, we also present a simple yet effective loss that works in the Fourier domain and applies discriminative weights to distinct spectral bands. Numerous experiments on benchmark datasets have demonstrated the superiority of our ZS-HPD over existing advanced methods.

Zero-shot image restoration provides a flexible way to handle diverse degradations without task-specific training. However, existing methods typically rely on stacked layers or pre-trained features to enhance degradation expression, while overlooking physically consistent priors. The insufficient degradation prompts impose the heavy training burden and high sampling costs during zero-shot diffusion. Moreover, the fixed inference trajectory often collapses to suboptimal solutions under complex corruptions. We observe that heterogeneous degradations can be reparameterized into a minimal set of physically coherent parameters for compact representation. Based on this insight, we first propose a unified physical zero-shot image restoration (UP-ZeroIR) framework that explicitly models heterogeneous degradations into a homogeneous all-in-one distribution. The distribution can be optimized directly in the latent space, enabling principled solution exploration and effective prompt adaptation. Besides, we introduce a dynamic quality-refinement strategy that adaptively adjusts the diffusion trajectory for robust globally optimal convergence. Extensive experiments demonstrate that our method achieves state-of-the-art performance across both single and mixed degradations. The code will be publicly available soon.

Self-supervised real-world image denoising remains a fundamental challenge, arising from the antagonistic trade-off between decorrelating spatially structured noise and preserving high-frequency details. Existing blind-spot network (BSN) methods rely on pixel-shuffle downsampling (PD) to decorrelate noise, but aggressive downsampling fragments fine structures, while milder downsampling fails to remove correlated noise. To address this, we introduce Next-Scale Prediction (NSP), a novel self-supervised paradigm that decouples noise decorrelation from detail preservation. NSP constructs cross-scale training pairs, where BSN takes low-resolution, fully decorrelated sub-images as input to predict high-resolution targets that retain fine details. As a by-product, NSP naturally supports super-resolution of noisy images without retraining or modification. Extensive experiments demonstrate that NSP achieves state-of-the-art self-supervised denoising performance on real-world benchmarks, significantly alleviating the long-standing conflict between noise decorrelation and detail preservation.

Shadow removal under diverse lighting conditions requires disentangling illumination from intrinsic reflectance—a challenge compounded when physical priors are not properly aligned. We propose PhaSR (Physically Aligned Shadow Removal), addressing this through dual-level prior alignment to enable robust performance from single-light shadows to multi-source ambient lighting. First, Physically Aligned Normalization (PAN) performs closed-form illumination correction via Gray-world normalization, log-domain Retinex decomposition, and dynamic range recombination, suppressing chromatic bias. Second, Geometric-Semantic Rectification Attention (GSRA) extends differential attention to cross-modal alignment, harmonizing depth-derived geometry with DINO-v2 semantic embeddings to resolve modal conflicts under varying illumination. Experiments show competitive performance in shadow removal with lower complexity and generalization to ambient lighting where traditional methods fail under multi-source illumination.

Image quality assessment (IQA) and image restoration are fundamental problems in low-level vision. Although IQA and restoration are closely connected conceptually, most existing work treats them in isolation. Recent advances in unified multimodal understanding-generation models demonstrate promising results and indicate that stronger understanding can improve generative performance. This motivates a single model that unifies IQA and restoration and explicitly studies how IQA can guide restoration, a setting that remains largely underexplored yet highly valuable. In this paper, we propose UARE, to our knowledge the first Unified vision-language model for image quality Assessment, Restoration, and Enhancement. Built on pretrained unified understanding and generation models, we introduce a two-stage training framework. First, a progressive, easy-to-hard schedule expands from single-type distortions to higher-order mixed degradations, enabling UARE to handle multiple degradations. Second, we perform unified fine-tuning of quality understanding and restoration with interleaved text-image data, aligning IQA signals with restoration objectives. Through multi-task co-training, UARE leverages IQA to boost restoration and enhancement performance. Extensive experiments across IQA, restoration, and enhancement tasks demonstrate the effectiveness of UARE. The code and models will be made available.

Inverse tone mapping (ITM) becomes significantly harder when the SDR input is produced by local tone mapping, which jointly applies global radiometric compression and spatially varying adaptations that distort dynamic range, contrast, and channel-wise color ratios. Existing ITM methods ignore this degradation structure and either regress HDR values directly or rely on a single-channel gain map, which scale luminance only and cannot restore the compressed dynamic range and wide color gamut.We introduce FastGaMer, a structured and resolution-agnostic ITM framework that explicitly mirrors this degradation process. Instead of regressing HDR values, we reconstruct a color gain map, which preserves per-channel amplification, simplifies learning, and enables proper gamut extension. Local and global degradations are inverted separately using dynamic bilateral grids and learnable 3D LUTs, followed by a lightweight neural modulator for global refinement and coherence. All high-resolution operations are network-free, yielding exceptional efficiency.To support color-GM supervision under realistic local TMO degradations, we create a dataset of over 8,000 4K SDR–GM pairs with an additional real-captured test set. FastGaMer outperforms prior lightweight ITM methods by +1.4 dB PQ-PSNR, reduces runtime by 70\%, and processes 4K images in only 6.2 ms, achieving both high accuracy and real-time performance.

Recent advances in learning-based video quality assessment (VQA) have achieved remarkable progress, yet the two fundamental components, model and data, are often studied in isolation.Model-centric approaches tend to design superior architectures over fixed and repeatedly used datasets, risking overfitting to benchmark-specific characteristics. In contrast, data-centric efforts emphasize constructing large-scale datasets through costly and time-consuming subjective experiments, typically overlooking the strengths and failure modes of existing VQA models. This separation limits progress, leading to brittle generalization and inefficient use of annotation resources.To bridge the gap, we introduce MDS-VQA, a model-informed data selection method that integrates model-centric and data-centric VQA. In its specific instantiation, a learned failure prediction module trained via a learning-to-rank formulation is combined with a content diversity measure based on deep semantic video features.Experiments across multiple VQA datasets demonstrate that MDS-VQA effectively spots diverse and challenging samples that expose model weaknesses.The selected videos are proven to be particularly informative for fine-tuning, offering a principled path toward constructing more challenging datasets and developing more generalizable and robust VQA models.

Novel view synthesis from low dynamic range (LDR) blurry images, which are common in the wild, struggles to recover high dynamic range (HDR) and sharp 3D representations in extreme lighting conditions. Although existing methods employ event data to address this issue, they ignore the sensor-physics mismatches between the camera output and physical world radiance, resulting in suboptimal HDR and deblurring results. To cope with this problem, we propose a unified sensor-physics grounded NeRF framework for sharp HDR novel view synthesis from single-exposure blurry LDR images and corresponding events. We utilize NeRF to directly represent the actual radiance of the 3D scene in the HDR domain and model raw HDR scene rays hitting the sensor pixels as in the physical world. A pixel-wise RGB mapping field is introduced to align the above NeRF-rendered HDR pixel values with the sensor-recorded LDR pixel values of the input images. A novel event mapping field is also designed to bridge the physical scene dynamics and actual event sensor output. The two mapping fields are jointly optimized with the NeRF network, leveraging the spatial and temporal dynamic information in events to enhance the sharp HDR 3D representation learning. Experiments on the collected and public datasets demonstrate that our method can achieve state-of-the-art deblurring HDR novel view synthesis results from single-exposure blurry LDR images and corresponding events.

Current dehazing methods face two intertwined challenges: (1): the misidentification of haze-related features due to domain-specific interference in both single-domain and empirically integrated multi-domain approaches, and (2): severe chromatic distortion caused by haze-induced inherent color entanglement. To overcome these limitations, we propose a unified framework centered on a $\textbf{Cross-domain Invariant Manifold}$ (CIM), which constructs a consistent latent representation space by aligning multi-domain features through shared scattering semantics. The manifold is optimized via $\textbf{consensus density-driven contrastive learning}$, effectively enhancing cross-domain consistency while eliminating domain-specific biases. Building upon this structured foundation, we further introduce a disentanglement-wise architecture, i.e.the $\textbf{Physics-Guided HSV Decomposition Network}$, that explicitly separates entangled color components to ensure robust color fidelity. Comprehensive experiments demonstrate that our CIM-D framework achieves state-of-the-art performance, effectively eliminating haze-induced color shifts and restoring natural scene appearance. The code will be made publicly available.

We introduce an unsupervised approach for segmenting multiscale subcellular objects in 3D volumetric cryo-electron tomography (cryo-ET) images. To this end, we address key challenges such as lack of annotated data, large data volumes, high heterogeneity of subcellular shapes and sizes, and high inter-domain variability of cellular cryo-ET images across different experiments and contexts. Our method requires users to only select a small number of slabs from a few representative tomograms in the dataset. The core of our method is extracting features for the corresponding slabs, leveraging a Stable Diffusion foundation model pretrained on mostly natural images. The feature extraction is followed by a novel heuristic-based feature aggregation strategy, and adaptive thresholding to segment the aggregated features. The resulting masks are refined with pretrained CellPose to split composite regions, and then utilized as pseudo-ground truth for training supervised deep learning models. We validated our unsupervised foundation-model based pipeline on publicly available cryo-ET benchmark datasets, demonstrating performance that closely approximates expert human annotations. This fully automated, data-driven framework enables the mining of multi-scale subcellular patterns, paving the way for accelerated biological discoveries from large-scale cellular cryo-ET datasets.

3D echocardiography provides superior cardiac quantification to traditional 2D echocardiography, which suffers from geometric idealizations and imaging plane misalignment. However, despite its advantages, clinical adoption of 3D echo remains limited due to logistical and visualization challenges. We propose a novel framework that reconstructs the 3D shape of the left ventricle (LV) throughout the cardiac cycle from sparse 2D echocardiographic views routinely acquired in clinical practice, without the need for external hardware or manual tracking. Our method integrates EchoPOSE, a new deep network that automatically estimates the 6D pose (position and orientation) of LV segmentations, with a graph-harmonic algorithm for 3D shape reconstruction. EchoPOSE employs a transformer-based architecture that combines local image features with global multi-view context, and introduces a geometry-aware loss to ensure spatial consistency across intersecting imaging planes. Trained and evaluated on large-scale synthetic data derived from 3D echocardiography and validated on prospectively acquired clinical echocardiograms, EchoPOSE achieves 3.78 mm and 8.65$^{\circ}$ pose errors, yielding 87.5 Dice reconstruction accuracy, 1.44\% ejection fraction error, and 3.03\% volume error, outperforming alternative deep learning techniques and classical clinical approaches. Notably, the framework remains robust under suboptimal imaging alignment, suggesting that EchoPOSE can reduce the sonography skills required for transducer positioning and allow minimally trained clinicians to perform echo scans.

Segment Anything Model (SAM)-based approaches have demonstrated remarkable potential for biomedical image segmentation. However, these methods often struggle to maintain spatial consistency in 3D electron microscopy (3D-EM) data and require extensive manual annotations. To this end, we propose Spatial-SAM, a spatially consistent and annotation-efficient framework that achieves high precision on 3D-EM data. Our method introduces two key innovations. First, we incorporate a 3D Signed Distance Field (SDF) memory mechanism that replaces the original memory in SAM2 with SDF representations precomputed by a 3D U-Net, providing richer geometric information and improving spatial consistency. Second, by combining the few-shot capability of SAM2 with a dual-track pseudo-label iterative optimization strategy, Spatial-SAM efficiently learns to segment large-scale 3D-EM datasets from minimal annotations. Experiments show that Spatial-SAM significantly outperforms existing semi-supervised methods and achieves performance comparable to state-of-the-art fully supervised approaches on multiple 3D-EM benchmarks, reducing annotation costs while preserving spatial consistency. The code will be publicly released upon acceptance.

Autoregressive models (ARMs) have long dominated the landscape of biomedical vision-language models (VLMs). Recently, masked diffusion models such as LLaDA have emerged as promising alternatives, yet their application in the biomedical domain remains largely underexplored. To bridge this gap, we introduce \textbf{LLaDA-MedV}, the first large language diffusion model tailored for biomedical image understanding through vision instruction tuning. LLaDA-MedV achieves relative performance gains of 7.855\% over LLaVA-Med and 1.867\% over LLaDA-V in the open-ended biomedical visual conversation task, and sets new state-of-the-art accuracy on the closed-form subset of three VQA benchmarks: 84.93\% on VQA-RAD, 92.31\% on SLAKE, and 95.15\% on PathVQA. Furthermore, a detailed comparison with LLaVA-Med suggests that LLaDA-MedV is capable of generating reasonably longer responses by explicitly controlling response length, which can lead to more informative outputs. We also conduct an in-depth analysis of both the training and inference stages, highlighting the critical roles of initialization weight selection, fine-tuning strategies, and the interplay between sampling steps and response repetition. All code and model weights will be released publicly to support future research.

Orthodontic treatment hinges on tooth alignment, which significantly affects occlusal function, facial aesthetics, and patients' quality of life. Current deep learning approaches predominantly predict transformation matrices for the misaligned tooth point cloud via point-to-point geometric constraints to achieve tooth alignment. Nevertheless, these matrices are likely to exhibit clinical-specific distributions, which deterministic constraints fail to capture. To address this, we introduce a new automatic tooth alignment method named TAlignDiff, which is assisted by diffusion-based transformation learning. TAlignDiff comprises two main components: a primary point cloud-based regression network (PRN) and a diffusion-based transformation matrix denoising module (DTMD). Geometry-constrained losses supervise PRN learning for point cloud-level alignment. DTMD, as an auxiliary module, learns the latent distribution of transformation matrices from clinical data. We integrate point cloud-based transformation regression and diffusion-based transformation modeling into a unified framework, allowing bidirectional feedback between geometric constraints and diffusion refinement. We validate our method on a challenge dataset from clinical practice and an extra orthodontic dataset. Its efficacy was confirmed through effective ablation studies and comparative analyses, highlighting its potential for application in orthodontic treatment.

Multi-rater medical image segmentation captures the inherent ambiguity of clinical interpretation, where diagnostic boundaries vary across experts and imaging devices. Existing approaches often reduce this diversity to consensus labels or treat rater differences as noise, resulting in overconfident and poorly calibrated models. We propose a harmonized probabilistic framework that disentangles acquisition artifacts from genuine annotator variability through adaptive feature conditioning and frequency-domain personalization. A lightweight Harmonizer Network implicitly models scanner-specific artifacts and performs dynamic feature modulation to standardize latent representations, ensuring that uncertainty reflects anatomy rather than noise. To represent rater-specific styles, we introduce High-Frequency Prompt Modules that operate in the spectral domain to encode annotator-dependent boundary precision and textural sensitivity. These prompts adaptively modulate harmonized features to produce personalized yet anatomically consistent segmentations. Furthermore, a Generalized Energy Distance (GED)–based regularization aligns the generative distribution with empirical annotation variability, promoting diversity where experts disagree and consensus where they converge. Experiments on LIDC-IDRI and NPC-170 show SOTA aggregated and individualized segmentation, with notable GED reductions and improved Dice scores, especially on noisy cases. Beyond accuracy, the model exhibits clinically meaningful uncertainty, confidence rises in agreement regions and declines in ambiguous areas, supporting its use as a reliable and interpretable tool for multi-expert clinical workflows.

Effective medical image analysis requires representations that capture both global anatomical structure and fine-grained tissue texture. Current self-supervised approaches exhibit limited capacity to address both requirements simultaneously. Invariance-based methods learn through augmentation consistency but face challenges in medical imaging where common augmentations may discard diagnostically relevant intensity patterns. Masked image modeling approaches employ high masking ratios to enforce holistic reasoning, yet inherently limit exposure to fine-grained texture. Recent work in general-domain vision demonstrates that generative and semantic objectives can mutually benefit each other, yet this paradigm remains unexplored for 3D medical imaging. We introduce Masked-Diffusion Autoencoders (MDAE), a self-supervised framework that imposes concurrent spatial masking and diffusion corruption, encouraging the model to learn complementary objectives: masked region reconstruction for structural coherence and visible region denoising for textural characteristics. This dual corruption enables the network to learn structure-texture representations within a unified time-conditioned objective. Evaluated on brain MRI across tumor classification, molecular marker detection, and dense segmentation benchmarks, MDAE consistently outperforms state-of-the-art baselines, with improvements most pronounced in cross-modal generalization tasks.

Brain tumor MRI segmentation is essential for clinical diagnosis and treatment planning, enabling accurate lesion detection and radiotherapy target delineation. However, tumor lesions occupy only a small fraction of the volumetric space, resulting in severe spatial sparsity, while existing segmentation networks often overlook clinically observed spatial priors of tumor occurrence, leading to redundant feature computation over extensive background regions. To address this issue, we propose PGR-Net (Prior-Guided Region Network)—an explicit ROI-aware framework that incorporates a data-driven spatial prior set to capture the distribution and scale characteristics of tumor lesions, providing global guidance for more stable segmentation. Leveraging these priors, PGR-Net introduces a hierarchical Top-K ROI decision mechanism that progressively selects the most confident lesion candidate regions across encoder layers to improve localization precision. We further develop the WinGS-ROI (Windowed Gaussian–Spatial Decay ROI) module, which uses multi-window Gaussian templates with a spatial decay function to produce center-enhanced guidance maps, thus directing feature learning throughout the network. With these ROI features, a windowed RetNet backbone is adopted to enhance localization reliability. Experiments on BraTS 2019, BraTS 2023, and MSD Task01 show that PGR-Net consistently outperforms existing approaches, achieving Dice scores of 89.02%, 91.82%, and 89.67% on the Whole Tumor (WT) region, respectively.

Despite the strong multimodal performance, large vision–language models (LVLMs) are vulnerable during fine-tuning to backdoor attacks, where adversaries insert trigger-embedded samples into the training data to implant behaviors that can be maliciously activated at test time. Existing defenses typically rely on retraining backdoored parameters (e.g., adapters or LoRA modules) with clean data, which is computationally expensive and often degrades model performance. In this work, we provide a new mechanistic understanding of backdoor behaviors in LVLMs: the trigger does not influence prediction through low-level visual patterns, but through abnormal cross-modal attention redistribution, where trigger-bearing visual tokens steal attention away from the textual context — a phenomenon we term attention stealing. Motivated by this, we propose CleanSight, a training-free, plug-and-play defense that operates purely at test time. CleanSight (i) detects poisoned inputs based on the relative visual–text attention ratio in selected cross-modal fusion layers, and (ii) purifies the input by selectively pruning the suspicious high-attention visual tokens to neutralize the backdoor activation. Extensive experiments show that CleanSight significantly outperforms existing pixel-based purification defenses across diverse datasets and backdoor attack types, while preserving the model’s utility on both clean and poisoned samples.

Pre-trained vision-language models (VLMs) have exhibited exceptional generalization capabilities in zero-shot tasks, yet remain vulnerable to adversarial examples. Conventional classification-guided adversarial fine-tuning often compromises the pre-trained cross-modal alignment, undermining the intricate visual-textual correspondence essential for zero-shot performance. To mitigate this, we introduce Alignment-Guided Fine-Tuning (AGFT), a novel framework that preserves semantic integrity while enhancing robustness. AGFT leverages the output distribution of pre-trained VLMs as the fine-tuning objective, thereby maintaining cross-modal semantic correspondence. Recognizing the divergence in feature alignment objectives between pre-trained and robust models, we further calibrate the output distribution by attenuating cross-modal feature similarity of robust models, all while safeguarding correspondence across images and diverse textual descriptions. This calibration ensures compatibility with robust feature representation without sacrificing generalization. Comprehensive experiments across diverse zero-shot datasets and settings demonstrate that AGFT achieves state-of-the-art performance, significantly improving the zero-shot adversarial robustness.

While multimodal large language models (MLLMs) have achieved remarkable success in recent advancements, their susceptibility to jailbreak attacks has come to light. In such attacks, adversaries exploit carefully crafted prompts to coerce models into generating harmful or undesirable content. Existing defense mechanisms often rely on external inference steps or safety alignment training, both of which are less effective and impractical when facing sophisticated adversarial perturbations in white-box scenarios. To address these challenges and bolster MLLM robustness, we introduce SAFEMLLM by adopting an adversarial training framework that alternates between an attack step for generating adversarial noise and a model updating step. At the attack step, SAFEMLLM generates adversarial perturbations through a newly proposed contrastive embedding attack (CoE-Attack), which optimizes token embeddings under a contrastive objective. SAFEMLLM then updates model parameters to neutralize the perturbation effects while preserving model utility on benign inputs. We evaluate SAFEMLLM across multiple MLLMs and six jailbreak methods spanning multiple modalities. Experimental results show that SAFEMLLM effectively defends against diverse attacks, maintaining robust performance and utilities.

Text-Image Person Re-Identification (TI-ReID) models are widely deployed in intelligent surveillance.Built on deep neural networks and vision–language models, TI-ReID models inherit their vulnerability to adversarial attacks, posing potential security risks.Yet their security issues have received far less attention than retrieval accuracy, and the robustness of TI-ReID to adversarial attacks remains largely unexplored.To fill this gap, we propose Reconstruction-residual based Targeted and Untargeted Attack (R$^2$TUA), which takes an image and an adversarial text prompt as input and generates perturbations that make TI-ReID models incorrectly match the perturbed image to the identity described by the adversarial prompt.To precisely inject identity attributes into perturbations and achieve fine-grained targeted attack, R$^2$TUA proposes Transformer-based Gradual Multimodal Fusion (TGMF) that fuses image and adversarial prompt progressively across layers with tunable cross-modal weight.In addition, we propose a fully differentiable Soft Clamp Function (SCF), which enables us to ensure perturbations remain inconspicuous while avoiding local gradient vanishing effects that would trap training into suboptimal local minima.To further align perturbed images with the adversarial text descriptions while leading them to mismatch their original descriptions, R$^2$TUA employs Push-Pull Losses (PPLs) and matching losses during training.Extensive evaluations across multiple datasets and models demonstrate the superior untargeted attack and targeted attack performance of R$^2$TUA. It also exhibits strong adaptability and transferability against black-box models, outperforming all related attacks across multiple tasks.

Vision-Language-Action (VLA) models are vulnerable to adversarial attacks, yet universal and transferable attacks remain underexplored, as most existing patches overfit to a single model and fail in black-box settings. To address this gap, we present a systematic study of **universal, transferable adversarial patches** against VLA-driven robots under unknown architectures, finetuned variants, and sim-to-real shifts. We introduce **UPA-RFAS** (Universal Patch Attack via Robust Feature, Attention, and Semantics), a unified framework that learns a single physical patch in a shared feature space while promoting cross-model transfer. UPA-RFAS combines (i) a feature-space objective with an $\ell_1$ deviation prior and repulsive InfoNCE loss to induce transferable representation shifts, (ii) a robustness-augmented two-phase min-max procedure where an inner loop learns invisible sample-wise perturbations and an outer loop optimizes the universal patch against this hardened neighborhood, and (iii) two VLA-specific losses: Patch Attention Dominance to hijack text$\to$vision attention and Patch Semantic Misalignment to induce image-text mismatch without labels. Experiments across diverse VLA models, manipulation suites, and physical executions show that UPA-RFAS consistently transfers across models, tasks, and viewpoints, exposing a practical patch-based attack surface and establishing a strong baseline for future defenses.

Vision-Language-Action (VLA) models are emerging as a cornerstone for robotics, with flow-matching policies like $\pi_0$ showing great promise in generating smooth, continuous actions. As these models advance, their unique action generation mechanism—the vector field dynamics—presents a critical yet unexplored security vulnerability, particularly backdoor vulnerabilities. Existing backdoor attacks designed for autoregressive discretization VLAs cannot be directly applied to this new continuous dynamics.We introduce FlowHijack, the first backdoor attack framework to systematically target the underlying vector-field dynamics of flow-matching VLAs. Our method combines a novel $\tau$-conditioned injection strategy, which manipulates the initial phase of the action generation, with a dynamics mimicry regularizer. Experiments demonstrate that FlowHijack achieves high attack success rates using stealthy, context-aware triggers where prior works failed. Crucially, it preserves benign task performance and, by enforcing kinematic similarity, generates malicious actions that are behaviorally indistinguishable from normal actions. Our findings reveal a significant vulnerability in continuous embodied models, highlighting the urgent need for defenses targeting the model's internal generative dynamics.

As vision-language models (VLMs) are increasingly deployed in open-world scenarios, they can be easily induced by visual jailbreak attacks to generate harmful content, posing serious risks to model safety and trustworthy usage.Recent activation steering methods inject directional vectors into model activations during inference to induce refusal behaviors and have demonstrated effectiveness.However, a steering vector may both enhance refusal ability and cause over-refusal, thereby degrading model performance on benign inputs.Moreover, due to the lack of theoretical interpretability, these methods still suffer from limited robustness and effectiveness.To better balance safety and utility, we propose \texttt{NullSteer}, a null-space projected activation defense framework.Our method constructs refusal directions within model activations through a linear transformation: it maintains zero perturbation within the benign subspace while dynamically inducing refusal along potentially harmful directions, thereby theoretically achieving safety enhancement without impairing the model’s general capabilities.Extensive experiments show that \texttt{NullSteer} significantly reduces harmful outputs under various jailbreak attacks (average ASR reduction over 15\% on MiniGPT-4) while maintaining comparable performance to the original model on general benchmarks.

Visual grounding aims to associate free-form textual queries with specific regions in an image. While recent Multimodal Large Language Models (MLLMs) have demonstrated promising capabilities in this domain, they primarily excel at object-level grounding and often struggle with part-level grounding—an essential requirement for fine-grained tasks such as robotic manipulation. In this work, we introduce a general approach that equips any open-source MLLMs with accurate 2D part-level point grounding, offering a more flexible alternative to conventional grounding representations. Our method leverages the attention mechanisms inherently present in MLLMs. By synthesizing text-conditioned, grounding-aware queries within intermediate layers via the proposed Q-Synth Module, we extract target-relevant attention patterns and refine them using a lightweight Attention-to-Point Decoder that converts these patterns into a point-centric heatmap for final prediction. Notably, all original MLLM parameters are frozen, ensuring full preservation of their pre-trained capabilities. Experiments show that our design consistently improves part-level grounding accuracy across datasets and can be seamlessly integrated into any open-source MLLMs.

Timely and accurate forecasts of severe weather events are essential for early warning and for constraining downstream analysis and decision-making. Since severe weather events prediction still depends on subjective, time-consuming expert interpretation, end-to-end “AI weather station” systems are emerging but face three major challenges: (1) scarcity of severe weather event samples; (2) imperfect alignment between high-dimensional meteorological data and textual warnings; (3) current multimodal language models cannot effectively process high-dimensional meteorological inputs or capture their complex spatiotemporal dependencies. To address these challenges, we introduce MP-Bench, the first large-scale multimodal dataset for severe weather events prediction, comprising 421,363 pairs of raw multi-year meteorological data and corresponding text caption, covering a wide range of severe weather scenarios. On top of this dataset, we develop a Meteorology Multimodal Large Model (MMLM) that directly ingests 4D meteorological inputs. In addition, it is designed to accommodate the unique characteristics of 4D meteorological data flow, incorporating three plug-and-play adaptive fusion modules that enable dynamic feature extraction and integration across temporal sequences, vertical pressure layers, and spatial dimensions. Extensive experiments on MP-Bench show that MMLM achieves strong performance across multiple tasks, demonstrating effective severe weather understanding and representing a key step toward automated, AI-driven severe weather events forecasting systems. Our source code and dataset will be made publicly available.

Crop yield prediction requires substantial data to train data-driven models. However, creating yield prediction datasets is constrained by high acquisition costs, heterogeneous data quality, and data privacy regulations. Consequently, existing datasets are scarce, low in quality, or limited to regional levels or single crop types. In this work, we release YieldSAT, a large, high-quality, and multimodal dataset for high-resolution crop yield prediction. YieldSAT spans various climate zones across two continents and multiple countries, including Argentina, Brazil, Uruguay, and Germany. The dataset was collected between 2016 and 2024 and comprises four major crop types—corn, rapeseed, soybeans, and wheat—across 2,176 curated fields. In total, over 12.2 million yield samples are available, each with a spatial resolution of $\SI{10}{m}$. Each field is paired with multispectral satellite imagery, resulting in 113,630 labeled satellite images, complemented by auxiliary environmental data. We demonstrate the potential of large-scale and high-resolution crop yield prediction as an image regression task by comparing various deep learning models and data fusion architectures. Furthermore, we highlight open challenges arising from severe distribution shifts in the ground truth data. To mitigate this, we explore a domain-informed Deep Ensemble that exhibits greater diversity in the weight space.

We present the first scalable framework for training sound source localization (SSL) models using synthetic data from text-to-X models. Although SSL has made notable progress, existing models remain constrained by limited-scale, uncurated real-world datasets that often suffer from semantic misalignment. Furthermore, the introduction of new SSL tasks and benchmarks has increased the need for more generalizable models. To address these challenges, we leverage synthetic data to create synthetic clones of the VGGSound dataset, enabling both fully synthetic and hybrid real–synthetic training. We demonstrate that synthetic data can effectively replace, refine, and scale real training datasets. Extensive experiments across multiple benchmarks show that synthetic data not only matches real data in performance but also enables significant improvements when combined with real samples. Our findings provide the first systematic evidence that synthetic data can serve as a scalable and effective approach for advancing SSL models.

Recent advances in large vision models (LVMs) have shifted from modality-specific designs toward unified architectures that jointly process images, videos, and 3D data. However, existing unified LVMs primarily pursue functional integration, while overlooking the deeper goal of cross-vision synergy: the ability to reason over complementary priors across visual modalities. To address this, we present PolyV, a unified LVM that achieves cross-vision synergy at both the architectural and training levels. Architecturally, PolyV adopts a sparse Mixture-of-Experts LVM coordinated by a dynamic modality router, allowing each expert to specialize in modality-specific priors while enabling bidirectional interaction and mutual refinement across modalities. Training-wise, a synergy-aware paradigm combines modality-specific pretraining with coarse-to-fine synergy tuning via knowledge distillation and object-/relation-level alignment. Extensive experiments on ten benchmarks spanning image, video, and 3D understanding, including synergy-focused datasets requiring spatial or temporal priors, demonstrate that PolyV consistently outperforms existing models, achieving over 10\% average improvement over its backbone. Overall, PolyV establishes a unified framework for synesthetic visual reasoning, advancing toward truly synergistic large vision models.

Multimodal Emotion Recognition in Conversations (MERC) aims to understand emotions expressed in each utterance by effectively integrating audio, text, and visual modalities. However, in real-world scenarios, unavoidable missing modalities often degrade multimodal interpretation performance. To address this, we propose \textbf{Hypergraph Diffusion and Evidence Fusion based Emotion Recognition (HyperEF)}, a novel framework designed to mitigate challenges arising from incomplete modalities in MERC. Specifically, to mitigate performance degradation caused by modality absence, we propose Masked Hypergraph Attention (MHGAT) conditioned diffusion model to recover latent features of missing modalities in the latent space. To ensure semantic consistency between recovered and available modalities within the same utterance, we introduce MHGAT that captures high-order semantic information from available modalities to guide the diffusion model’s denoising process. Furthermore, to disentangle and model the complex uncertainties inherent in MERC, we propose Dual Channel Evidence Fusion (DCEF), which estimates uncertainty at both feature source level and discriminative level, thereby achieving adaptive evidence fusion. Extensive comparative experiments and interpretability demonstrate the superior performance of our model in emotion recognition, as well as the contribution of each module within the model.

Major state-of-the-art unified tokenizers predominantly adopt pixel reconstruction and feature alignment as pretext tasks, they leave key domains largely unexplored such as architecture, supervised objectives and tasks interaction, potentially resulting in limited performance. We systematically investigate the critical factors of a unified visual tokenizer and propose a novel framework that strengthens synergy between understanding and generation in various aspects. Our initial analysis focus on properties of frontier vision models, confirming inherent conflict in contrastive learning style models for unifying generation and understanding, and demonstrate distinct convergence behavior of codebooks. To address the above bottleneck, we hierarchically decouple the conflicting proxy tasks, enriching the diversity of semantic features supervision to enhance thesemantic and low-level capabilities. Subsequently, we further introduce attention-prioritized mapping strategy, which guides fine-grained generation with powerful semantic prior. Our method achieves rFID of 0.33 and zero-shot accuracy of 80.9\% on ImageNet at 256$\times$256 resolution, surpassing VILA-U by 7.6\% and outperforms continuous embedding of SigLIP. When applied to discrete unified MLLMs, our 7B model exceeds TokenFlow-13B by 3.1\% in understanding and achieve SOTA performance in GenAI-Bench and MJHQ-30K.

Recent Multimodal Large Language Models (MLLMs) have significantly advanced e-commerce product understanding. However, they still face three challenges: (i) the modality imbalance induced by modality mixed training; (ii) underutilization of the intrinsic alignment relationships among visual and textual information within a product; and (iii) limited handling of noise in e-commerce multimodal data. To address these, we propose MOON2.0, a dynamic modality-balanced Multimodal representation learning framework for e-commerce product understanding. It comprises: (1) a Modality-driven Mixture-of-Experts (MoE) that adaptively processes input samples by their modality composition, enabling Multimodal Joint Learning to mitigate the modality imbalance; (2) a Dual-level Alignment method to better leverage semantic alignment properties inside individual products; and (3) an MLLM-based Image-text Co-augmentation strategy that integrates textual enrichment with visual expansion, coupled with Dynamic Sample Filtering to improve training data quality. We further release MBE2.0, a co-augmented Multimodal representation Benchmark for E-commerce representation learning and evaluation at https://huggingface.co/datasets/ZHNie/MBE2.0. Experiments show that MOON2.0 delivers state-of-the-art zero-shot performance on MBE2.0 and multiple public datasets. Furthermore, attention-based heatmap visualization provides qualitative evidence of improved multimodal alignment of MOON2.0.

The development of affective multimodal language models (MLMs) has long been constrained by a gap between low-level perception and high-level interaction, leading to fragmented affective capabilities and limited generalization. To bridge this gap, we propose a cognitively inspired three-level hierarchy that organizes affective tasks according to their cognitive depth—perception, understanding, and interaction—and provides a unified conceptual foundation for advancing affective modeling. Guided by this hierarchy, we introduce Nano-EmoX, a small-scale multitask MLM, and P2E (Perception-to-Empathy), a curriculum-based training framework. Nano-EmoX integrates a suite of omni-modal encoders, including an enhanced facial encoder and a fusion encoder, to capture key multimodal affective cues and improve cross-task transferability. The outputs are projected into a unified language space via heterogeneous adapters, empowering a lightweight language model to tackle diverse affective tasks. Concurrently, P2E progressively cultivates emotional intelligence by aligning rapid perception with chain-of-thought-driven empathy. To the best of our knowledge, Nano-EmoX is the first compact MLM (2.2B) to unify six core affective tasks across all three hierarchy levels, achieving state-of-the-art or highly competitive performance across multiple benchmarks, demonstrating excellent efficiency and generalization.

Recent multimodal large language models (MLLMs) such as GPT-4o and Qwen3-Omni show strong perception but struggle in multi-speaker, dialogue-centric settings that demand agentic reasoning tracking who speaks, maintaining roles, and grounding events across time. These scenarios are central to multimodal audio-video understanding, where models must jointly reason over audio and visual streams in applications such as conversational video assistants and meeting analytics. We introduce $AMusE$, a benchmark designed around tasks that are inherently agentic, requiring models to decompose complex audio-visual interactions into planning, grounding, and reflection steps. It evaluates MLLMs across three modes zero-shot, guided, and agentic and six task families, including spatio-temporal speaker grounding and multimodal dialogue summarization. Across all modes, current models exhibit weak multi-speaker reasoning and inconsistent behavior under both non-agentic and agentic evaluation. Motivated by the inherently agentic nature of these tasks and recent advances in LLM agents, we propose $RAFT$, a data-efficient agentic alignment framework that integrates reward optimization with intrinsic multimodal self-evaluation as reward and selective parameter adaptation for data and parameter efficient updates. Using $RAFT$, we achieve up to $39.52\%$ relative improvement in accuracy on our benchmark. Together, $AMusE$ and $RAFT$ provide a practical platform for examining agentic reasoning in multimodal models and improving their capabilities. To facilitate further research we will publicly release our code and benchmark.

Multimodal Sentiment Analysis (MSA) aims to integrate textual, acoustic, and visual information to predict sentiment polarity. With the emergence of Large Language Models (LLMs), existing studies commonly employ learnable queries to compress audio–visual representations and feed them as soft prompts into LLMs for MSA. However, due to the implicit learning mechanism of the learnable queries, these learnable queries lack explicit guidance regarding how each query encodes sentiment semantics. To address this issue, we propose a prototype-as-prompt framework that maps audio–visual representations into a fixed set of multimodal sentiment prototypes. These prototypes are then used as soft prompts to guide the LLM in performing MSA. Concretely, we first compress both textual and non-textual features into multimodal prototypes using a resampling-based strategy. We further introduce a sentiment-aware prototype learning that explicitly binds multimodal prototypes with sentiment semantics. To ensure both cross-modal consistency and intra-modal diversity of multimodal sentiment prototypes, we design a cross-modal prototype alignment constraint and a distance-weighted prototype diversity constraint. Extensive experiments across three LLMs and four benchmark datasets show that PaP achieves superior performance with only 0.09\%–0.26\% of trainable parameters, highlighting its effectiveness and parameter efficiency.

Fine-tuning Vision-Language Models (VLMs) trained on large-scale datasets of natural image-text pairs has demonstrated impressive performance for various downstream tasks. However, their fine-tuning for remote sensing (RS) tasks faces dual barriers: (1) Data-level barrier caused by the fundamental modality gap between natural imagery and RS data, and (2) Task-level barrier stemming from the requirement for multi-source interaction modeling capabilities. This paper proposes a Cross-modal Fusion Interactive Prompt Tuning (CF-IPT) method to fine-tune CLIP for multi-source RS image classification tasks. It aims to leverage the prompt learning framework to transfer the alignment target of the text branch shifts from natural images to multi-source RS images. Specifically, we design a Multi-source Interactive Fusion–guided Spectral-Spatial Prompt Generation (MFPG) module, which enables cross-modal feature interaction to generate a prompt matrix that preserves the original spectral and spatial information while performing adaptive multi-scale fusion to address the multi-source image adaptation problem. Subsequently, a Spectral–Spatial Prompt–guided Visual–Text Prompt Interaction (V-TPI) Strategy is proposed, which leverages spectral–spatial prompt matrices to guide visual–textual prompt interaction and inject RS–specific information into both branches of CLIP, ultimately enabling multi-source RS image–text representation alignment. The proposed approach performs the downstream task of multi-source RS image classification with merely 0.76\% of CLIP’s parameters. It is evaluated on several widely used datasets, demonstrating the effectiveness of the proposed approach.

Deep learning models for medical image analysis often act as “black boxes,” seldom aligning with clinical guidelines or explicitly linking decisions to supporting evidence. This is especially critical in Alzheimer's disease (AD), where predictions should be grounded in both anatomical and clinical findings. We present EMAD, a vision–language framework that generates structured AD diagnostic reports with each claim explicitly grounded in multimodal evidence. EMAD uses a hierarchical Sentence–Evidence–Anatomy (SEA) Grounding mechanism: (i) sentence-to-evidence grounding links generated sentences to clinical evidence phrases, and (ii) evidence-to-anatomy grounding localizes corresponding structures on 3D brain MRI. To reduce dense annotation requirements, we propose GTX-Distill, which transfers grounding behavior from a teacher trained with limited supervision to a student operating on model-generated reports. We further introduce Executable-Rule GRPO, a reinforcement fine-tuning scheme with verifiable rewards that enforces clinical consistency, protocol adherence, and reasoning–diagnosis coherence. On the AD-MultiSense dataset, EMAD achieves state-of-the-art diagnostic accuracy and produces more transparent, anatomically faithful reports than existing methods; we will release code and grounding annotations to support future research in trustworthy medical vision–language models.

Multimodal learning often grapples with the challenge of low-quality data, which predominantly manifests as two facets: modality imbalance and noisy corruption. While these issues are often studied in isolation, we argue that they share a common root in the predictive uncertainty towards the reliability of individual modalities and instances during learning. In this paper, we propose a unified framework, termed Conformal Predictive Self-Calibration (CPSC), which leverages conformal prediction to equip the model with the ability to perform self-guided calibration on-the-fly. The core of our proposed CPSC lies in a novel self-calibrating training loop that seamlessly integrates two key modules: (1) Representation Self-Calibration, which decomposes unimodal features into components, selectively fuses the most robust ones identified by a conformal predictor to enhance feature resilience. (2) Gradient Self-Calibration, which recalibrates the gradient flow during backpropagation based on instance-wise reliability scores, steering the optimization towards more trustworthy directions. Furthermore, we also devise a self-update strategy for the conformal predictor to ensure the entire system co-evolves consistently throughout the training process. Extensive experiments on six benchmark datasets under both imbalanced and noisy settings demonstrate that our CPSC framework consistently outperforms existing state-of-the-art methods. Code will be made publicly available.

While existing incomplete multi-view multi-label learning methods have achieved promising performance, few studies have focused on the issue of multi-view imbalance. Existing methods using gradient modulation or alternating optimization strategies alleviate this problem but often oversimplify the interaction between views, resulting in persistently performance. In response to the challenge, we propose the Cross-view Distillation and Adaptive Masking (CDAM) framework, a novel approach designed to achieve balanced multi-view optimization for the challenging double incomplete multi-view multi-label learning tasks. First, to overcome the performance bottleneck of views, we design a cross-view distillation module. This module aligns low-quality student representations with high-quality teacher representations, thereby effectively mitigating the multi-view imbalance problem. Second, recognizing that distillation may not rectify all low-quality views, we introduce a subsequent adaptive masking module to perform an explicit quality assessment. This module dynamically identifies and masks out any remaining unreliable representations before multi-view fusion, thus preventing low-quality information from corrupting the fused representation. Extensive comparisons with nine state-of-the-art methods on six datasets validate the effectiveness and stability of our method.

Audio-visual segmentation (AVS) aims to accurately segment sounding objects in video frames by leveraging audio-visual correspondence cues. However, it remains challenging due to the intrinsic semantic incompleteness within a single modality and the semantic gap between audio and visual representations. Traditional feature-fusion-based decoding approaches struggle to suppress fusion noise effectively, while recent methods that incorporate data-dependent priors often increase the complexity of modeling audio-visual correlations, leading to poor cross-domain generalization. To address these issues, we propose a novel adaptive contrastive and prototype learning framework, BYOAVP, for AVS. Specifically, we design a Self-Supervised Audio Enhancement (SSAE) module that introduces contrastive learning to adaptively align audio representations with gradient-blocked visual embeddings, thus narrowing the semantic gap between modalities. Furthermore, a Dynamic Prototype Constraint (DPC) module is developed to refine pixel-wise category perception via momentum-based prototype updating, while enhancing the localization of sounding regions through cross-modal interaction. Extensive experiments demonstrate that our method achieves state-of-the-art performance across two AVS benchmarks and six sub-tasks, exhibiting strong robustness and generalization ability.

Dataset distillation compresses large training sets into compact synthetic datasets while preserving downstream performance. As modern systems increasingly operate on paired vision–language inputs, multimodal distillation must preserve representation quality and cross-modal alignment under tight compute and memory budgets, yet prior methods often require heavy computes and overlook their correlations. To address this, we present Multimodal Distribution Matching (MDM), a geometry-aware framework for efficient and generalizable multimodal distillation. Specifically, MDM integrates complementary components at the data, model, and loss levels. At the data level, it initializes synthetic image–text pairs by sampling from clusters in the joint embedding space. At the model level, it forms a mixed teacher by interpolating independently fine-tuned models in weight space according to their angular deviation from the pretrained anchor. At the loss level, it matches joint distributions on the unit hypersphere using a geometry-aware matching objective that exploits the joint features in the cross-modal alignment and discrepancy directions along with symmetric contrastive learning. Across image–text retrieval benchmarks with cross-architecture evaluation, MDM yields compact synthetic sets that preserve multimodal semantics, substantially reduce distillation cost, and remain robust across architectures.

Vision–language models (VLMs) have achieved strong performance in visual question answering (VQA), yet they remain constrained by static training data. Retrieval-Augmented Generation (RAG) mitigates this limitation by enabling access to up-to-date, culturally grounded, and multilingual information; however, multilingual multimodal RAG remains largely underexplored. We introduce M4-RAG, a massive-scale benchmark covering 42 languages and 56 regional dialects and registers, comprising over 80,000 culturally diverse image–question pairs for evaluating retrieval-augmented VQA across languages and modalities. To balance realism with reproducibility, we build a controlled retrieval environment containing millions of carefully curated multilingual documents relevant to the query domains, approximating real-world retrieval conditions while ensuring consistent experimentation. Our systematic evaluation reveals that although RAG consistently benefits smaller VLMs, it fails to scale to larger models and often even degrades their performance, exposing a critical mismatch between model size and current retrieval effectiveness. M4-RAG provides a foundation for advancing next-generation RAG systems capable of reasoning seamlessly across languages, modalities, and cultural contexts.

Our goal is to train a generative model of 3D hand motions, conditioned on natural language descriptions specifying motion characteristics such as handshapes, locations, finger/hand/arm movements. To this end, we automatically build pairs of 3D hand motions and their associated textual labels with unprecedented scale. Specifically, we leverage a large-scale sign language video dataset, along with noisy pseudo-annotated sign categories, which we translate into hand motion descriptions via an LLM that utilizes a dictionary of sign attributes, as well as our complementary motion-script cues. This data enables training a text-conditioned hand motion diffusion model (HandMDM), that is robust across domains such as unseen sign categories from the same sign language, but also signs from another sign language and non-sign hand movements. We contribute extensive experimental investigation of these scenarios and will make our trained models and data publicly available to support future research in this relatively new field.

Recent progress in robot learning has been driven by large-scale datasets and powerful visuomotor policy architectures, yet policy robustness remains limited by the substantial cost of collecting diverse demonstrations, particularly for spatial generalization in manipulation tasks. To reduce repetitive data collection, we present Real2Edit2Real, a framework that generates new demonstrations by bridging 3D editability with 2D visual data through a 3D control interface. Our approach first reconstructs scene geometry from multi-view RGB observations with a metric-scale 3D reconstruction model. Based on the reconstructed geometry, we perform depth-reliable 3D editing on point clouds to generate new manipulation trajectories while geometrically correcting the robot poses to recover physically consistent depth, which serves as a reliable condition for synthesizing new demonstrations. Finally, we propose a multi-conditional video generation model guided by depth as the primary control signal, together with action, edge, and ray maps, to synthesize spatially augmented multi-view manipulation videos. Experiments on four real-world manipulation tasks demonstrate that policies trained on data generated from only $1$–$5$ source demonstrations can match or outperform those trained on 50 real-world demonstrations, improving data efficiency by up to $10$-$50\times$. Moreover, experimental results on height and texture editing demonstrate the framework’s flexibility and extensibility, indicating its potential to serve as a unified data generation framework.

Hand–Object Interaction (HOI) remains a core challenge in digital human video synthesis, where models must generate physically plausible contact and preserve object identity across frames. Although recent HOI reenactment approaches have achieved progress, they are typically trained and evaluated in-domain and fail to generalize to complex, in-the-wild scenarios. In contrast, all-in-one video editing models exhibit broader robustness but still struggle with HOI-specific issues such as inconsistent object appearance. In this paper, we present GenHOI, a lightweight augmentation to pretrained video generation models that injects reference-object information in a temporally balanced and spatially selective manner. For temporal balancing, we propose Head-Sliding RoPE, which assigns head-specific temporal offsets to reference tokens, distributing their influence evenly across frames and mitigating the temporal decay of 3D RoPE to improve long-range object consistency. For spatial selectivity, we design a two-level spatial attention gate that concentrates object-conditioned attention on HOI regions and adaptively scales its strength, preserving background realism while enhancing interaction fidelity. Extensive qualitative and quantitative evaluations on unseen, in-the-wild scenes demonstrate that GenHOI significantly outperforms state-of-the-art HOI reenactment and all-in-one video editing competitors.

Understanding hand-object interaction from monocular videos is crucial for immersive and dexterous interactions in AR/VR and robotic applications. However, existing monocular reconstruction methods primarily assume rigid grasping and static object geometry. When applied to articulated manipulations, the continuous joint rotations and frequent component deformations introduce a strong coupling between shape and motion, leading to severe ambiguity and instability in articulation optimization under monocular observation. To address this challenge, we propose a Clay-to-Stone dual-phase framework, modeling the articulated manipulation at hierarchical granularities, enabling a progression from flexible semantic exploration to structured articulation recovery. In the CLAY phase, our method performs fine-grained control over geometric deformation, guided by inter-part semantic correlation learning. As semantic and motion priors emerge, the STONE phase enforces rigid constraints to consolidate articulated structures and explicitly estimates motion parameters. Experiments on a real-world manipulation dataset show that our method achieves state-of-the-art reconstruction quality and plausible articulation modeling from monocular videos.

Compositional video generation aims to synthesize multiple instances with diverse appearance and motion, which is widely applicable in real-world scenarios. However, current approaches mainly focus on binding semantics, neglecting to understand diverse motion categories specified in prompts. In this paper, we propose a \textbf{motion factorization} framework that decomposes complex motion into three primary categories: motionlessness, rigid motion, and non-rigid motion. Specifically, our framework follows a \textit{planning before generation} paradigm. (1) During planning, we reason about motion laws on the motion graph to obtain frame-wise changes in the shape and position of each instance. This alleviates semantic ambiguities in the user prompt by organizing it into a structured representation of instances and their interactions. (2) During generation, we modulate the synthesis of distinct motion categories in a disentangled manner. Conditioned on the motion cues, guidance branches stabilize appearance in motionless regions, preserve rigid-body geometry, and regularize local non-rigid deformations. Crucially, our two modules are model-agnostic, which can be seamlessly incorporated into various diffusion model architectures. Extensive experiments demonstrate that our framework achieves impressive performance in motion synthesis on real-world benchmarks. \textcolor[RGB]{237,0, 140}{Our code will be released soon}.

Recent advancements in video generation highlight that realistic audio-visual synchronization is crucial for engaging content creation. However, existing video editing methods largely overlook audio-visual synchronization and lack the fine-grained spatial and temporal controllability required for precise instance-level edits. In this paper, we propose AVI-Edit, a framework for audio-sync video instance editing. We propose a granularity-aware mask refiner that iteratively refines coarse user-provided masks into precise instance-level regions. We further design a self-feedback audio agent to curate high-quality audio guidance, providing fine-grained temporal control. To facilitate this task, we additionally construct a large-scale dataset with instance-centric correspondence and comprehensive annotations. Extensive experiments demonstrate that AVI-Edit outperforms state-of-the-art methods in visual quality, condition following, and audio-visual synchronization.

Autoregressive (AR) language models rely on causal tokenization, but extending this paradigm to vision remains non-trivial. Current visual tokenizers either flatten 2D patches into non-causal sequences or enforce heuristic orderings that misalign with the "next-token prediction" pattern. Recent diffusion autoencoders similarly fall short: conditioning the decoder on all tokens lacks causality, while applying nested dropout mechanism introduces imbalance. To address these challenges, we present CaTok, a 1D causal image tokenizer with a MeanFlow decoder. By selecting tokens over time intervals and binding them to the MeanFlow objective, CaTok learns causal 1D representations that support both fast one-step generation and high-fidelity multi-step sampling, while naturally capturing diverse visual concepts across token intervals. To further stabilize and accelerate training, we propose a straightforward regularization REPA-A, which aligns encoder features with Vision Foundation Models (VFMs). Experiments demonstrate that CaTok achieves state-of-the-art results on ImageNet reconstruction, reaching 22.72 PSNR and 0.681 SSIM with fewer training epochs, and the AR model attains performance comparable to leading approaches.

First-Frame Propagation (FFP) offers a promising paradigm for controllable video editing, but existing methods are hampered by a reliance on cumbersome run-time guidance. We identify the root cause of this limitation as the inadequacy of current training datasets, which are often too short, low-resolution, and lack the task diversity required to teach robust temporal priors. To address this foundational data gap, we first introduce FFP-300K, a new large-scale dataset comprising 300K high-fidelity video pairs at 720p resolution and 81 frames in length, constructed via a principled two-track pipeline for diverse local and global edits. Building on this dataset, we propose a novel framework designed for true guidance-free FFP that resolves the critical tension between maintaining first-frame appearance and preserving source video motion. Architecturally, we introduce Adaptive Spatio-Temporal RoPE (AST-RoPE), which dynamically remaps positional encodings to disentangle appearance and motion references. At the objective level, we employ a self-distillation strategy where an identity propagation task acts as a powerful regularizer, ensuring long-term temporal stability and preventing semantic drift. Comprehensive experiments on the EditVerseBench benchmark demonstrate that our method significantly outperforming existing academic and commercial models by receiving about 0.2 PickScore and 0.3 VLM score improvement against these competitors.

Large-scale video generation models have shown remarkable potential in modeling photorealistic appearance and lighting interactions in real-world scenes. However, a closed-loop framework that jointly understands intrinsic scene properties (e.g., albedo, normal, material, and irradiance), leverages them for video synthesis, and supports editable intrinsic representations remains unexplored. We present V-RGBX, the first end-to-end framework for intrinsic-aware video editing. V-RGBX unifies three key capabilities: (1) video inverse rendering into intrinsic channels, (2) photorealistic video synthesis from these intrinsic representations, and (3) keyframe-based video editing conditioned on intrinsic channels. At the core of V-RGBX is an interleaved conditioning mechanism that enables intuitive, physically grounded video editing through user-selected keyframes, supporting flexible manipulation of any intrinsic modality. Extensive qualitative and quantitative results show that V-RGBX produces temporally consistent, photorealistic videos while propagating keyframe edits across sequences in a physically plausible manner. We demonstrate its effectiveness in diverse applications, including object appearance editing and scene-level relighting, surpassing the performance of prior methods.

Pose-guided video generation refers to controlling the motion of subjects in generated video through a sequence of poses. It enables precise control over subject motion and has important applications in animation. However, current pose-guided video generation methods are limited to accepting only human poses as input, thus generalizing poorly to pose of other subjects. To address this issue, we propose PoseAnything, the first universal pose-guided video generation framework capable of handling both human and non-human characters, supporting arbitrary skeletal inputs. To enhance consistency preservation during motion, we introduce Part-aware Temporal Coherence Module, which divides the subject into different parts, establishes part correspondences, and computes cross-attention between corresponding parts across frames to achieve fine-grained part-level consistency. Additionally, we propose Subject and Camera Motion Decoupled CFG, a novel guidance strategy that, for the first time, enables independent camera movement control in pose-guided video generation, by separately injecting subject and camera motion control information into the positive and negative anchors of CFG. Furthermore, we present XPose, a high-quality public dataset containing 50,000 non-human pose-video pairs, along with an automated pipeline for annotation and filtering. Extensive experiments demonstrate that Pose-Anything significantly outperforms state-of-the-art methods in both effectiveness and generalization.

Autoregressive (AR) models have achieved remarkable success in natural language processing, yet their application to image generation faces significant challenges. When implementing VQ-based decoders for autoregressive image generation, the generated images typically preserve semantic information but struggle with fine-grained details. Recent hybrid AR image generation approaches address these issues by integrating diffusion models as decoder heads, enabling more high-fidelity generation. However, the diffusion-based denoising process introduces significant computational overhead during inference.To accelerate hybrid AR image generation, we propose the Lookahead Decoding Strategy, which integrates the strengths of autoregressive and diffusion models by separating the process into two complementary branches: semantic prediction and detail refinement. The autoregressive branch captures high-level semantic structures while refining coarse predictions made by the parallel. Furthermore, we introduce Guided Diffusion Sampling to steer the diffusion denoising trajectory, significantly reducing the number of denoising steps. Extensive experiments demonstrate that our approach provides an effective solution for accelerating hybrid AR image generation models.

Decoder-only autoregressive image generation typically relies on fixed-length tokenization schemes whose token counts grow quadratically with resolution, substantially increasing the computational and memory demands of attention. We present DPAR, a novel decoder-only autoregressive model that dynamically aggregates image tokens into a variable number of patches for efficient image generation. Our work is the first to demonstrate that next-token prediction entropy from a lightweight and unsupervised autoregressive model provides a reliable criterion for merging tokens into larger patches based on information content. DPAR makes minimal modifications to the standard decoder architecture, ensuring compatibility with multimodal generation frameworks and allocating more compute to generation of high-information image regions. Further, we demonstrate that training with dynamically sized patches yields representations that are robust to patch boundaries, allowing DPAR to scale to larger patch sizes at inference. DPAR reduces token count by 1.81x and 2.06x on Imagenet 256 and 384 generation resolution respectively, leading to a reduction of up to 40\% FLOPs in training costs. Further, our method exhibits faster convergence and improves FID by up to 27.1\% relative to baseline models.

Recent advances in subject-driven video generation with large diffusion models have enabled personalized content synthesis conditioned on user-provided subjects.However, existing methods lack fine-grained temporal control over subject appearance and disappearance, which are essential for applications such as compositional video synthesis, storyboarding, and controllable animation.We propose AlcheMinT, a unified framework that introduces explicit timestamps conditioning for subject-driven video generation.Our approach introduces a novel positional encoding mechanism that unlocks the encoding of temporal intervals, associated in our case with subject identities, while seamlessly integrating with the pretrained video generation model positional embeddings.Additionally, we incorporate subject-descriptive text tokens to strengthen binding between visual identity and video captions, mitigating ambiguity during generation.Through token-wise concatenation, AlcheMinT avoids any additional cross-attention modules and incurs negligible parameter overhead.We establish a benchmark evaluating multiple subject identity preservation, video fidelity, and temporal adherence.Experimental results demonstrate that AlcheMinT achieves visual quality matching state-of-the-art video personalization methods, while, for the first time, enabling precise temporal control over multi-subject generation within videos.

This paper focuses on the alignment of flow-matching models with human preference. A promising way is fine-tuning by directly backpropagating reward signals through the differentiable generation process of flow matching. However, backpropagating through long trajectories results in prohibitive memory costs and gradient explosion. Therefore, direct-gradient methods struggle to update early generation steps, which are crucial for determining the global structure of the final image. To address this issue, we introduce LeapAlign, a fine-tuning method that reduces computational cost and enables direct gradient propagation from reward to early-step latents.Specifically, we shorten the long trajectory into only two steps by designing two consecutive leaps, each skipping multiple ODE sampling steps and predicting future latents in a single step. By randomizing the start and end timesteps of the leaps, LeapAlign leads to efficient and stable model updates at any generation step. To better use such shortened trajectories, we assign higher training weights to those that are more consistent with the long generation path. To further enhance gradient stability, we reduce the weights of terms with large gradients, instead of completely removing them as done in previous works. When fine-tuning the Flux model, LeapAlign consistently outperforms state-of-the-art GRPO-based and direct-gradient methods across various metrics, achieving superior image quality and image–text alignment.

Autoregressive (AR) video generative models rely on video tokenizers that compress pixels into discrete token sequences. The length of these token sequences is crucial for balancing reconstruction quality against downstream generation computational cost. Traditional video tokenizers apply a uniform token assignment across temporal blocks of different videos, often wasting tokens on simple, static, or repetitive segments while underserving dynamic or complex ones. To address this inefficiency, we introduce EVATok, a framework to produce Efficient Video Adaptive Tokenizers. Our framework estimates optimal token assignments for each video to achieve the best quality-cost trade-off, develops lightweight routers for fast prediction of these optimal assignments, and trains adaptive tokenizers that encode videos based on the assignments predicted by routers. We demonstrate that EVATok delivers substantial improvements in efficiency and overall quality for video reconstruction and downstream AR generation. Enhanced by our advanced training recipe that integrates video semantic encoders, EVATok achieves superior reconstruction and state-of-the-art class-to-video generation on UCF-101, with at least 24.4\% savings in average token usage compared to the prior state-of-the-art LARP and our fixed-length baseline.

Visual foundation models play an increasingly important role in the training efficiency of flow-based models by inducing structured latent space through alignments, distillations, adapters, and even replacements of visual encoders. When structured latent space improves training efficiency by lowering the complexity of the target (latent) distribution, the efficiency can be further boosted by a data-adaptive multimodal source (noise) distribution that globally shortens the distance to the target (latent) distribution, and a mode-dependent coupling between source and target samples to move probability mass locally. To this end, we propose an efficient source and coupling co-design algorithm termed Mixture-Modeling Flow Matching (MM-FM). Under linear conditional flow objective and multimodal target assumption, our theoretical results reveal straighter and shorter sampling trajectories and smaller Lipschitz constant for learning complexity relative to isotropic Gaussian with independent coupling. In our ImageNet256x256 experiments with multimodal DINOv2-B latents, we observe superior convergence and state-of-the-art unconditional generation FID=2.74 with autoguidance in only 80 epochs.

Image Chain-of-Thought (Image-CoT) is a test-time scaling paradigm that improves image generation by extending inference time. Most Image-CoT methods focus on text-to-image (T2I) generation. Unlike T2I generation, image editing is goal-directed: the solution space is constrained by the source image and instruction. This mismatch causes three challenges when applying Image-CoT to editing: inefficient resource allocation with fixed sampling budgets, unreliable early-stage verification using general MLLM scores, and redundant edited results from large-scale sampling. To address this, we propose ADaptive Edit-CoT (ADE-CoT), an on-demand test-time scaling framework to enhance editing efficiency and performance. It incorporates three key strategies: (1) a difficulty-aware resource allocation that assigns dynamic budgets based on estimated edit difficulty; (2) edit-specific verification in early pruning that uses region localization and caption consistency to select promising candidates; and (3) depth-first opportunistic stopping, guided by an instance-specific verifier, that terminates when intent-aligned results are found. Extensive experiments on three SOTA editing models (Step1X-Edit, BAGEL, FLUX.1 Kontext) across three benchmarks show that ADE-CoT achieves superior performance-efficiency trade-offs. With comparable sampling budgets, ADE-CoT obtains better performance with more than 2× speedup over Best-of-N.

Recent advances in image editing models have shown remarkable progress. A common architectural design couples a multimodal large language model (MLLM) encoder with a diffusion decoder, as seen in systems such as Step1X-Edit and Qwen-Image-Edit, where the MLLM encodes both the reference image and the instruction but remains frozen during training.In this work, we demonstrate that unlocking the reasoning capabilities of MLLM can further push the boundaries of editing models. Specifically, we explore two reasoning mechanisms, thinking and reflection, which enhance instruction understanding and editing accuracy.Based on that, our proposed framework enables image editing in a thinking–editing–reflection loop: the thinking mechanism leverages the world knowledge of MLLM to interpret abstract instructions, while the reflection reviews editing results, automatically corrects unintended manipulations, and identifies the stopping round.Extensive experiments demonstrate that our reasoning approach achieves significant performance gains, with improvements of ImgEdit(+4.4\%), GEdit(+3.1\%), and Kris(+11.5\%) when initializing our DiT from the Step1X-Edit(ReasonEdit-S), and also outperforms previous open-source methods on both GEdit and Kris when integrated with Qwen-Image-Edit(ReasonEdit-Q).Code and checkpoints will be open-sourced.

Reconstructing video content from EEG (electroencephalogram) is a research task of significant scientific importance. However, due to inter-subject differences in physiological states and variations in signal acquisition configurations, this task faces the challenge of inconsistent cross-subject generation.To address this, we propose a Subject Adversarial and Mapping Network (SAM-Net). In SAM-Net, we first introduce a Hybrid Region-Temporal (HRT) Encoder to conduct inter-channel semantic interactions guided by brain regions and aggregate temporal semantics across different time scales. Secondly, we propose a Centered-progressive Subject Adversarial (C-SA) Mechanism to gradually narrow the metric distance between different subjects, thereby obtaining a unified and stable semantic representation. Thirdly, we design a New2Source Mapper to align the EEG distribution of new subjects with that of multiple known subjects. Finally, we adopt a keyframe-guided continuous semantic generation paradigm to drive the production of coherent and high-quality videos. Extensive experiments validate the competitive performance of our SAM-Net in cross-subject EEG-to-Video generation tasks, as well as its excellent performance in generation tasks involving new subjects.

Recent proprietary models such as Sora‑2 demonstrate promising progress in generating multi‑shot videos conditioned on multiple reference characters. However, academic research on this problem remains limited. We study this task and identify a core challenge: when reference images exhibit highly similar appearances, the model often suffers from reference confusion, where semantically similar tokens degrade the model’s ability to retrieve the correct context. To address this, we introduce PoCo (Position Embedding as a Context Controller), which incorporates position encoding as additional context control beyond semantic retrieval. By employing side information of tokens, PoCo enables precise token‑level matching while preserving implicit semantic consistency modeling. Building on PoCo, we develop a multi‑reference and multi‑shot video generation model capable of accurately controlling characters with extremely similar visual traits. Extensive experiments demonstrate that PoCo improves cross‑shot consistency and reference fidelity compared with various baselines.

Generating high-fidelity human videos that match user-specified identities is important yet challenging in the field of generative AI.Existing methods often rely on an excessive number of training parameters and lack compatibility with other AIGC tools.In this paper, we propose $\textbf{Stand-In}$, a lightweight and plug-and-play framework for identity preservation in video generation.Specifically, we introduce a conditional image branch into the pre-trained video generation model.Identity control is achieved through restricted self-attentions with conditional position mapping.Thanks to these designs, which greatly preserve the pretrained prior of the video generation model, our approach is able to outperform other full-parameter training methods in video quality and identity preservation, even with just $\sim$1\% additional parameters and only 2000 training pairs.Moreover, our framework can be seamlessly integrated for other tasks, such as subject-driven video generation, pose-referenced video generation, stylization, and face swapping.Code and dataset will be available to the community.

Recent diffusion and flow matching models have demonstrated strong capabilities in image generation and editing by progressively removing noise through iterative sampling. While this enables flexible inversion for semantic-preserving edits, few-step sampling regimes suffer from poor forward process approximation, leading to degraded editing quality. Existing few-step inversion methods often rely on pretrained generators and auxiliary modules, limiting scalability and generalization across different architectures.To address these limitations, we propose BiFM (Bidirectional Flow Matching), a unified framework that jointly learns generation and inversion within a single model. BiFM directly estimates average velocity fields in both ''image $\to$ noise'' and ''noise $\to$ image'' directions, constrained by a shared instantaneous velocity field derived from either predefined schedules or pretrained multi-step diffusion models. Additionally, BiFM introduces a novel training strategy using continuous time-interval supervision, stabilized by a bidirectional consistency objective and a lightweight time-interval embedding. This bidirectional formulation also enables one-step inversion and can integrate seamlessly into popular diffusion and flow matching backbones.Across diverse image editing and generation tasks, BiFM consistently outperforms existing few-step approaches, achieving superior performance and editability. Our code and models will be released upon acceptance.

Recent progress in video diffusion models has enabled remarkable generative fidelity, yet leveraging these priors for restoration remains limited by the strong coupling between conditional and unconditional branches in standard classifier-free guidance. We introduce a training-free framework that enhances distorted and low-resolution videos by decoupling these signals in time. Our proposed Decoupled Time Guidance (DTG) evaluates the unconditional branch at a cleaner diffusion timestep, providing a lookahead prior that preserves geometry while suppressing replication of warped content. This temporal bias is annealed throughout sampling, allowing the model to transition from structure correction to detail refinement without retraining. Combined with any off-the-shelf restoration module in a plug-and-play manner, our approach improves perceptual coherence and restores plausible structure in AI-generated and real-world videos alike. To facilitate evaluation, we curate GenWarp480, a benchmark of 4000 distorted 480p videos synthesized from diverse text-to-video models. GenWarp480 focuses on characteristic generative degradations such as warped faces, body misalignments, and spatial artifacts, providing a purpose-built testbed for assessing robustness to generative errors. Extensive experiments demonstrate that our method achieves significant improvements in structural fidelity and temporal stability without any model training.

Recent advances in video generation have been remarkable, enabling models to produce visually compelling videos with synchronized audio. While existing video generation benchmarks provide comprehensive metrics for visual quality, they lack convincing evaluations for audio-video generation, especially for models aiming to generate synchronized audio-video outputs. To address this gap, we introduce VABench, a comprehensive and multi-dimensional benchmark framework designed to systematically evaluate the capabilities of synchronous audio-video generation. VABench encompasses three primary task types: text-to-audio-video (T2AV), image-to-audio-video (I2AV), and stereo audio-video generation. It further establishes two major evaluation modules covering 15 dimensions. These dimensions specifically assess pairwise similarities (text-video, text-audio, video-audio), audio-video synchronization, lip-speech consistency, and carefully curated audio and video question-answering (QA) pairs, among others. Furthermore, VABench covers seven major content categories: animals, human sounds, music, environmental sounds, synchronous physical sounds, complex scenes, and virtual worlds. We provide a systematic analysis and visualization of the evaluation results, aiming to establish a new standard for assessing video generation models with synchronous audio capabilities and to promote the comprehensive advancement of the field.

We present a method for harmonizing the lighting of a foreground video to match a target background scene, adjusting shadows, color tone, and illumination intensity (relightful harmonization). Unlike images, acquiring labeled data for videos, where identical motions are recorded under different lighting conditions, is practically infeasible and non-scalable. While one way to create such paired data is to apply existing image-based harmonization models frame by frame to a video, the resulting outputs often suffer from significant temporal jitters. We overcome this problem by introducing a novel lighting deflickering model that can stabilize the global and local lighting flickering artifacts. Our video diffusion model learns from these upgraded deflickered data with a volume of real and synthetic videos to generate high-quality video harmonization results. We further propose an asymmetric alpha mask conditioning technique to learn the clean boundaries from real videos. Experiments demonstrate that our model achieves strong temporal coherence, naturalness, cleaner boundaries, and physically meaningful lighting behavior, while maintaining strong relighting expressiveness compared to prior image-based and video-based harmonization methods.

In this work, we propose DiT360, a DiT-based framework that performs hybrid training on perspective and panoramic data for panoramic image generation. We attribute the main challenges in preserving geometric fidelity and photorealism to the scarcity of large-scale, high-quality real-world panoramic data, in contrast to prior methods that emphasize model design. Basically, DiT360 has several key modules for inter-domain transformation and intra-domain augmentation, applied at both the pre-VAE image level and the post-VAE token level. At the image level, we incorporate cross-domain knowledge through perspective image guidance and panoramic refinement, which enhance perceptual quality while regularizing diversity and photorealism. At the token level, hybrid supervision is applied across multiple modules, which include circular padding for boundary continuity, yaw loss for rotational robustness, and cube loss for distortion awareness. Extensive experiments on text-to-panorama, inpainting, and outpainting tasks demonstrate that our method achieves better boundary consistency and image fidelity across eleven quantitative metrics. Our code will be released publicly.

Next-scale prediction paradigm visual autoregressive (VAR) models have demonstrated significant potential for image super-resolution. However, their practical application is constrained by a rigid, size-specific design. This limitation stems from their reliance on memorizing fixed, absolute scaling schedules, which necessitates a distinct model for each target resolution. We introduce DVAR, a Dynamic Visual AutoRegressive framework that overcomes this fundamental bottleneck. Instead of memorizing these rigid schedules, DVAR learns a canonical scaling dynamic. This dynamic effectively decouples the logic of relative scaling from the absolute target size, thereby preserving a single set of proportions between generative steps that can be applied uniformly to any size. Furthermore, we introduce a dynamic sampling scheduler to mitigate the teacher-forcing problem with negligible computational overhead. By leveraging the geometric proximity of visual tokens in the codebook, it efficiently simulates the model's predictive error distribution to bridge the training-inference gap. To our knowledge, DVAR is the first framework to grant VAR models size-flexibility, breaking their one-to-one dependency on a fixed resolution. Extensive evaluations demonstrate that DVAR achieves superior visual quality over existing Real-ISR methods, proving that a flexible, purely autoregressive approach is a viable path to state-of-the-art image super-resolution.

Recent advances in diffusion-based controllable visual generation have led to remarkable improvements in image quality. However, these powerful models are typically deployed on cloud servers due to their large computational demands, raising serious concerns about user data privacy. To enable secure and efficient on-device generation, we explore in this paper controllable diffusion models built upon linear attention architectures, which offer superior scalability and efficiency, even on edge devices. Yet, our experiments reveal that existing controllable generation frameworks, such as ControlNet and OminiControl, either lack the flexibility to support multiple heterogeneous condition types or suffer from slow convergence on such linear-attention models. To address these limitations, we propose a novel controllable diffusion framework tailored for linear attention backbones like SANA. The core of our method lies in a unified gated conditioning module working in a dual-path pipeline, which effectively harmonizes multi-type conditional inputs, such as image, semantic, and spatial cues, while maintaining training stability. Extensive experiments on multiple tasks and benchmarks demonstrate that our approach achieves state-of-the-art controllable generation performance based on linear-attention models, surpassing existing methods in terms of fidelity and controllability. Codes will be available.

Video diffusion models (DMs) have enabled high-quality video synthesis, but their computation costs scale quadratically with sequence length due to the nature of self-attention. While linear attention offers a more efficient alternative, fully replacing quadratic attention demands costly pretraining. This is largely because linear attention lacks sufficient expressiveness and struggles with the complex spatiotemporal dynamics inherent to video generation. In this paper, we present LinVideo, an efficient data-free post-training framework that replaces a target number of self-attention modules with linear attention while preserving performance. First, we observe a significant disparity in the replaceability of different layers. Instead of manual or heuristic choices, we frame layer selection as a binary classification problem and propose a selective transfer, which automatically and progressively converts layers to linear attention with minimal performance impact. Additionally, to overcome the ineffectiveness and even inefficiency of existing objectives in optimizing this challenge transfer process, we introduce an anytime distribution matching (ADM) objective that aligns the distributions of samples across any timestep along the sampling trajectory. This objective is highly efficient and recovers model performance. Extensive experiments show that LinVideo achieves a $\mathbf{1.43\text{-}1.71\times}$ speedup while preserving generation quality, and the 4-step distilled models further reduce latency by $\mathbf{15.9\text{-}20.9\times}$ with only a minor drop in visual quality.

Hybrid CNN-Transformer architectures achieve strong results in image super-resolution, but scaling attention windows or convolution kernels significantly increases computational cost, limiting deployment on resource-constrained devices. We present UCAN, a lightweight network that unifies convolution and attention to expand the effective receptive field efficiently. UCAN combines window-based spatial attention with a Hedgehog Attention mechanism to model both local texture and long-range dependencies, and introduces a distillation-based large-kernel module to preserve high-frequency structure without heavy computation. In addition, we employ cross-layer parameter sharing to further reduce complexity. On Manga109 ($4\times$), UCAN-L achieves 31.63 dB PSNR with only 48.4G MACs, surpassing recent lightweight models. On BSDS100, UCAN attains 27.79 dB, outperforming methods with significantly larger models. Extensive experiments show that UCAN achieves a superior trade-off between accuracy, efficiency, and scalability, making it well-suited for practical high-resolution image restoration. The code will be published later.

Hardware constraints make it challenging to simultaneously acquire hyperspectral images (HSIs) with both high spatial and high spectral resolutions. A promising solution is to fuse low-resolution HSI (LR-HSI) with high-resolution multispectral images (HR-MSI) to generate high-resolution HSI (HR-HSI). Recently, diffusion models have introduced possibilities for HSI super-resolution, but suffer from low-efficiency sampling, detail-limited generation, and insufficient denoising. To address these issues, we propose an Edge-aware Multimodal Residual Diffusion Model (EMR-Diff). Specifically, multimodal residual mechanism is introduced to facilitate efficient information transfer among HR-MSI, LR-HSI, and HR-HSI, significantly improving the fusion efficiency. Edge-aware noise strategy is designed by exploiting the edge information of HR-MSI, which guides the model to prioritize high-frequency detail reconstruction by applying stronger noise perturbations to edge regions. In addition, we propose a Bilateral Attention Fusion UNet and design a multi-scale supervision mechanism to enable progressive reconstruction and collaborative optimization of spectral and spatial features. Extensive experiments demonstrate that our method achieves superior performance over existing approaches in both quantitative metrics and visual quality.

Digital zoom on smartphones relies on learning-based super-resolution (SR) models that operate on RAW sensor images, but obtaining sensor-specific training data is challenging due to the lack of ground-truth images. Synthetic data generation via ``unprocessing'' pipelines offers a potential solution by simulating the degradations that transform high-resolution (HR) images into their low-resolution (LR) counterparts. However, these pipelines can introduce domain gaps due to incomplete or unrealistic degradation modeling. In this paper, we demonstrate that principled and carefully designed degradation modeling can enhance SR performance in real-world conditions. Instead of relying on generic priors for camera blur and noise, we model device-specific degradations through calibration and unprocess publicly available rendered images into the RAW domain of different smartphones. Using these image pairs, we train a single-image RAW-to-RGB SR model and evaluate it on real data from a held-out device. Our experiments show that accurate degradation modeling leads to noticeable improvements, with our SR model outperforming baselines trained on large pools of arbitrarily chosen degradations. We will make our calibrated kernels and noise models publicly available, to facilitate research on image enhancement for mobile photography.

Recent advances in diffusion-based real-world image super-resolution (Real-ISR) have demonstrated remarkable perceptual quality, yet the balance between fidelity and controllability remains a problem: multi-step diffusion-based methods suffer from generative diversity and randomness, resulting in low fidelity, while one-step methods lose control flexibility due to fidelity-specific finetuning. In this paper, we present ODTSR, a one-step diffusion transformer based on Qwen-Image that performs Real-ISR considering fidelity and controllability simultaneously: a newly introduced visual stream receives low-quality images (LQ) with adjustable noise (Control Noise), and the original visual stream receives LQs with consistent noise (Prior Noise), forming the Noise-hybrid Visual Stream (NVS) design. ODTSR further employs Fidelity-aware Adversarial Training (FAA) to enhance controllability and achieve one-step inference. Extensive experiments demonstrate that ODTSR not only achieves state-of-the-art (SOTA) performance on generic Real-ISR, but also enables prompt controllability on challenging scenarios such as real-world scene text image super-resolution (STISR) of Chinese characters without training on specific datasets.

Real-image super-resolution (Real-ISR) seeks to recover HR images from LR inputs with mixed, unknown degradations. While diffusion models surpass GANs in perceptual quality, they under-reconstruct high-frequency (HF) details due to a low-frequency (LF) bias and a depth-wise “low-first, high-later” hierarchy. We introduce FRAMER, a plug-and-play training scheme that exploits diffusion priors without changing the backbone or inference. At each denoising step, the final-layer feature map teaches all intermediate layers. Teacher and student feature maps are decomposed into LF/HF bands via FFT masks to align supervision with the model’s internal frequency hierarchy. For LF, an Intra Contrastive Loss (IntraCL) stabilizes globally shared structure. For HF, an Inter Contrastive Loss (InterCL) sharpens instance-specific details using random-layer and in-batch negatives. Two adaptive modulators, Frequency-based Adaptive Weight (FAW) and Frequency-based Alignment Modulation (FAM), reweight per-layer LF/HF signals and gate distillation by current similarity. Across U-Net and DiT backbones (e.g., Stable Diffusion 2, 3), FRAMER consistently improves PSNR/SSIM and perceptual metrics (LPIPS, NIQE, MANIQA, MUSIQ). Ablations validate the final-layer teacher and random-layer negatives. The code and project page will be publicly released.

Diffusion-based methods have shown great promise in single image super-resolution (SISR); however, existing approaches often produce blurred fine details due to insufficient guidance in the high-frequency domain. To address this issue, we propose a High-Frequency Guided Diffusion Network based on Wavelet Decomposition (HDW-SR), which replaces the conventional U-Net backbone in diffusion frameworks. Specifically, we perform diffusion only on the residual map, allowing the network to focus more effectively on high-frequency information restoration. We then introduce wavelet-based downsampling in place of standard CNN downsampling to achieve multi-scale frequency decomposition, enabling sparse cross-attention between the high-frequency subbands of the pre-super-resolved image and the low-frequency subbands of the diffused image for explicit high-frequency guidance. Moreover, a Dynamic Thresholding Block (DTB) is designed to refine high-frequency selection during the sparse attention process. During upsampling, the invertibility of the wavelet transform ensures low-loss feature reconstruction. Experiments on both synthetic and real-world datasets demonstrate that HDW-SR achieves competitive super-resolution performance, excelling particularly in recovering fine-grained image details. The code will be available after acceptance.

Text-conditioned human motion in-betweening leverages keyframes for spatio-temporal control, with text providing high-level semantic guidance for the transitions. However, existing methods are unable to establish a coherent alignment between textual semantics and the spatio-temporal constraints provided by keyframes, often resulting in insufficiently constrained motions with unintended behavior. Moreover, they struggle with precise spatial control, often generating motions that deviate from keyframe constraints. To address these issues, we propose a multi-level diffusion framework that integrates textual semantics with implicit cues from keyframe sequences to modulate global motion dynamics, while leveraging individual keyframes to guide local transitions around them. During inference, to ensure strict keyframe adherence, we propose a novel trajectory refinement strategy that adjusts the root positions of the generated motion, followed by diffusion imputation to refine the poses of the generated keyframes. Additionally, our framework enables semantics-preserving motion editing, allowing for plausible modifications while retaining the original motion semantics. Extensive experiments demonstrate that our method generates high-quality motions that strictly satisfy keyframe constraints while achieving precise semantic alignment.

State-of-the-art whole-body pose estimators often lack robustness, producing anatomically implausible predictions in challenging scenes. We posit this failure stems from spurious correlations learned from visual context, a problem we formalize using a Structural Causal Model (SCM). The SCM identifies visual context as a confounder that creates a non-causal backdoor path, corrupting the model's reasoning. We introduce the Causal Intervention Graph Pose (CIGPose) framework to address this by approximating the true causal effect between visual evidence and pose. The core of CIGPose is a novel Causal Intervention Module: it first identifies confounded keypoint representations via predictive uncertainty and then replaces them with learned, context-invariant canonical embeddings. These deconfounded embeddings are processed by a hierarchical graph neural network that reasons over the human skeleton at both local and global semantic levels to enforce anatomical plausibility. Extensive experiments show CIGPose achieves a new state-of-the-art on COCO-WholeBody. Notably, our CIGPose-x model achieves 67.0\% AP, surpassing prior methods that rely on extra training data. With the additional UBody dataset, CIGPose-x is further boosted to 67.5\% AP, demonstrating superior robustness and data efficiency.

We present Pressure2Motion, a novel motion capture algorithm that reconstructs human motion from a ground pressure sequence and text prompt. At inference time, Pressure2Motion requires only a pressure mat, eliminating the need for specialized lighting setups, cameras, or wearable devices, making it suitable for privacy-preserving, low-light, and low-cost motion capture scenarios. Such a task is severely ill-posed due to the indeterminacy of pressure signals with respect to full-body motion. To address this issue, we introduce Pressure2Motion, a generative model that leverages pressure features as input and utilizes a text prompt as a high-level guiding constraint to resolve ambiguities. Specifically, our model adopts a dual-level feature extractor to accurately interpret pressure data, followed by a hierarchical diffusion model that discerns broad-scale movement trajectories and subtle posture adjustments. Both the physical cues gained from the pressure sequence and the semantic guidance derived from descriptive texts are leveraged to guide the motion estimation with precision. To the best of our knowledge, Pressure2Motion is a pioneering work in leveraging both pressure data and linguistic priors for motion reconstruction, and the established MPL benchmark is the first benchmark for this novel motion capture task. Experiments show that our method generates high-fidelity, physically plausible motions, establishing a new state of the art for this task. The codes and benchmarks will be publicly released upon publication.

We present BioCoach, a biomechanics-grounded vision–language framework for fitness coaching from streaming video. BioCoach fuses two signals, visual appearance and 3D skeletal kinematics, through a novel three-stage pipeline: an exercise-specific degree-of-freedom selector that focuses analysis on salient joints; a structured biomechanical context that pairs individualized morphometrics with cycle and constraint analysis; and a vision–biomechanics conditioned feedback module that applies cross-attention to generate precise, actionable text. Using parameter-efficient training that freezes the vision and language backbones, BioCoach yields transparent, personalized reasoning rather than pattern matching. To enable learning and fair evaluation, we augment QEVD-fit-coach with biomechanics-oriented feedback to create QEVD-bio-fit-coach, and we introduce a biomechanics-aware LLM judge metric. BioCoach delivers clear gains on QEVD-bio-fit-coach across lexical and judgment metrics while maintaining temporal triggering; on the original QEVD-fit-coach, it improves text quality and correctness with near-parity timing, demonstrating that explicit kinematics and constraints are key to accurate, phase-aware coaching.

Humans rarely plan whole-body interactions with objects at the level of explicit whole-body movements. High-level intentions, such as affordance, define the goal, while coordinated balance, contact, and manipulation can emerge naturally from underlying physical and motor priors. Scaling such priors is key to enabling humanoids to compose and generalize loco-manipulation skills across diverse contexts while maintaining physically coherent whole-body coordination. To this end, we introduce InterPrior, a scalable framework that learns a unified control policy, i.e., interaction motion prior through large-scale imitation pretraining and post-training by reinforcement learning. InterPrior first distills a full-reference imitation expert into a versatile, goal-conditioned variational policy that reconstructs motion from multi-modal and partially specified goal cues. A targeted diversity process, combining data augmentation and physical perturbations, broadens exposure to varied contact and object conditions, producing a motion prior that generalizes beyond the training data. To address the vast configuration space of large-scale human-object interaction, a reinforcement learning finetuning enhances unseen goal competence, enabling recovery from unsuccessful grasp. The resulting policy acts as a reusable motion prior that can absorb new behaviors, including interactions with unseen objects. We also show its effectiveness in user-interactive control and across different embodiments.

Diffusion-based motion generation has advanced rapidly, but current methods still struggle with long-horizon consistency, style control, and multi-condition guidance. A major reason is the fused-conditioning design, where semantic, stylistic, and temporal signals share a single pathway, causing interference and limiting controllability.We propose MoCoDiff, a controlable autoregressive diffusion framework that introduces Injection Modulation Controllers (IMC). IMC is a lightweight, modality-specific linear modulation modules that inject text, style, and history signals through separate conditioning paths. IMC preserves the simplicity of a frozen backbone while avoiding the entanglement inherent to fused conditioning, enabling more stable and interpretable multi-condition control.To further enhance long-range synthesis, we develop a controllable autoregressive diffusion model equipped with Temporal IMC (TIMC), which applies history as a timestep-dependent corrective signal. This controllable formulation actively suppresses drift, enforces smooth transitions across motion segments, and significantly improves temporal coherence over extended sequences.Experiments show that MoCoDiff achieves state-of-the-art style fidelity, transition quality, and efficiency, while supporting flexible and interpretable multi-condition motion synthesis without retraining.

Pedestrian trajectory prediction is crucial for applications such as autonomous driving and social robots. Recently, language model (LM)–based trajectory prediction has offered both prediction accuracy and interpretability. However, the L2 loss commonly used in trajectory prediction cannot be directly applied to LM optimization, resulting in degraded prediction performance. Moreover, current LM-based trajectory prediction methods lack explicit expressions of social interactions, and their scene descriptions are overly simplistic, making it challenging to impose practical scene constraints. To address these issues, we propose Write-to-Walk (W2W). First, we construct a pedestrian trajectory dataset with explicit interaction semantics and generate parsable prompts based on observed trajectories and interaction cues (companion/following/obstacle), alleviating the lack of interaction semantics in prompts. Afterward, a T5-Small backbone is trained in a two-stage manner: (1) Full-parameter supervised fine-tuning with cross-entropy loss for language learning, endowing W2W with the capability for formatted question answering; (2) Reinforcement Learning (RL) to optimize W2W, where a reward function that integrates ADE error and off-road penalties strengthens scene constraints, producing future trajectories consistent with the scene context and further improving prediction accuracy. Experiments on the benchmarking datasets (ETH/UCY and SDD) demonstrate that W2W outperforms LM-based prediction methods on ADE/FDE metrics and achieves competitive results compared with SOTA trajectory prediction methods. Meanwhile, the interpretability of LMs further enhances W2W’s prospects for deployment in safety-critical domains, such as autonomous driving.

Text-to-motion synthesis aims to generate natural and expressive human motions from textual descriptions. While existing approaches primarily focus on generating holistic motions from text descriptions, they struggle to accurately reflect actions involving specific body parts. Recent part-wise motion generation methods attempt to resolve this but face two critical limitations: (i) they lack explicit mechanisms for aligning textual semantics with individual body parts, and (ii) they often generate incoherent full-body motions due to integrating independently generated part motions. To overcome these issues and resolve the fundamental trade-off in existing methods, we propose ParTY, a novel framework that enhances part expressiveness while generating coherent full-body motions. ParTY comprises: (1) Part-Guided Network, which first generates part motions to obtain part guidance, then uses it to generate holistic motions; (2) Part-aware Text Grounding, which diversely transforms text embeddings and appropriately aligns them with each body part; and (3) Holistic-Part Fusion, which adaptively fuses holistic motions and part motions. Extensive experiments, including part-level and coherence-level evaluations, demonstrate that ParTY achieves substantial improvements over previous methods.

Generating realistic human-human interactions is a challenging task that requires not only high-quality individual body and hand motions, but also coherent coordination among all interactants. Due to limitations in available data and increased learning complexity, previous methods tend to ignore hand motions, limiting the realism and expressivity of the interactions. Additionally, current diffusion-based approaches generate entire motion sequences simultaneously, limiting their ability to capture the reactive and adaptive nature of human interactions. To address these limitations, we introduce Interact2Ar, the first end-to-end text-conditioned autoregressive diffusion model for generating full-body, human-human interactions. Interact2Ar incorporates detailed hand kinematics through dedicated parallel branches, enabling high-fidelity full-body generation. Furthermore, we introduce an autoregressive pipeline coupled with a novel memory technique that facilitates adaptation to the inherent variability of human interactions using efficient large context windows. The adaptability of our model enables a series of downstream applications, including temporal motion composition, real-time adaptation to disturbances, and extension beyond dyadic to multi-person scenarios. To validate the generated motions, we introduce a set of robust evaluators and extended metrics designed specifically for assessing full-body interactions. Through quantitative and qualitative experiments, we demonstrate the state-of-the-art performance of Interact2Ar.

Generative models excel at motion synthesis for a fixed number of agents but struggle to generalize with variable agents. Based on limited, domain-specific data, existing methods employ autoregressive models to generate motion recursively, which suffer from inefficiency and error accumulation. We propose Unified Motion Flow (UMF), which consists of Pyramid Motion Flow (P-Flow) and Semi-Noise Motion Flow (S-Flow). UMF decomposes the number-free motion generation into a single-pass motion prior generation stage and multi-pass reaction generation stages. Specifically, UMF utilizes a unified latent space to bridge the distribution gap between heterogeneous motion datasets, enabling effective unified training. For motion prior generation, P-Flow operates on hierarchical resolutions conditioned on different noise levels, thereby mitigating computational overheads. For reaction generation, S-Flow learns a joint probabilistic path that adaptively performs reaction transformation and context reconstruction, alleviating error accumulation. Extensive results and user studies demonstrate UMF’s effectiveness as a generalist model for multi-person motion generation from text. We will release the code.

Reconstructing the three-dimensional distribution of dark matter from weak-lensing observations is a central but highly ill-posed inverse problem in cosmology. Unlike standard 3D reconstruction with multiple viewpoints, we observe the universe from a single line of sight, through noisy shape distortions of galaxies with uncertain distances, so meaningful recovery of the 3D matter field requires strong prior assumptions. Existing methods either produce point estimates with handcrafted priors or use neural ensembles for approximate Bayesian uncertainty, and struggle to capture the non-Gaussian, filamentary structure of the cosmic web. With the advent of new high-resolution cosmological simulations, we now have an alternative source of prior knowledge that captures the nonlinear statistics of structure formation with far greater fidelity than analytic prescriptions. We leverage these simulations to build a new dataset \texttt{Conicus3D}, which enables us to learn a data-driven diffusion-model prior capturing the full 3D distribution of dark matter structure across cosmic time. Building on recent plug-and-play approaches, we modify a diffusion-based posterior sampling scheme to the 3D weak-lensing setting, combining the learned prior with a differentiable physical forward model. On realistic simulations targeting a modern weak lensing survey, our approach yields substantially improved 2D and 3D reconstruction accuracy over baseline methods. Moreover, it produces posterior samples whose statistics closely track the underlying simulations, while remaining robust to moderate shifts in cosmology.

Recently, synthetic palmprints have been increasingly used as substitutes for real data to train recognition models. To be effective, such synthetic data must reflect the diversity of real palmprints, including both style variation and geometric variation. However, existing palmprint generation methods mainly focus on style translation, while geometric variation is either ignored or approximated by simple handcrafted augmentations. In this work, we propose FlowPalm, an optical-flow-driven palmprint generation framework capable of simulating the complex non-rigid deformations observed in real palms. Specifically, FlowPalm estimates optical flows between real palmprint pairs to capture the statistical patterns of geometric deformations. Building on these priors, we design a progressive sampling process that gradually introduces the geometric deformations during diffusion while maintaining identity consistency. Extensive experiments on six benchmark datasets demonstrate that FlowPalm significantly outperforms state-of-the-art palmprint generation approaches in downstream recognition tasks. Notably, FlowPalm achieves a higher TAR at FAR=1e-4 than the best generative model does at FAR=1e-3.

Robotic manipulation from visual observations remains challenging due to the lack of 3D consistent representations that can generalize across diverse viewpoints and sensor configurations. Existing approaches often rely on masked autoencoders or neural scene representations, which fail to capture cross view correspondences. Crucially, while multi-view diffusion models have recently shown tremendous success in 3D aware generative synthesis, their powerful representations offer a promising direction for achieving viewpoint robust visuomotor control. In this paper, we introduce DiffuView, a novel framework that learns unified 3D aware representations through multi-view diffusion pretraining and deploys them for imitation learning. Specifically, DiffuView models the conditional generation of target views given source observations within a diffusion framework, enabling the network to implicitly recover scene geometry and enforce view consistency. The pretrained diffusion network is then utilized as a powerful visual backbone for an action policy, allowing robust control under varying viewpoints and visual conditions. We evaluate DiffuView on two challenging benchmarks, MetaWorld and Libero. Extensive experiments in both simulation and realworld scenarios demonstrate that DiffuView achieves superior generalization, improving success rates under viewpoint shifts by nearly 20\% compared with existing methods.

Diffusion Transformers (DiTs) have greatly advanced text-to-image generation, but models still struggle to generate the correct spatial relations between objects as specified in the text prompt. In this study, we adopt a mechanistic interpretability approach to investigate how a DiT can generate correct spatial relations between objects. We train, from scratch, DiTs of different sizes with different text encoders to learn to generate images containing two objects whose attributes and spatial relations are specified in the text prompt. We find that, although all the models can learn this task to near-perfect accuracy, the underlying mechanisms differ drastically depending on the choice of text encoder. When using random text embeddings, we find that the spatial-relation information is passed to image tokens through a two-stage circuit, involving two cross-attention heads that separately read the spatial relation and single-object attributes in the text prompt. When using a pretrained text encoder (T5), we find that the DiT uses a different circuit that leverages information fusion in the text tokens, reading spatial-relation and single-object information together from a single text token. We further show that, although the in-domain performance is similar for the two settings, their robustness to out-of-domain perturbations differs, potentially suggesting the difficulty of generating correct relations in real-world scenarios.

Ill-posed inverse problems are fundamental in many domains, ranging from astrophysics to medical imaging. Emerging diffusion models provide a powerful prior for solving these problems. Existing maximum-a-posteriori (MAP) or posterior sampling approaches, however, rely on different computational approximations, leading to inaccurate or suboptimal samples. To address this issue, we introduce a new approach to solving MAP problems with diffusion model priors using a dual ascent optimization framework. Our framework achieves better image quality as measured by various metrics for image restoration problems, it is more robust to high levels of measurement noise, it is faster, and it estimates solutions that represent the observations more faithfully than the state of the art.

Diffusion Transformer (DiT) models have achieved unprecedented quality in image and video generation, yet their iterative sampling process remains computationally prohibitive. To accelerate inference, feature caching methods have emerged by reusing intermediate representations across timesteps. However, existing caching approaches treat all feature components uniformly. We reveal that DiT feature spaces contain distinct principal and residual subspaces with divergent temporal behavior: the principal subspace evolves smoothly and predictably, while the residual subspace exhibits volatile, low-energy oscillations that resist accurate prediction. Building on this insight, we propose SVD-Cache, a subspace-aware caching framework that decomposes diffusion features via Singular Value Decomposition (SVD), applies exponential moving average (EMA) prediction to the dominant low-rank components, and directly reuses the residual subspace. Extensive experiments demonstrate that SVD-Cache achieves near-lossless across diverse models and methods, including 5.55$\times$ speedup on FLUX and HunyuanVideo, and compatibility with model acceleration techniques including distillation, quantization and sparse attention. Our code is in supplementary material and will be released on Github.

Pan-sharpening, a fundamental image preprocessing technique in remote sensing, aims to generate spatially and spectrally enriched multispectral imagery by integrating complementary information from texture-rich panchromatic (PAN) images and paired low-resolution multispectral (LRMS) counterparts. Although recent generative diffusion models have achieved impressive fusion quality, these performance gains often come with substantial computational costs, rendering them impractical for resource-constrained scenarios common in remote sensing applications. This work introduces a function-space diffusion model built upon a neural operator architecture that achieves compelling performance with promising efficiency. Specifically, our framework replaces the standard attention-based denoising backbone with a Galerkin-type neural operator, significantly reducing computational complexity while maintaining excellent representational capacity. Furthermore, by explicitly integrating pixel-wise spatial-spectral consistency residuals into each reverse diffusion step, our method establishes a fine-grained, closed-loop guidance mechanism that dynamically calibrates spatial details and spectral fidelity throughout the generation process. Extensive experiments on multiple benchmark datasets demonstrate the effectiveness of our approach over state-of-the-art methods.

Ocean dynamics drive global climate patterns and extreme weather events, making accurate spatiotemporal forecasting essential for climate monitoring and marine operations. Traditional Global Ocean Forecasting Systems (GOFSs) offer high accuracy predictions, yet remain computationally expensive and fail to fully leverage growing historical data. Recent deep learning models have achieved notable success, but still face three fundamental challenges: (1) they homogenize ocean variables despite strong physical coupling via equation-of-state relationships; (2) they neglect spherical geometry, resulting in severe distortions at high latitudes; and (3) they struggle to model multi-scale temporal dynamics. We introduce PhyOceanCast, a physics-informed diffusion model that overcomes these limitations through two key innovations. First, the Spherical Graph Attention Network for Multi-scale Ocean Coupling (SGAN-MOC) preserves spherical topology while enabling cross-variable interactions via heterogeneous encoding and k-hop-constrained attention. Second, the Physics-Informed Wavelet Temporal Coherence (PWTC) module that decomposes ocean dynamics across multiple scales with advection-diffusion constraints. PhyOceanCast forecasts 145 ocean variables, including temperature, salinity, and velocity fields, across 36 depth levels plus sea surface height. Extensive experiments demonstrate superior performance over diffusion, transformer, and hybrid baselines, promising a new paradigm for global ocean canonical variable forecasting. Code is available at supplementary materials.

We present DAWN (Diffusion is All We Need for robot control), a unified diffusion-based framework for language-conditioned robotic manipulation that bridges high-level motion intent and low-level robot action via structured pixel motion representation. In DAWN, both the high-level and low-level controllers are modeled as diffusion processes, yielding a fully trainable, end-to-end system with interpretable intermediate motion abstractions. DAWN achieves state-of-the-art results on the challenging CALVIN benchmark, demonstrating strong multi-task performance, and further validates its effectiveness on MetaWorld. Despite the substantial domain gap between simulation and reality and limited real-world data, we demonstrate reliable real-world transfer with only minimal finetuning, illustrating the practical viability of diffusion-based motion abstractions for robotic control. Our results show the effectiveness of combining diffusion modeling with motion-centric representations as a strong baseline for scalable and robust robot learning. Visualization page: https://anonymous.4open.science/w/DAWN

Large Vision-Language Models (LVLMs) excel at captioning, visual question answering, and robotics by combining vision and language, yet they often miss obvious objects or hallucinate nonexistent ones in atypical scenes. We examine these failures through the lens of uncertainty, focusing on contextual incongruity, where objects appear unexpectedly or fail to appear in expected contexts, and show that such cases increase recognition difficulty for state-of-the-art LVLMs. To study this regime, we introduce the Object Recognition in Incongruous Context (ORIC) framework, which constructs incongruous object-context pairs through two complementary strategies: (1) LLM-guided sampling to identify hard-to-recognize objects present in the image and (2) CLIP-guided sampling to mine plausible but absent ones. Applied to MSCOCO, ORIC produces ORIC-Bench and ORIC-style training data. Evaluating 18 LVLMs and 2 open-vocabulary detectors reveals substantial performance drops and bias patterns under incongruous contexts. Fine-tuning Qwen3-VL-8B-Instruct with Visual Reinforcement Fine-Tuning on 600 ORIC-style samples improves results on ORIC-Bench, AMBER, and HallusionBench. Overall, we show that contextual incongruity is a key source of uncertainty and provide tools for more reliable LVLMs.

Document QA requires not only accurate answers but also identifying where each answer is grounded on the page. Most models treat the task as text-only generation, while existing answer grounding methods generate coarse bounding boxes that fail to capture curved text. We introduce M3Grounder, a hybrid vision–language and segmentation architecture that formulates document grounding as pixel-level segmentation. It produces fine-grained evidence masks refined by a bleed-suppression loss to prevent spillover. M3Grounder autoregressively generates answer text interleaved with [GROUND] tokens that link individual answer spans to their corresponding evidence regions. Also, M3Grounder grounds evidence hierarchically across phrase, line, and block levels using an enclosure loss that enforces spatial containment. We release GroundingDocQA dataset (200K documents, 2M multi-span and multi-granular QA pairs with pixel-level grounding masks), built through a data engine that handles complex layouts, curved-text, and graphics-rich documents. We also release GroundingDocQA-Bench, a diverse and challenging human-verified benchmark. M3Grounder sets a new state of the art in grounded DocVQA, advancing from coarse boxes to hierarchical, fine-grained and contextually grounded evidence.

Early children's developmental trajectories set up a natural goal for sample-efficient pretraining of vision foundation models. We introduce BabyVLM-V2, a developmentally grounded framework for infant-inspired vision–language modeling that extensively improves upon BabyVLM-V1 through a longitudinal, multifaceted pretraining set, a versatile model, and, most importantly, DevCV Toolbox for cognitive evaluation. The pretraining set maximizes coverage while minimizing curation of a longitudinal, infant-centric audiovisual corpus, yielding video-utterance, image-utterance, and multi-turn conversational data that mirror infant experiences. DevCV Toolbox adapts all vision-related measures of the recently released NIH Baby Toolbox into a benchmark suite of ten multimodal tasks, covering spatial reasoning, memory, and vocabulary understanding aligned with early children's capabilities. Experimental results show that a compact model pretrained from scratch can achieve competitive performance on DevCV Toolbox, outperforming GPT-4o on some tasks. We hope the principled, unified BabyVLM-V2 framework will accelerate research in developmentally plausible pretraining of vision foundation models.

Document parsing has recently advanced with multimodal large language models (MLLMs) that directly map document images to structured outputs. Traditional cascaded pipelines depend on precise layout analysis and often fail under casually captured or non-standard conditions. Although end-to-end approaches mitigate this dependency, they still exhibit repetitive, hallucinated, and structurally inconsistent predictions—primarily due to the scarcity of large-scale, high-quality full-page (document-level) end-to-end parsing data and the lack of structure-aware training strategies. To address these challenges, we propose a data–training co-design framework for robust end-to-end document parsing. A Realistic Scene Synthesis strategy constructs large-scale, structurally diverse full-page end-to-end supervision by composing layout templates with rich document elements, while a Document-Aware Training Recipe introduces progressive learning and structure-token optimization to enhance structural fidelity and decoding stability. We further build Wild-OmniDocBench, a benchmark derived from real-world captured documents for robustness evaluation. Integrated into a 1B-parameter MLLM, our method achieves superior accuracy and robustness across both scanned/digital and real-world captured scenarios. All models, data synthesis pipelines, and benchmarks will be publicly released to advance future research in document understanding.

Understanding mid-level road semantics—the structural and contextual cues that bridge low-level perception and high-level planning—is essential for reliable autonomous driving and digital map construction. However, existing benchmarks primarily target perception tasks such as detection or segmentation, overlooking the reasoning capabilities required to infer road topology and dynamic scene structure. To address this gap, we present RoadSceneBench, a lightweight yet information-rich benchmark designed to evaluate and advance visual reasoning in complex road environments. Unlike large-scale perception datasets, RoadSceneBench emphasizes relational understanding and structural consistency, encouraging models to capture the underlying logic of real-world road scenes. Furthermore, to enhance reasoning reliability, we propose Hierarchical Relational Reward Propagation with Temporal Consistency (HRRP-T)—a training framework for Vision-Language Models (VLMs) in which reward signals adaptively promote spatial coherence and semantic alignment throughout the reasoning process. This paradigm enables models to move beyond static recognition toward geometry-aware and temporally consistent reasoning. Extensive experiments demonstrate that our method achieves state-of-the-art performance across diverse road configurations. RoadSceneBench thus provides a compact yet powerful foundation for studying mid-level road semantics and fostering structure-aware autonomous perception. Our dataset will be made publicly available in the final version.

Counting is a core capability for multimodal large language models (MLLMs), yet there is no unified counting dataset to rigorously evaluate this ability across image, text, and audio. We present UNICBench, a unified multimodal, multi‑level counting benchmark and evaluation toolkit with accurate ground truth, deterministic numeric parsing, and stratified reporting. The corpus comprises 5,300 images (5,508 QA), 872 documents (5,888 QA), and 2,069 audio clips (2,905 QA), annotated with a three‑level capability taxonomy and difficulty tags. Under a standardized protocol with fixed splits/prompts/seeds and modality‑specific matching rules, we evaluate 45 state‑of‑the‑art MLLMs across modalities. Results show strong performance on some basic counting tasks but significant gaps on reasoning and the hardest partitions, highlighting long‑tail errors and substantial headroom for improving general counting. UNICBench offers a rigorous and comparable basis for measurement and a public toolkit to accelerate progress.

Image captions serve as efficient surrogates for visual content in multimodal systems such as retrieval, recommendation, multi-step agentic inference pipelines. Yet current evaluation practices miss a fundamental question: Can captions stand-in for images in real downstream tasks? We propose a utility-based benchmark, CaptionQA, to evaluate model-generated captions, where caption quality is measured by how well it supports downstream tasks. CaptionQA is an extensible domain-dependent benchmark covering 4 domains, Natural, Document, E-commerce, and Embodied AI, each with fine-grained taxonomies (25 top-level and 69 subcategories) that identify useful information for domain-specific tasks. CaptionQA builds 33,027 densely annotated multiple-choice questions (50.3 per image on average) that explicitly require visual information to answer, providing a comprehensive probe of caption utility. In our evaluation protocol, an LLM answers these questions using captions alone, directly measuring whether captions preserve image-level utility and are utilizable by a downstream LLM. Evaluating state-of-the-art MLLMs reveals substantial gaps between the image and its caption utility. Notably, models nearly identical on traditional image-QA benchmarks lower by up to 32\% in caption utility. We release CaptionQA along with an open-source pipeline for extension to new domains.

Humans constantly reason about 3D proximity, the relations between their body and surrounding objects, to guide perception and action in daily life. Whether multimodal large language models (MLLMs) can perform such embodied 3D reasoning remains unclear. To this end, we introduce EgoProx, a benchmark for egocentric 3D proximity reasoning. We organize our tasks along a cognitive chain, covering intention, exploration, exploitation, and chain-of-actions reasoning. We also design an agent based data engine that produces diverse and consistent QA pairs at scale. We benchmark prevailing MLLMs on EgoProx and conduct additional analyses with dataset specific and task specific instruction tuning. We observe large cross-domain gains, indicating that current MLLMs contain some spatial knowledge; however, they still struggle to effectively leverage it for spatial reasoning VQA.

Despite the remarkable progress of Multimodal Large Language Models (MLLMs) in 2D vision-language tasks, their application to complex 3D scene manipulation remains underexplored. In this paper, we bridge this critical gap by tackling three key challenges in 3D object arrangement task using MLLMs. First, to address the weak visual grounding of MLLMs, which struggle to link programmatic edits with precise 3D outcomes, we introduce an MCP-based API. This shifts the interaction from brittle raw code manipulation to more robust, function-level updates. Second, we augment the MLLM's 3D scene understanding with a suite of specialized visual tools to analyze scene state, gather spatial information, and validate action outcomes. This perceptual feedback loop is critical for closing the gap between language-based updates and precise 3D-aware manipulation. Third, to manage the iterative, error-prone updates, we propose a collaborative multi-agent framework with designated roles for planning, execution, and verification. This decomposition allows the system to robustly handle multi-step instructions and recover from intermediate errors. We demonstrate the effectiveness of our approach on a diverse set of 25 complex object arrangement tasks, where it significantly outperforms existing baselines.

Understanding a 3D scene immediately with its exploration is essential for embodied tasks, where an agent must construct and comprehend the 3D representation in an online and nearly real-time manner. In this study, we propose EmbodiedSplat, an online feed-forward 3DGS for open-vocabulary scene understanding that enables simultaneous online 3D reconstruction and 3D semantic understanding from the streaming images. Unlike existing open-vocabulary 3DGS methods, our objectives are two-fold: 1) Reconstructs the semantic-embedded 3DGS of the entire scene from over 300 streaming images in an online manner. 2) Highly generalizable to novel scenes with feed-forward design and supports nearly real-time 3D semantic reconstruction when combined with real-time 2D models. To achieve these objectives, we propose an Online Sparse Coefficients Field with a CLIP Global Codebook where it binds the 2D CLIP embeddings to each 3D Gaussian while minimizing memory consumption and preserving the full semantic generalizability of CLIP. Furthermore, we generate 3D geometric-aware CLIP features by aggregating the partial point cloud of 3DGS through 3D U-Net to compensate the 3D geometric prior to 2D-oriented language embeddings. Our code will be publicly available on paper acceptance.

Large Multimodal Models (LMMs) that process 3D data typically rely on heavy, pre-trained visual encoders to extract geometric features. While recent 2D LMMs have begun to eliminate such encoders for efficiency and scalability, extending this paradigm to 3D remains challenging due to the unordered and large-scale nature of point clouds. This leaves a critical unanswered question: How can we design an LMM that tokenizes unordered 3D data effectively and efficiently without a cumbersome encoder? We introduce Fase3D, the first efficient encoder-free Fourier-based 3D LMM. Fase3D tackles the challenges of scalability and permutation invariance with a novel tokenizer that combines point cloud serialization and the Fast Fourier Transform (FFT) to approximate self-attention. This design enables an effective and computationally minimal architecture, built upon three key innovations: First, we represent large scenes compactly via structured superpoints. Second, our space-filling curve serialization followed by an FFT enables efficient global context modeling and graph-based token merging. Lastly, our Fourier-augmented LoRA adapters inject global frequency-aware interactions into the LLMs at a negligible cost. Fase3D achieves performance comparable to encoder-based 3D LMMs while being significantly more efficient in computation and parameters. We will release code and models publicly.

Multimodal Large Language Models (MLLMs) have significantly advanced vision–language understanding. However, even state-of-the-art models struggle with geometric reasoning, revealing a critical bottleneck: the extreme scarcity of high-quality image–text pairs. Human annotation is prohibitively expensive, while automated methods fail to ensure both fidelity and training effectiveness. Existing approaches either passively adapt to available images or employ inefficient random exploration with post-hoc filtering, decoupling generation from learning needs.We propose Socratic-Geo, a fully autonomous framework that dynamically couples data synthesis with model learning through multi-agent interaction. The Teacher agent generates parameterized Python scripts with reflective feedback (Reflect for solvability, RePI for visual validity), ensuring image–text pair purity. The Solver agent optimizes reasoning through preference learning, with failure paths guiding the Teacher’s targeted augmentation. Independently, the Generator learns image generation capabilities on accumulated “image–code–instruction” triplets, distilling programmatic drawing intelligence into visual generation.Starting from only 108 seed problems, Socratic-Solver achieves 49.11 on six geometric benchmarks using one-quarter of baseline data, surpassing strong baselines by 2.43 points. Socratic-Generator achieves 42.4% on GenExam, establishing new state-of-the-art for open-source models, surpassing Seedream-4.0 (39.8%) and approaching Gemini-2.5-Flash-Image (43.1%).

Humans commonly reason about object affordance through observed interactions in images or videos, and once formed, such knowledge can be generically generalized to novel objects. Inspired by this principle, we advocate for a novel framework that leverages emerging multimodal large language models (MLLMs) for interaction intention-driven 3D affordance grounding, namely HAMMER. Instead of generating explicit object attribute descriptions or relying on off-the-shelf 2D segmenters, we alternatively aggregate the interaction intention depicted in the reference image into a contact-aware embedding and guide the model to infer textual affordance labels, ensuring it thoroughly excavates object semantics and contextual cues. We further devise a hierarchical cross-modal integration mechanism to fully exploit the complementary information from the MLLM for 3D representation refinement and introduce a multi-granular geometry lifting module that infuses spatial characteristics into the extracted intention embedding, thus facilitating accurate 3D affordance localization. Extensive experiments on public datasets and our newly constructed corrupted benchmark reveal the superiority and robustness of our method in seen and unseen scenarios compared to existing approaches. The code and models will be made publicly available.

Spatial intelligence in vision-language models (VLMs) attracts research interest with the practical demand to reason in the 3D world. Despite promising results, most existing methods follow the conventional 2D pipeline in VLMs and use pixel-aligned representations for the vision modality. However, correspondence-based models with implicit 3D scene understanding often fail to achieve spatial consistency, and representation-based models with 3D geometric priors lack efficiency in vision sequence serialization. To address this, we propose a Proxy3D method with compact yet comprehensive 3D proxy representations for the vision modality. Given only video frames as input, we employ semantic and geometric encoders to extract scene features and then perform their semantic-aware clustering to obtain a set of proxies in the 3D space. For representation alignment, we further curate the SpaceSpan dataset and apply multi-stage training to adopt the proposed 3D proxy representations with the VLM. When using shorter sequences for vision information, our method achieves competitive or state-of-the-art performance in 3D visual question answering, visual grounding and general spatial intelligence benchmarks.

Achieving unified 3D perception and reasoning across tasks such as segmentation, retrieval, and relation understanding remains challenging, as existing methods are either object-centric or rely on costly training for inter-object reasoning. We present a novel framework that constructs a hierarchical language-distilled Gaussian scene and its 3D semantic scene graph without scene-specific training. A Gaussian pruning mechanism refines scene geometry, while a robust multi-view language alignment strategy aggregates noisy 2D features into accurate 3D object embeddings. On top of this hierarchy, we build an open-vocabulary 3D scene graph with Vision Language-derived annotations and Graph Neural Network-based relational reasoning. Our approach enables efficient and scalable open-vocabulary 3D reasoning by jointly modeling hierarchical semantics and inter/intra-object relationships, validated across tasks including open-vocabulary segmentation, scene graph generation, and relation-guided retrieval.

Leveraging 3D information within Multimodal Large Language Models (MLLMs) has recently shown significant advantages for indoor scene understanding. However, existing methods, including those using explicit ground-truth 3D positional encoding and those grafting external 3D foundation models for implicit geometry, struggle with the trade-off in 2D-3D representation fusion, leading to suboptimal deployment. To this end, we propose 3D-Implicit Depth Emergence, a method that reframes 3D perception as an emergent property derived from geometric self-supervision rather than explicit encoding. Our core insight is the Implicit Geometric Emergence Principle: by strategically leveraging privileged geometric supervision through mechanisms like a fine-grained geometry validator and global representation constraints, we construct an information bottleneck. This bottleneck forces the model to maximize the mutual information between visual features and 3D structures, allowing 3D awareness to emerge naturally within a unified visual representation. Unlike existing approaches, our method enables 3D perception to emerge implicitly, disentangling features in dense regions and, crucially, eliminating depth and pose dependencies during inference with zero latency overhead. This paradigm shift from external grafting to implicit emergence represents a fundamental rethinking of 3D knowledge integration in visual-language models. Extensive experiments demonstrate that our method surpasses SOTA on multiple 3D scene understanding benchmarks. Our approach achieves a 55\% reduction in inference latency while maintaining strong performance across diverse downstream tasks, underscoring the effectiveness of meticulously designed auxiliary objectives for dependency-free 3D understanding. Source code will be made publicly available upon acceptance.

Recent work in 3D scene understanding has begun to shift from purely spatial analysis to the more complex challenge of functional scene understanding.However, existing methods often consider functional relationships between object pairs in isolation, failing to capture the scene-wide interdependencies that humans use to resolve ambiguity. We introduce FunFact, a framework for constructing probabilistic open-vocabulary functional 3D scene graphs from posed RGB-D images. FunFact first builds an object- and part-centric 3D map and uses foundation models to propose semantically plausible functional relations. These candidates are converted into factor graph variables and constrained by both LLM-derived common-sense priors and geometric priors. This formulation enables joint probabilistic inference over all functional edges and their uncertainties, yielding substantially better-calibrated confidence scores. To benchmark this setting, we also introduce FunThor, a synthetic dataset based on AI2THOR with part-level geometry and systematically-defined rule-based functional annotations. Experiments on SceneFun3D, FunGraph3D, and FunThor show that FunFact improves node and relation discovery recall and significantly reduces calibration error for ambiguous relations, highlighting the benefits of holistic probabilistic modeling for functional scene understanding. We will release the code and dataset to facilitate future research.

In this paper, we present PSC-AVDN, a training-free framework for Aerial Vision-and-Dialog Navigation that integrates a three-stage Parsing-Search-Confirmation reasoning pipeline with a Structured Spatial Memory (SSM) module. The parsing stage converts ambiguous instructions into stable geometric cues, Search-CoT conducts stepwise high-altitude target exploration, and Confirmation-CoT performs fine-grained verification to resolve visual ambiguity and confirm the final target. Meanwhile, SSM integrates multi-scale visual observation, spatial visual memory, and structured geometric memory to provide global spatial context and long-horizon consistency.Extensive experiments on the AVDH and AVDH-Full datasets show that PSC-AVDN sets new state-of-the-art performance in the training-free setting, matching or surpassing several finetuned methods. We believe this framework offers a principled way to combine explicit CoT-style reasoning with structured spatial memory for scalable and generalizable aerial embodied navigation in the future.

Despite recent advances, Vision Language Models (VLMs) still struggle to grasp the dynamics of the world. We note that the ability to reason about 3D motion, challenging in itself, is further complicated by two factors. First, VLMs observe motion indirectly via its projection on 2D images. Second, existing datasets fail to disentangle object and camera motion. To address these, we present a QA generation pipeline that focuses on motion-related scene understanding. We take particular care of the entanglement of camera and object motion by casting tracking in both the traditional way and in a novel, fixed reference system, dubbed True-Motion Tracking, which provides an intuitive description of motion. From this pipeline, we generate large-scale 400K training samples and a 2.2K-sample benchmark. Training existing models on our dataset yields performance improvements on an external benchmark, validating the effectiveness of our method.

A fundamental challenge in embodied AI is verifying if agents build internal models of spatial structure or merely learn to mimic task-specific expert trajectories. This is critical as foundational approaches rooted in action-centric tasks (e.g., VLN) and reasoning-centric tasks (e.g., EQA) often share a common limitation: they lack a learning signal that forces them to encode fine-grained spatial relationships (like topology or distance) over long-range, fragmented experiences. To address this, we first propose LASAR, an architecture featuring a dual-memory system designed to maintain both episodic experiences and a semantic cognitive map. We then introduce Spatio-temporal Contextual Representation Learning (ST-CRL), a contrastive objective designed to train this architecture. ST-CRL leverages spatio-temporal cues from cognitive queries generated through annotated spatio-temporal context in simulation to build sample pairs, thereby forming the internal cognitive map from the agent's experiences. Experiments demonstrate that our method achieves 2\%-3.5\% gains in both zero-shot generalization on standard VLN-CE and VSI-Bench benchmarks. We also demonstrate that our proposed cognitive map has high self-consistency.

This paper studies unsupervised cross-domain image retrieval (UCDIR), which aims to retrieve images of the same category across different domains without relying on labeled data. Existing methods typically utilize pseudo-labels, derived from clustering algorithms, as supervisory signals for intra-domain representation learning and cross-domain feature alignment. However, these discrete pseudo-labels often fail to provide accurate and comprehensive semantic guidance. Moreover, the alignment process frequently overlooks the entanglement between domain-specific and semantic information, leading to semantic degradation in the learned representations and ultimately impairing retrieval performance. This paper addresses the limitations by proposing a Text-Phase Synergy Network with Dual Priors (TPSNet). Specifically, we first employ CLIP to generate a set of class-specific prompts per domain, termed as domain prompt, serving as a text prior that offers more precise semantic supervision. In parallel, we further introduce a phase prior, represented by domain-invariant phase features, which is integrated into the original image representations to bridge the domain distribution gaps while preserving semantic integrity. Leveraging the synergy of these dual priors, TPSNet significantly outperforms state-of-the-art methods on both the Office-Home and DomainNet datasets. Notably, when using ViT-B as the image encoder, TPSNet achieves an average $P@15$ improvement of 24.48\% on Office-Home and a 13.86\% improvement in $P@200$ on DomainNet. The code will be released.

Text-video retrieval tasks have seen significant improvements due to the recent development of large-scale vision-language pre-trained models. Traditional methods primarily focus on video representations or cross-modal alignment, while recent works shift toward enriching text expressiveness to better match the rich semantics in videos. However, these methods use only interactions between text and frames/video, and ignore rich interactions among the internal frames within a video, so the final expanded text cannot capture frame contextual information, leading to disparities between text and video. In response, we introduce Energy-Aware Fine-Grained Relationship Learning Network (EagleNet) to generate accurate and context-aware enriched text embeddings. Specifically, the proposed Fine-Grained Relationship Learning mechanism (FRL) first constructs a text-frame graph by the generated text candidates and frames, then learns relationships among texts and frames, which are finally used to aggregate text candidates into an enriched text embedding that incorporates frame contextual information. To further improve fine-grained relationship learning in FRL, we design Energy-Aware Matching (EAM) to model the energy of text-frame interactions and thus accurately capture the distribution of real text-video pairs. Moreover, for more effective cross-modal alignment and stable training, we replace the conventional softmax-based contrastive loss with the sigmoid loss. Extensive experiments have demonstrated the superiority of EagleNet across MSRVTT, DiDeMo, MSVD, and VATEX.

Table structure recognition in document AI faces significant challenges due to layout inconsistencies, merged cells, and complex nested structures, which is further exacerbated by the scarcity of large, diverse annotated datasets. In this paper, we present PIX-TAB (Efficient PIXel-Precise TABle Structure Recognition Approach) that provides exact, pixel-level structure using a small, lightweight model that can run on-device. The approach is language-agnostic, as it allows adding support for a new languages simply by replacing the Optical Character Recognition (OCR) model without modifying to the core structure recognition model. Key innovations include: position-aware pixel-precise tokens for deterministic cell reconstruction; speculative decoding for faster sequence generation, and training-only box supervision to stabilize spatial grounding; region-based image segmentation. To mitigate data scarcity we propose a pipeline for generating a large synthetic table dataset. Experimental results validate each component. To address the limitations of existing evaluation methods we introduce $TEDS_{struct100}$ and $TEDS_{100}$ metrics. Speculative decoding approach significantly improves recognition speed while maintaining accuracy. Finally, the combined techniques enable a mobile-optimized model that is more than three times faster than the full-size version.

The rapid proliferation of generative components, such as LoRAs, has created a vast but unstructured ecosystem. Existing discovery methods depend on unreliable user descriptions or biased popularity metrics, hindering usability. We present CARLoS, a large-scale framework for characterizing LoRAs without requiring additional metadata. Analyzing over 650 LoRAs, we employ them in image generation over a variety of prompts and seeds, as a credible way to assess their behavior. Using CLIP embeddings and their difference to a base-model generation, we concisely define a three part representation: \textit{Directions}, defining semantic shift; \textit{Strength}, quantifying the significance of the effect; and \textit{Consistency}, quantifying how stable the effect is. Using these representations, we develop an efficient retrieval framework that semantically matches textual queries to relevant LoRAs while filtering overly strong or unstable ones, outperforming textual baselines in automated and human evaluations. While retrieval is our primary focus, the same representation also supports analyses linking Strength and Consistency to legal notions of substantiality and volition, key considerations in copyright, positioning CARLoS as a practical system with broader relevance for LoRA analysis.

Camouflaged scene understanding (CSU) has attracted significant attention due to its broad practical implications. However, in this field, robust image-text cross-modal alignment remains under-explored, hindering deeper understanding of camouflaged scenarios and their related applications. To this end, we focus on the typical image-text retrieval task, and formulate a new task dubbed ``camouflage-aware image-text retrieval'' (CA-ITR). We first construct a dedicated camouflage image-text retrieval dataset (CamoIT), comprising $\sim$10.5K samples with multi-granularity textual annotations. Benchmark results conducted on CamoIT reveal the underlying challenges of CA-ITR for existing cutting-edge retrieval techniques, which are mainly caused by objects' camouflage properties as well as those complex image contents. As a solution, we propose a camouflage-expert collaborative network (CECNet), which features a dual-branch visual encoder: one branch captures holistic image representations, while the other incorporates a dedicated model to inject representations of camouflaged objects. A novel confidence-conditioned graph attention (C\textsuperscript{2}GA) mechanism is incorporated to exploit the complementarity across branches. Comparative experiments show that CECNet achieves a $\sim$29\% CA-ITR accuracy boost, surpassing seven representative retrieval models. Our dataset and code will be made publicly available.

In remote sensing (RS) cross-modal retrieval, most existing methods employ contrastive learning as their primary optimization objective, aligning anchors with positive counterparts and distinguishing them from negative samples. To improve negative sampling, these approaches typically set thresholds on cross-modal similarity scores, designating negatives that exceed the threshold as false negative samples (FNS). However, dependence on a single cross-modal similarity threshold is fragile because it fails to account for the cross-modal semantic overlaps and gaps. To address these challenges, we introduce TriSim, a novel image-text retrieval framework that constructs a tri-dimensional negative similarity space $<$img-img, img-txt, txt-txt$>$ to mitigate the influence of FNS issue. Specifically, considering that FNS appear as anomalies in this space, Extreme Value Theory (EVT) is applied to model the statistical behavior of the tail distribution for FNS selection. Two complementary tail selection strategies are developed: one identifies samples distant from the dense ellipsoidal center, and the other targets upper-right high-similarity extremes. The selected tail samples are regarded as FNS and modeled using a generalized Pareto distribution, with probabilistic weights assigned in the triplet loss. To further refine the selected FNS, intra-modal saliency differences are computed to generate masks that guide the learning of a gain matrix, which amplifies highly discriminative regions and suppresses ambiguous ones. Extensive experiments on two benchmarks demonstrate the superiority of the proposed TriSim framework in mitigating the influence of false negatives in RS image-text retrieval.

Many real-world applications in digital forensics, urban monitoring, and environmental analysis require jointly reasoning about visual appearance, geolocation, and time. Beyond standard geo-localization and time-of-capture prediction, these applications increasingly demand more complex capabilities, such as retrieving an image captured at the same location as a query image but at a specified target time. We formalize this problem as Geo-Time Aware Image Retrieval and curate a diverse benchmark of 4.5M paired image–location–time triplets for training and 85k high-quality triplets for evaluation. We then propose TIGeR, a multi-modal-transformer–based model that maps image, geolocation, and time into a unified geo-temporal embedding space. TIGeR supports flexible input configurations (single-modality and multi-modality queries) and uses the same representation to perform (i) geo-localization, (ii) time-of-capture prediction, and (iii) geo-time–aware retrieval. By better preserving underlying location identity under large appearance changes, TIGeR enables retrieval based on where and when a scene is, rather than purely on visual similarity. Extensive experiments show that TIGeR consistently outperforms strong baselines and state-of-the-art methods by up to 16% on time-of-year, 8% time-of-day prediction, and 14% in geo-time aware retrieval recall, highlighting the benefits of unified geo-temporal modeling.

We introduce Mistake Attribution (MATT), a task for fine-grained understanding of human mistakes in egocentric video. Unlike prior mistake understanding work, which lacks fine-grained output, MATT concretely attributes mistakes to the input instruction text or the attempt video.MATT determines what part of the instruction is violated (semantic role), when the deviation becomes irreversible (the Point-of-No-Return, PNR), and where the mistake appears in the PNR frame. We develop MisEngine, a data engine that automatically constructs attribution-rich mistake samples from existing datasets and inherits their annotations. Applied to large egocentric corpora, MisEngine yields EPIC-KITCHENS-M and Ego4D-M—two datasets that are up to two orders of magnitude larger than prior mistake datasets. We then present MisFormer, a unified attention-based model for mistake attribution across semantic (what), temporal (when), and spatial (where) dimensions, trained using MisEngine supervision. Experiments on our new datasets and prior benchmarks show that MisFormer outperforms strong video-language, temporal localization, hand-object interaction, and mistake-detection baselines.

The task of video geolocalization aims to determine the precise GPS coordinates of a video’s origin and map its trajectory; with applications in forensics, social media, and exploration. Existing classification-based approaches operate at a coarse city-level granularity and fail to capture fine-grained details, while image retrieval methods are impractical on a global scale due to the need for extensive image galleries which are infeasible to compile. Comparatively, constructing a gallery of GPS coordinates is straightforward and inexpensive. We propose VidTAG, a dual-encoder framework that performs frame-to-GPS retrieval using both self-supervised and language-aligned features. To address temporal inconsistencies in video predictions, we introduce the TempGeo module, which aligns frame embeddings, and the GeoRefiner module, an encoder-decoder architecture that refines GPS features using said aligned frame embeddings. Evaluations on Mapillary (MSLS) and GAMa datasets demonstrate our model’s ability to generate temporally consistent trajectories and outperform baselines, achieving a 20% improvement at the 1 km threshold over GeoCLIP. We also beat current State-of-the-Art by 25% on global coarse grained video geolocalization (CityGuessr68k). Our approach enables fine-grained video geolocalization and lays a strong foundation for future research. Code and models will be released publicly.

When obtaining visual illustrations from text descriptions, today’s methods take a description with a single text context—a caption, or an action description—and retrieve or generate the matching visual context. However, prior work does not permit visual illustration of multistep descriptions, e.g. a cooking recipe or a gardening instruction manual, and simply handling each step description in isolation would result in an incoherent demonstration. We propose Stitch-a-Demo, a novel retrieval-based method to assemble a video demonstration from a multistep description. The resulting video contains clips, possibly from different sources, that accurately reflect all the step descriptions, while being visually coherent. We formulate a training pipeline that creates large-scale weakly supervised data containing diverse procedures and injects hard negatives that promote both correctness and coherence. Validated on in-the-wild instructional videos, Stitch-a-Demo achieves state-of-the-art performance, with gains up to 29% as well as dramatic wins in a human preference study.

Egocentric Action Anticipation aims to infer future actions from videos, which is crucial for embodied AI systems. However, its advancement is hindered by the inherent stochasticity of the future, which introduces significant prediction uncertainty. Prevailing methods typically adopt an end-to-end approach to model holistic spatiotemporal contexts, yet they often lack explicit semantic reasoning capabilities, making it difficult to handle open-ended future uncertainties.To address these challenges, we propose a Prototypical Action Reasoning Framework Facilitated by Vision-Language Alignment (PAR-VLA), which leverages the semantic alignment capability of vision-language models to learn disentangled visual prototype for verbs and nouns. These prototypes serve as robust semantic anchors, transforming the unconstrained temporal prediction problem into a conditional forecasting task guided by well-defined semantic concepts. Our multi-stage framework first extracts visually-grounded and text-aligned prototype groups from a VLM, learning multiple prototypes per category to capture intra-class diversity. Subsequently, a novel Prototypical Action Reasoning-guided Verb-Noun Encoding branch dynamically retrieves the most relevant verb and noun concepts based on visual observations and explicitly models their interactions to guide temporal anticipation. Furthermore, we introduce Dual-Stream Symbiotic Predictive Decoders to more finely capture the interdependencies between verbs and nouns during the prediction process. Experiments Results demonstrate that PAR achieves state-of-the-art performance and exhibits a strong capability in dealing with future uncertainty.

Precise Event Spotting aims to localize fast-paced actions or events in videos with high temporal precision, a key task for applications in sports analytics, robotics, and autonomous systems. Existing methods typically process all frames uniformly, overlooking the inherent spatio-temporal redundancy in video data. This leads to redundant computation on non-informative regions while limiting overall efficiency. To remain tractable, they often spatially downsample inputs, losing fine-grained details crucial for precise localization. To address these limitations, we propose \textbf{AdaSpot}, a simple yet effective framework that processes low-resolution videos to extract global task-relevant features while adaptively selecting the most informative region-of-interest in each frame for high-resolution processing. The selection is performed via an unsupervised, task-aware strategy that maintains spatio-temporal consistency across frames and avoids the training instability of learnable alternatives. This design preserves essential fine-grained visual cues with a marginal computational overhead compared to low-resolution-only baselines, while remaining far more efficient than uniform high-resolution processing. Experiments on standard PES benchmarks demonstrate that AdaSpot achieves state-of-the-art performance under strict evaluation metrics (\eg, $+3.96$ and $+2.26$ mAP$@0$ frames on Tennis and FineDiving), while also maintaining strong results under looser metrics. Code will be released upon publication.

We introduce the ``single-life” learning paradigm, where we train a distinct vision model exclusively on egocentric videos captured by one individual. We leverage the multiple viewpoints naturally captured within a single life to learn a visual encoder in a self-supervised manner. Our experiments demonstrate three key findings. First, models trained independently on different lives develop a highly aligned geometric understanding. We demonstrate this by training visual encoders on distinct datasets each capturing a different life, both indoors and outdoors, as well as introducing a novel cross-attention-based metric to quantify the functional alignment of the internal representations developed by different models. Second, we show that single-life models learn generalizable geometric representations that effectively transfer to downstream tasks, such as depth estimation, in unseen environments. Third, we demonstrate that training on up to 30 hours from one week of the same person's life leads to comparable performance to training on 30 hours of diverse web data, highlighting the strength of single-life representation learning. Overall, our results establish that the shared structure of the world, both leads to consistency in models trained on individual lives, and provides a powerful signal for visual representation learning.

Despite rapid developments and widespread applications of MLLM agents, they still struggle with long-form video understanding (LVU) tasks, which are characterized by high information density and extended temporal spans. Recent research on LVU agents demonstrates that simple task decomposition and collaboration mechanisms are insufficient for long-chain reasoning tasks. Moreover, directly reducing the time context through embedding-based retrieval may lose key information of complex problems. In this paper, we propose Symphony, a multi-agent system, to alleviate these limitations. By emulating human cognition patterns, Symphony decomposes LVU into fine-grained subtasks and incorporates a deep reasoning collaboration mechanism enhanced by reflection, effectively improving the reasoning capability. Additionally, Symphony provides a VLM-based grounding approach to analyze LVU tasks and assess the relevance of video segments, which significantly enhances the ability to locate complex problems with implicit intentions and large temporal spans. Experimental results show that Symphony achieves state-of-the-art performance on LVBench, LongVideoBench, VideoMME, and MLVU, with a 5.0% improvement over the prior state-of-the-art method on LVBench.

Long-form video understanding remains challenging due to the extended temporal structure and dense multimodal cues. Despite recent progress, many existing approaches still rely on hand-crafted reasoning pipelines or employ token-consuming video preprocessing to guide MLLMs in autonomous reasoning. To overcome these limitations, we introduce VideoARM, an agentic reasoning-over-hierarchical-memory paradigm for long-form video understanding. Instead of static, exhaustive preprocessing, VideoARM performs adaptive, on-the-fly agentic reasoning and memory construction. Specifically, VideoARM performs an adaptive and continuous loop of observing, thinking, acting, and memorizing, where a Controller autonomously invokes tools to interpret the video in a coarse-to-fine manner, thereby substantially reducing token consumption. In parallel, a hierarchical multimodal memory continuously captures and updates multi-level clues throughout the operation of the agent, providing precise contextual information to support the Controller in decision-making. Experiments on prevalent benchmarks demonstrate that VideoARM outperforms the state-of-the-art method, DVD, while significantly reducing token consumption for long-form videos.

Frame selectoin is crucial due to high frame redundancy and limited context windows when applying Large Vision-Language Models (LVLMs) to long videos. Current methods typically select frames with high relevance to a given query, resulting a disjointed set of frames that disregard the narrative structure of video. In this paper, we introduce $\textbf{W}$avelet-based $\textbf{F}$rame $\textbf{S}$election by Detecting $\textbf{S}$emantic $\textbf{B}$oundary ($\textbf{WFS-SB}$), a training-free framework that presents a new perspective: effective video understanding hinges not only on high relevance but, more importantly, on capturing semantic shifts—pivotal moments of narrative change that are essential to comprehending the holistic storyline of video. However, a direct detection of abrupt changes in the query-frame similarity signal is often unreliable due to high-frequency noise arising from model uncertainty and transient visual variations.To address this, we leverage the wavelet transform, which provides an ideal solution through its multi-resolution analysis in both time and frequency domains. By applying this transform, we decompose the noisy signal into multiple scales and extract a clean semantic change signal from the coarsest scale. We identify the local extrema of this signal as semantic boundaries, which segment the video into coherent clips. Building on this, WFS-SB comprises a two-stage strategy: first, adaptively allocating a frame budget to each clip based on a composite importance score; and second, within each clip, employing the Maximal Marginal Relevance approach to select a diverse yet relevant set of frames. Extensive experiments show that WFS-SB significantly boosts LVLM performance, e.g., improving accuracy by $\textbf{5.5\\% on VideoMME, 9.5\\% on MLVU, and 6.2\\% on LongVideoBench}$, consistently outperforming state-of-the-art methods.

Video question answering (VideoQA) is a challenging task that requires integrating spatial, temporal, and semantic information to capture the complex dynamics of video sequences. Although recent advances have introduced various approaches for video understanding, most existing methods still rely on locating relevant frames to answer questions rather than reasoning through the evolving storyline as humans do. Humans naturally interpret videos through coherent storylines, an ability that is crucial for making robust and contextually grounded predictions. To address this gap, we propose SVAgent, a storyline-guided cross-modal multi-agent framework for VideoQA. The storyline agent progressively constructs a narrative representation based on frames suggested by a refinement suggestion agent that analyzes historical failures. In addition, cross-modal decision agents independently predict answers from visual and textual modalities under the guidance of the evolving storyline. Their outputs are then evaluated by a meta-agent to align cross-modal predictions and enhance reasoning robustness and answer consistency. Experimental results demonstrate that SVAgent achieves superior performance and interpretability by emulating human-like storyline reasoning in video understanding.

Adapting image-pretrained backbones to video typically relies on time-domain adapters tuned to a single temporal scale. Our experiments show that these modules pick up static image cues and very fast flicker changes, while overlooking medium-speed motion. Capturing dynamics across multiple time-scales is, however, crucial for fine-grained temporal analysis (i.e., opening vs. closing bottle).To address this, we introduce Frame2Freq -- a family of frequency-aware adapters that perform spectral encoding during image-to-video adaptation of pretrained Vision Foundation Models (VFMs), improving fine-grained action recognition. Frame2Freq uses Fast Fourier Transform (FFT) along time and learns frequency-band specific embeddings that adaptively highlight the most discriminative frequency ranges. Across five fine-grained activity recognition datasets, Frame2Freq outperforms prior PEFT methods and even surpasses fully fine-tuned models on four of them. These results provide encouraging evidence, that frequency analysis methods are a powerful tool for modeling temporal dynamics in image-to-video transfer.

The internal organization of vision–language models (VLMs) remains poorly understood, particularly how they distribute and fuse information across layers. We take a topology-first perspective and analyze VLMs through the interaction graphs induced by neuron–neuron correlations, treating each layer as a structured computational network rather than a sequence of token transformations. Operating solely on these graphs, we show that global connectivity patterns are strongly predictive of model behavior across grounded reasoning, counting, and hallucination tasks. Modality-separated graphs reveal that cross-modal fusion strengthens sharply in mid-to-late layers, while contrastive graph alignment exposes how multimodal training reorganizes topology relative to text-only backbones. Targeted interventions on high-degree neurons further demonstrate their causal influence, indicating that VLMs route multimodal reasoning through sparse but structurally critical hubs. These results highlight interaction topology as a powerful, model-agnostic lens for interpreting and comparing multimodal transformers.

Vision-language models (VLMs) have achieved remarkable success across diverse tasks. However, concerns about their trustworthiness persist, particularly regarding tendencies to lean more on textual cues than visual evidence and the risk of producing ungrounded or fabricated responses. To address these issues, we propose $\textbf{Saliency-R1}$, a framework for improving the interpretability and faithfulness of VLMs reasoning. Specifically, we introduce a novel saliency map technique that efficiently highlights critical image regions contributing to generated tokens without additional computational overhead. This can further be extended to trace how visual information flows through the reasoning process to the final answers, revealing the alignment between the thinking process and the visual context. We use the overlap between the saliency maps and human-annotated bounding boxes as the reward function, and apply Group Relative Policy Optimization (GRPO) to align the salient parts and critical regions, encouraging models to focus on relevant areas when conduct reasoning. Experiments show Saliency-R1 improves reasoning faithfulness, interpretability, and overall task performance. The dataset, code and pretrained models of this paper will be released.

It has been demonstrated in various contexts that monotonicity leads to better explainability in neural networks. However, not every function can be well approximated by a monotone neural network.We demonstrate that monotonicity can still be used in two ways to boost explainability. First, we use an adaptation of the decomposition of a trained ReLU network into two monotone and convex parts, thereby overcoming numerical obstacles from an inherent blowup of the weights in this procedure. Our proposed saliency methods -- SplitCAM and SplitLRP --improve onstate of the art results on both VGG16 and Resnet18 networks on ImageNet-S across all Quantus saliency metric categories.Second, we exhibit that training a model as the difference between two monotone neural networks results in a system with strong self-explainability properties.

Visual counterfactual explanations aim to reveal the minimal semantic modifications that can alter a model’s prediction, providing causal and interpretable insights into deep neural networks. However, existing diffusion-based counterfactual generation methods are often computationally expensive, slow to sample, and imprecise in localizing the modified regions. To address these limitations, we propose MaskDiME, a simple, fast, yet effective diffusion framework that unifies semantic consistency and spatial precision through localized sampling. Our approach adaptively focuses on decision-relevant regions to achieve localized and semantically consistent counterfactual generation while preserving high image fidelity. Our training-free framework, MaskDiME, performs inference over \textbf{30× faster} than the baseline and achieves comparable or state-of-the-art performance across five benchmark datasets spanning diverse visual domains, establishing a practical and generalizable solution for efficient counterfactual explanation. Our code will be publicly available upon acceptance.

How do video understanding models acquire their answers? Although current Vision Language Models (VLMs) reason over complex scenes with diverse objects, action performances, and scene dynamics, understanding and controlling their internal processes remains an open challenge. Motivated by recent advancements in text-to-video (T2V) generative models, this paper introduces a logits-to-video (L2V) task alongside a model-independent approach, TRANSPORTER, to generate videos that capture the underlying rules behind VLMs' predictions. Given the high-visual-fidelity produced by T2V models, TRANSPORTER learns an optimal transport coupling to VLM's high-semantic embedding spaces. In turn, logit scores define embedding directions for conditional video generation. TRANSPORTER generates videos that reflect caption changes over diverse object attributes, action adverbs, and scene context. Quantitative and qualitative evaluations across VLMs demonstrate that L2V can provide a fidelity-rich, novel direction for model interpretability that has not been previously explored.

Humans do not just see attribute similarity---we also see relational similarity. An apple is like a peach because both are reddish fruit, but the Earth is also like a peach: its crust, mantle, and core correspond to the peach’s skin, flesh, and pit. This ability to perceive and recognize relational similarity, is arguable by cognitive scientist to be what distinguishes humans from other species. %intelligentYet, all widely used visual similarity metrics today (e.g., LPIPS, CLIP, DINO) focus solely on perceptual attribute similarity and fail to capture the rich, often surprising relational similarities that humans perceive.How can we go beyond the visible content of an image to capture its relational properties? How can we bring images with the same relational logic closer together in representation space?To answer these questions, we first formulate relational image similarity as a measurable problem: two images are relationally similar when their internal relations or functions among visual elements correspond, even if their visual attributes differ.We then curate 114k image–caption dataset in which the captions are anonymized---describing the underlying relational logic of the scene rather than its surface content. Using this dataset, we finetune a Vision Language Model to measure the relational similarity between images. This model serves as the first step toward connecting images by their underlying relational structure rather than their visible appearance.Our study shows that while relational similarity has a lot of real-world applications, existing image similarity models fail to capture it---revealing a critical gap in visual computing.

Existing convolutional learning methods for 3D point cloud data are divided into two paradigms: point-based methods that preserve geometric precision but often face performance challenges, and voxel-based methods that achieve high efficiency through quantization at the cost of geometric fidelity. This loss of precision is a critical bottleneck for tasks such as point cloud registration. We propose PointCNN++, a novel architectural design that fundamentally mitigates this precision-performance trade-off. It generalizes sparse convolution from voxels to points, treating voxel-based convolution as a specialized, degraded case of our more general point-based convolution. First, we introduce a point-centric convolution where the receptive field is centered on the original, high-precision point coordinates. Second, to make this high-fidelity operation performant, we design a computational strategy that operates natively on points. We formulate the convolution on native points as a Matrix-Vector Multiplication and Reduction (MVMR) problem, for which we develop a dedicated, highly-optimized GPU kernel. Experiments demonstrate that PointCNN++ uses an order of magnitude less memory and is several times faster than representative point-based methods. Furthermore, when used as a simple replacement for the voxel-based backbones it generalizes, it significantly improves point cloud registration accuracies while proving both more memory-efficient and faster. PointCNN++ shows that preserving geometric detail and achieving high performance are not mutually exclusive, paving the way for a new class of 3D learning with high fidelity and efficiency.

In many real world applications of non-rigid shape matching, the shapes are subject to topological noise (i.e. varying genus). In this paper, we propose a novel formulation based on Markov Random Fields (MRF) that can handle these cases with topological noise. The solutions to our optimisation problem can be approximated efficiently using the alpha expansion algorithm, which gives rise to theoretical approximation guarantees. In particular, we cast non-rigid 3D shape matching as a multi-labelling problem in which each triangle of the source shape is assigned a label that represents the matching to a specific surface element on the target shape. We propose a novel pairwise term that imposes that our matching prefers solutions in which neighbouring triangles on the source shape remain close on the target shape. Further, by exploiting the specific structure of our label space, we show that the alpha expansion algorithm can be customised to gain significant speed-ups, while maintaining its approximation guarantees. We test our formalism on various shape matching datasets including settings in which shapes have topological artefacts.

Modern neural architectures for 3D point cloud processing contain both convolutional layers and attention blocks, but the best way to assemble them remains unclear. We analyse the role of different computational blocks in 3D point cloud networks and find an intuitive behaviour: convolution is adequate to extract low-level geometry at high-resolution in early layers, where attention is expensive without bringing any benefits; attention captures high-level semantics and context in low-resolution, deep layers more efficiently. Guided by this design principle, we propose a new, improved 3D point cloud backbone that employs convolutions in early stages and switches to attention for deeper layers. To avoid the loss of spatial layout information when discarding redundant convolution layers, we introduce a novel, training-free 3D positional encoding, PointROPE. The resulting LitePT model has 3.6× fewer parameters, runs 2× faster, and uses 2× less memory than the state-of-the-art Point Transformer V3, but nonetheless matches or even outperforms it on a range of tasks and datasets.

While existing point-cloud-registration methods can well handle high-overlap scenarios of two point clouds, they often struggle with low-overlap scenarios, due to inevitable geometric/semantic ambiguities in the non-overlapping regions. In this paper, we introduce SuP, a novel framework that reformulates low-overlap registration as a high-overlap sub-cloud pairs (anchor pairs) mining problem. Central to SuP is our Dual-phase Sub-cloud Anchor Mining (DSAM) module, which first subdivides the source and target point clouds into multiple sub-clouds, followed by introducing a dual-phase weighting pipeline: 1) an efficient overlap-guided prior-weighting scheme (OPS) that leverages feature salience to identify candidate anchor pairs, and 2) a multi-scale post-weighting network (MPN) that exploits neighborhood feature consensus to further identify anchor pairs. Subsequently, final correspondences are generated through a merge-to-match module using the anchor pairs. To train DSAM, we design an alignment-aware weighting loss that uses on-the-fly alignment errors as supervision. Comprehensive experiments on the color-enhanced 3DMatch and 3DLoMatch demonstrate that SuP significantly outperforms state-of-the-art methods, achieving higher registration recall and more accurate alignment, especially under challenging low-overlap conditions.

Point clouds are a fundamental 3D representation in computer vision, enabling a wide range of perception tasks. However, real-world point clouds often suffer from degradations such as incompleteness, noise, outliers, and irregular density, caused by sensor limitations or occlusions. Recovering clean and detailed shapes from such degraded data is crucial for downstream applications. While existing learning-based methods achieve progress on individual tasks like completion or denoising, they typically rely on global bottleneck features, which lose fine-grained geometry and remain sensitive to varying input quality. We propose a unified 3D restoration network that directly takes point clouds as input and adaptively reconstructs high-quality geometry under diverse degradation scenarios. At the core of our approach is a Pseudo-Query module, implemented within a Transformer backbone, which reformulates geometric translation into two cooperative stages to enhance structural clarity, robustness, and local detail preservation. Extensive experiments on curated benchmarks demonstrate that our approach surpasses state-of-the-art performance in general 3D restoration. It effectively handles complex combinations of completion, deformation, and denoising degradations. With this work, we provide a novel unified, point-only backbone for robust 3D restoration, paving the way for more versatile 3D perception.

LiDAR semantic segmentation must remain robust under various sensor and environmental corruptions to be reliable in safety-critical applications.Existing test-time adaptation methods, including approaches based on pseudo-labels and normalization statistics, have shown promising results but can still struggle under severe distribution shifts.To complement these approaches, we propose a geometry-aware test-time training framework that leverages an auxiliary self-supervised objective.Our method is based on geometric inlier discrimination (GeoID), which injects synthetic off-manifold points into the input and trains the model to distinguish geometry-consistent inliers from synthetically displaced outliers, enabling adaptation on unlabeled test data.To further stabilize this process under real corruptions, we introduce bidirectional unreliable point filtering (BiUPF), which uses inlier scores from the source-trained model to filter out unreliable regions on both original and synthetic points, focusing updates on high-confidence samples.Experiments on two large-scale corruption benchmarks, SemanticKITTI-C and nuScenes-C, show that our method consistently outperforms strong test-time adaptation baselines and improves robustness across diverse LiDAR corruptions.Code will be released.

Outlier rejection for correspondence-based point cloud registration confronts two fundamental challenges in real-world scenarios. First, low-overlap regions yield sparse and fragmented inlier distributions that are difficult to discover using conventional one-step global search strategies. Second, large-scale scenes present dense correspondence inputs that impose stringent requirements on the accuracy-efficiency trade-off of search algorithms. To this end, we propose a hierarchical multi-hop graph search framework that progressively refines correspondences to address these challenges. Our method constructs a compatibility graph with transformation-invariant embeddings to predict correspondence confidence, establishing the foundation for cluster-balanced seed sampling that ensures comprehensive coverage across fragmented regions. These strategically selected seeds subsequently drive hierarchical multi-hop expansion, progressively discovering inliers through multi-resolution graph layers while circumventing the high complexity of exhaustive global search. Finally, distribution-aware ranking jointly evaluates geometric consistency and spatial coverage to select well-distributed transformations from multiple hypotheses. Experiments on 3DMatch, 3DLoMatch, and KITTI demonstrate that our method significantly outperforms state-of-the-art methods in both low-overlap and large-scale scenarios.

World models, which simulate environmental dynamics and generate sensor observations, are gaining increasing attention in autonomous driving. However, progress in LiDAR-based world models has lagged behind those built on camera videos or occupancy data, primarily due to two core challenges: the inherent disorder of point clouds and the difficulty of distinguishing dynamic objects from static structures. To address these issues, we propose GEM: a Generative LiDAR world model that leverages dEformable Mamba architecture, significantly improving fidelity and imaginative capability. Specifically, leveraging the structural similarity between sequential laser scanning and Mamba's processing mechanism, we first tokenize LiDAR sweeps into compact representations via a custom LiDAR scene tokenizer. After unsupervised disentanglement of tokenized features via a dynamic-static separator, a tri-path deformable Mamba is introduced to perform selective scanning and adaptive gating fusion over the disentangled features, leading to enhanced spatial-temporal understanding of the world evolution. Optionally, a planner and a BEV layout controller can be integrated to explore the model's capability for autonomous rollout and it's potential to generate "what-if" scenarios. Extensive experiments show that GEM achieves state-of-the-art performances across diverse benchmarks and evaluation settings, demonstrating its superiority and effectiveness.

Current collaborative perception systems have significantly improved 3D object detection performance.However, widely used LiDAR and camera systems often suffer performance degradation under adverse weather conditions.The weather-robust 4D radar provides a promising solution to address this challenge.Nevertheless, the effective fusion of sparse 4D radar measurements with degraded LiDAR data remains a significant challenge due to cross-modal corruption and information loss. In this work, we propose a novel hybrid robust collaborative perception framework (HRCP), designed to improve the collaborative perception performance under adverse weather conditions through LiDAR-4D radar fusion.Specifically, we introduce a hybrid collaboration strategy that considers their distinct physical properties and differently processes them during information transmission.Additionally, we propose a bidirectional cross-modal gating (BCMG) module that enables LiDAR and 4D radar to mutually validate feature reliability, ensuring consistent cross-modal representation, and an adaptive feature enhancement (AFE) module that enables comprehensive refinement of degraded and suppressed regions to mitigate information loss.Extensive experimental results demonstrate that our method outperforms previous state-of-the-art approaches under adverse weather conditions.

Imaging system design is a complex, time-consuming, and largely manual process; LiDAR design, ubiquitous in mobile devices, autonomous vehicles, and aerial imaging platforms, adds further complexity through unique spatial and temporal sampling requirements. In this work, we propose a framework for automated, task-driven LiDAR system design under arbitrary constraints. To achieve this, we represent LiDAR configurations in a continuous six-dimensional design space and learn task-specific implicit densities in this space via flow-based generative modeling. We then synthesize new LiDAR systems by modeling sensors as parametric distributions in 6D space and fitting these distributions to our learned implicit density using expectation-maximization, enabling efficient, constraint-aware LiDAR system design. We validate our method on diverse tasks in 3D vision, enabling automated LiDAR system design across real-world-inspired applications in face scanning, robotic tracking, and object detection.

3D shape anomaly detection is a crucial task for industrial inspection and geometric analysis. Existing deep learning approaches typically learn representations of normal shapes and identify anomalies via out-of-distribution feature separation or decoder-based reconstruction. They often fail to generalize across diverse anomaly types and scales, such as global geometric errors (e.g., planar shifts, surface misalignments), and are sensitive to noisy or incomplete local points during training. To address these limitations, we propose a hierarchical point–patch anomaly scoring network that jointly models regional part features and local point features for robust anomaly reasoning. An adaptive patchification module integrates self-supervised decomposition to capture complex structural deviations. Beyond evaluations on public benchmarks (Anomaly-ShapeNet and Real3D-AD), we release an industrial test set with real CAD models exhibiting planar, angular, and structural defects. Experiments on public and industrial datasets show superior AUC-ROC and AUC-PR performance, including over 50\% point-level improvement on the new industrial anomaly type and average object-level gains of 7\% on Real3D-AD and 4\% on Anomaly-ShapeNet, demonstrating strong robustness and generalization.

Text-to-video diffusion models have enabled open-ended video synthesis, but often struggle with generating the correct number of objects specified in a prompt. We introduce NUMINA, a training-free identify-then-guide framework for improved numerical alignment. NUMINA identifies prompt–layout inconsistencies by selecting discriminative self- and cross-attention heads to derive a countable latent layout. It then refines this layout conservatively and modulates cross-attention to guide regeneration. On the introduced CountBench, NUMINA improves counting accuracy by up to 7.4\% on Wan2.1-1.3B, and by 4.9\% and 5.5\% on 5B and 14B models, respectively. Furthermore, CLIP alignment is improved while maintaining temporal consistency. These results demonstrate that structural guidance complements seed search and prompt enhancement, offering a practical path toward count-accurate text-to-video diffusion. The code will be made available.

While model merging has demonstrated remarkable success across diverse domains for large language models (LLMs), its application to vision-language models (VLMs) remains largely underexplored. Recent methods attempt to enhance VLM reasoning capabilities by integrating specialized LLM parameters through layer-wise merging. However, existing paradigms suffer from two critical limitations: (1) strict positional correspondence, which enforces rigid one-to-one layer alignment, and (2) uniform merging weights applied indiscriminately across all layers. These constraints fail to account for substantial functional disparities between corresponding layers in VLMs and LLMs, potentially misaligning incompatible layers and leading to detrimental parameter combinations.To address these, we propose Chain-of-Merging (CoM) framework that adaptively adjusts merging plans for different images and questions, including two key stages: (1) Adaptive Layer Matching, which identifies optimal layer pairings based on structural and semantic matching scores while filtering incompatible pairings, and (2) Dynamic Weight Merging, which determines layer-specific merging weights based on matching scores and employs spherical linear interpolation to minimize memory overhead.Extensive experiments demonstrate that CoM achieves substantial performance improvements, with Qwen2.5-VL-7B + Qwen2.5-Math-7B attaining a 4.4\% average improvement on mathematical reasoning benchmarks while enhancing general visual understanding, significantly outperforming existing training-free methods.

Gaze estimation is instrumental in modern virtual reality (VR) systems. Despite significant progress in remote-camera gaze estimation, VR gaze research remains constrained by data scarcity—particularly the lack of large-scale, accurately labeled datasets captured with the off-axis camera configurations typical of modern headsets. Gaze annotation is difficult since fixation on intended targets cannot be guaranteed. To address these challenges, we introduce VRGaze—the first large-scale off-axis gaze estimation dataset for VR—comprising 2.1 million near-eye infrared images collected from 68 participants. We further propose GazeShift, an attention-guided unsupervised framework for learning gaze representations without labeled data. Unlike prior redirection-based methods that rely on multi-view or 3D geometry, GazeShift is tailored to near-eye infrared imagery, achieving effective gaze–appearance disentanglement in a compact, real-time model. A lightweight few-shot calibration can optionally adapt embeddings to individual users, achieving 1.84° mean error on VRGaze under per-person calibration and 7.15° on MPIIGaze under person-agnostic calibration, with a tenfold reduction in parameters and 5 ms runtime on a VR headset GPU. Quantitative robustness analyses confirm invariance to illumination variations, demonstrating a label-efficient and deployable solution for VR gaze estimation.VRGaze and GazeShift are released under \url{https://github.com/gazeshift3/gazeshift}.

Test-time prompt tuning (TPT) has emerged as a promising technique for enhancing the adaptability of vision-language models by optimizing textual prompts using unlabeled test data. However, prior studies have revealed that TPT often produces poorly calibrated models, raising concerns about the reliability of their predictions. Recent works address this issue by incorporating additional regularization terms that constrain model outputs, which improve calibration but often degrade performance. In this work, we reveal that these regularization strategies implicitly encourage optimization toward flatter minima, and that the sharpness of the loss landscape around adapted prompts is a key factor governing calibration quality. Motivated by this observation, we introduce Flatness-aware Prompt Pretraining (FPP), a simple yet effective pretraining framework for TPT that initializes prompts within flatter regions of the loss landscape prior to adaptation. We show that simply replacing the initialization in existing TPT pipelines---without modifying any other components---is sufficient to improve both calibration and performance. Notably, FPP requires no labeled data and avoids any additional computational costs during test-time tuning, making it highly practical for real-world deployment. The code will be released.

Recent research has indicated that the embeddings generated by contrastive language-image training like CLIP may not be ideal for image-only tasks. The main theory is that the inter-modal (language-image) alignment loss ignores intra-modal (image-image) alignment, leading to poorly calibrated similarities between images.In this study, we question this intra-modal misalignment hypothesis.We reexamine the theoretical arguments and techniques that seek to demonstrate the misalignment.Our findings reveal that neither the distribution of cosine similarities nor few-shot or retrieval metrics serve as reliable indicators of misalignment.In fact, these metrics yield similar results for language-image trained models (CLIP, SigLIP) and image-image trained models (DINO, SigLIP2), which indicates there is no intra-modal misalignment stemming from contrastive language-image training.We argue the observed phenomena can be explained without assuming a fundamental flaw in the image embedding space.Experiments on the commonly studied intra-modal tasks retrieval and few-shot classification confirm that addressing supposed misalignment is unnecessary for achieving strong performance.

Following the success of Group Relative Policy Optimization (GRPO) in foundation LLMs, an increasing number of works have sought to adapt GRPO to Visual Large Language Models (VLLMs) for visual perception tasks (e.g., detection and segmentation). However, much of this line of research rests on a long-standing yet unexamined assumption: training paradigms developed for language reasoning can be transferred seamlessly to visual perception. Our experiments show that this assumption is not valid, revealing intrinsic differences between reasoning-oriented and perception-oriented settings. Using reasoning segmentation as a representative case, we surface two overlooked factors: (i) the need for a broader output space, and (ii) the importance of fine-grained, stable rewards. Building on these observations, we propose Dr.Seg, a simple, plug-and-play GRPO-based framework consisting of a Look-to-Confirm mechanism and a Distribution-Ranked Reward module, requiring no architectural modifications and integrating seamlessly with existing GRPO-based VLLMs. Extensive experiments demonstrate that Dr.Seg improves performance in complex visual scenarios while maintaining strong generalization. Code, data, and models will be publicly released upon acceptance.

Mixture-of-Experts (MoE) has become a prevalent backbone for large vision-language models (VLMs), yet how modality-specific signals should guide expert routing remains under-explored. Existing routing strategies are either hand-crafted or modality-agnostic, relying on idealized priors that ignore the layer-dependent modality fusion patterns in MoE-VLMs and provide little guidance for expert specialization. We propose Soft Modality-guided Expert Specialization (SMoES), which consists of dynamic soft modality scores that capture layer-dependent fusion patterns, an expert binning mechanism aligned with expert-parallel deployment, and an inter-bin mutual information regularization that encourages coherent modality specialization. Our method leverages attention-based or Gaussian-statistics modality scores to optimize mutual information regularization. Experiments across four MoE-based VLMs and 16 benchmarks demonstrate improvement on both effectiveness and efficiency: 0.9\% and 4.2\% average gain on multimodal and language-only tasks, 56.1\% reduction in EP communication overhead, and 12.3\% throughput improvement under realistic deployment. These results validate that aligning routing with modality-aware expert specialization unlocks MoE-VLM capacity and efficiency.

Multimodal large language models (MLLMs) achieve remarkable progress in cross-modal perception and reasoning, yet a fundamental question remains unresolved: should the vision encoder be fine-tuned or frozen? Despite the success of models such as LLaVA and Qwen-VL, inconsistent design choices and heterogeneous training setups hinder a unified understanding of visual fine-tuning (VFT) in MLLMs. Through a configuration-aligned benchmark, we find that existing VFT methods fail to consistently outperform the frozen baseline across multimodal tasks. Our analysis suggests that this instability arises from visual preference conflicts, where the context-agnostic nature of vision encoders induces divergent parameter updates under diverse multimodal context. To address this issue, we propose the Context-aware Visual Fine-tuning (CoVFT) framework, which explicitly incorporates multimodal context into visual adaptation. By integrating a Context Vector Extraction (CVE) and a Contextual Mixture-of-Experts (CoMoE) module, CoVFT decomposes conflicting optimization signals and enables stable, context-sensitive visual updates. Extensive experiments on 12 multimodal benchmarks demonstrate that CoVFT achieves state-of-the-art performance with superior stability. Notably, fine-tuning a 7B MLLM with CoVFT surpasses the average performance of its 13B counterpart, revealing substantial untapped potential in visual encoder optimization within MLLMs. Code will be released.

Vision-language models (VLMs) have achieved impressive performance across a wide range of multimodal reasoning tasks, but they often struggle to disentangle fine-grained visual attributes and reason about underlying causal relationships. In-context learning (ICL) offers a promising avenue for VLMs to adapt to new tasks, but its effectiveness critically depends on the selection of demonstration examples. Existing retrieval-augmented approaches typically rely on passive similarity-based retrieval, which tends to select correlated but non-causal examples, amplifying spurious associations and limiting model robustness. We introduce CIRCLES (Composed Image Retrieval for Causal Learning Example Selection), a novel framework that actively constructs demonstration sets by retrieving counterfactual examples through targeted, attribute-guided composed image retrieval. By incorporating counterfactual-style examples, CIRCLES enables VLMs to reason about the causal relations between attributes and outcomes, moving beyond superficial correlations and fostering more robust and causally grounded reasoning. Comprehensive experiments on four diverse datasets demonstrate that CIRCLES consistently outperforms existing methods across multiple architectures, especially on small-scale models, with pronounced gains under information scarcity. Furthermore, CIRCLES retrieves more diverse and causally informative examples, providing qualitative insights into how models leverage in-context demonstrations for improved reasoning.

Previous representation and generation approaches for the B-rep relied on graph-based representations that disentangle geometric and topological features through decoupled computational pipelines, thereby precluding the application of sequence-based generative frameworks, such as transformer architectures that have demonstrated remarkable performance. In this paper, we propose BrepARG, the first attempt to encode B-rep's geometry and topology into a holistic token sequence representation, enabling sequence-based B-rep generation with an autoregressive architecture. Specifically, BrepARG encodes B-rep into 3 types of tokens: geometry and position tokens representing geometric features, and face index tokens representing topology. Then the holistic token sequence is constructed hierarchically, starting with constructing the geometry blocks (i.e., faces and edges) using the above tokens, followed by geometry block sequencing. Finally, we assemble the holistic sequence representation for the entire B-rep. We also construct a transformer-based autoregressive model that learns the distribution over holistic token sequences via next-token prediction, using a multi-layer decoder-only architecture with causal masking. Experiments demonstrate that BrepARG achieves state-of-the-art (SOTA) performance. BrepARG validates the feasibility of representing B-rep as holistic token sequences, opening new directions for B-rep generation.

Vector glyphs are the atomic units of digital typography, yet most learning-based pipelines still depend on carefully curated exemplar sheets and raster-to-vector postprocessing, which limits accessibility and editability. We introduce VecGlypher, a single multimodal language model that generates high-fidelity vector glyphs directly from text descriptions or image exemplars. Given a style prompt, optional reference glyph images, and a target character, VecGlypher autoregressively emits SVG path tokens, avoiding raster intermediates and producing editable, watertight outlines in one pass. A typography-aware data and training recipe makes this possible: (i) a large-scale continuation stage on 39K noisy Envato fonts to master SVG syntax and long-horizon geometry, followed by (ii) post-training on 2.5K expert-annotated Google Fonts with descriptive tags and exemplars to align language and imagery with geometry; preprocessing normalizes coordinate frames, canonicalizes paths, de-duplicates families, and quantizes coordinates for stable long-sequence decoding. On cross-family OOD evaluation, VecGlypher substantially outperforms both general-purpose LLMs and specialized vector-font baselines for text-only generation, while image-referenced generation reaches a state-of-the-art performance, with marked gains over DeepVecFont-v2 and DualVector. Ablations show that model scale and the two-stage recipe are critical and that absolute-coordinate serialization yields the best geometry. VecGlypher lowers the barrier to font creation by letting users design with words or exemplars, and provides a scalable foundation for future multimodal design tools.

The proliferation of neural radiance field (NeRF) research requires significant efforts to reimplement papers before building upon them. We introduce NERFIFY, a multi-agent framework that reliably converts NeRF research papers into trainable Nerfstudio plugins, in contrast to generic paper-to-code methods and frontier models like GPT-5 that usually fail to produce runnable code. NERFIFY achieves domain-specific executability through six key innovations: (1) Context-free grammar (CFG): LLM synthesis is constrained by Nerfstudio formalized as a CFG, ensuring generated code satisfies architectural invariants. (2) Graph-of-Thought code synthesis: Specialized multi-file-agents generate repositories in topological dependency order, validating contracts and errors at each node. (3) Compositional citation recovery: Agents automatically retrieve and integrate components (samplers, encoders, proposal networks) from citation graphs of references. (4) Visual feedback: Artifacts are diagnosed through PSNR-minima ROI analysis, cross-view geometric validation, and VLM-guided patching to iteratively improve quality. (5) Knowledge enhancement: Beyond reproduction, methods can be improved with novel regularizers or architectural optimizations. (6) Benchmarking: An evaluation framework is designed for NeRF paper-to-code synthesis across 30 diverse papers. On papers without public implementations, NERFIFY achieves visual quality matching expert human code (±0.5 dB PSNR, ±0.2 SSIM) while reducing implementation time from weeks to minutes. NERFIFY demonstrates that a domain-aware design enables code translation for complex vision papers, potentiating accelerated and democratized reproducible research. Code, data and implementations will be publicly released.

Diagrams are widely used in daily life. However, offline diagrams typically exist in the form of images, lacking structured data representation, which significantly limits their reusability and editability. Current research mainly focuses on supporting basic query tasks for online diagrams and does not meet the semantic understanding and interaction requirements for complex offline diagrams. Although large language models (LLMs) possess powerful reasoning and knowledge integration capabilities, their performance in processing offline diagrams is unsatisfactory due to the inability to accurately understand the structure and content of offline diagrams. To address these issues,we propose DiagramDiff, a framework consisting of a high-precision diagram reconstruction model and an instance-level diagram element recognition model. The framework converts offline diagrams into standardized data structures, enabling LLMs to transition from being unable to understand offline diagrams to becoming intelligent assistants capable of performing tasks such as semantic reasoning, logical validation, and efficient diagram editing. We have constructed a dataset containing diagrams and their corresponding question and answering(Q&A) and editing tasks. Experiments demonstrate that DiagramDiff achieves state-of-the-art performance in diagram reconstruction and recognition tasks, significantly enhancing LLMs' understanding and interaction capabilities with offline diagrams.

While existing generation and unified models excel at general image generation, they struggle with tasks requiring deep reasoning, planning, and precise data-to-visual mapping abilities beyond general scenarios. To push beyond the existing limitations, we introduce a new and challenging task: creative table visualization, requiring the model to generate an infographic that faithfully and aesthetically visualizes the data from a given table. To address this challenge, we propose ShowTable, a pipeline that synergizes MLLMs with diffusion models via a progressive self-correcting process. The MLLM acts as the central orchestrator for reasoning the visual plan and judging visual errors to provide refined instructions, the diffusion execute the commands from MLLM, achieving high-fidelity results. To support this task and our pipeline, we introduce three automated data construction pipelines for training different modules. Furthermore, we introduce TableVisBench, a new benchmark with 800 challenging instances across 5 evaluation dimensions, to assess performance on this task. Experiments demonstrate that our pipeline, instantiated with different models, significantly outperforms baselines, highlighting its effective multi-modal reasoning, generation, and error correction capabilities.

Jiangnan gardens, a prominent style of Chinese classical gardens, hold great potential as digital assets for film and game production and digital tourism. However, manual modeling of Jiangnan gardens heavily relies on expert experience for layout design and asset creation, making the process time-consuming. To address this gap, we propose GardenDesigner, a novel framework that encodes aesthetic principles for Jiangnan garden construction and integrates a chain of agents based on procedural modeling. The water-centric terrain and explorative pathway rules are applied by terrain distribution and road generation agents. Selection and spatial layout of garden assets follow the aesthetic and cultural constraints. Consequently, we propose asset selection and layout optimization agents to select and arrange objects for each area in the garden. Additionally, we introduce GardenVerse for Jiangnan garden construction, including expert-annotated garden knowledge to enhance the asset arrangement process. To enable interaction and editing, we develop an interactive interface and tools in Unity, in which non-expert users can construct Jiangnan gardens via text input within one minute. Experiments and human evaluations demonstrate that GardenDesigner can generate diverse and aesthetically pleasing Jiangnan gardens.

We introduce ShadowDraw, a framework that transforms ordinary 3D objects into shadow–drawing compositional art. Given a 3D object, our system predicts scene parameters---including object pose and lighting---together with a partial line drawing, such that the cast shadow completes the drawing into a recognizable image. To this end, we optimize scene configurations to reveal meaningful shadows, employ shadow contour to guide line drawing generation, and adopt automatic evaluation to enforce shadow-drawing coherence and visual quality. Experiments show that ShadowDraw produces compelling results across diverse inputs, from real-world scans and curated datasets to generative assets, and naturally extends to multi-object scenes, animations, and physical deployments. Our work provides a practical pipeline for creating shadow–drawing art and broadens the design space of computational visual art, bridging the gap between algorithmic design and artistic storytelling.For more results and an end-to-end real-world demonstration of our pipeline, please refer to the project page in the supplementary material.

Engineering diagrams are the core carriers of technical information in industrial contexts, where the pressing demand for their digitization from industrial sectors has driven great advancements in related research domains. However, existing research still suffers from three limitations. Firstly, the detection of symbols, lines, and texts typically involves multiple independent models, resulting in cumbersome workflows. In addition, high-resolution diagrams often impose an excessive computational cost on existing models. Moreover, parsing frameworks solely based on object detection can merely localize component positions, yet fail to capture the topological connection semantics and structured knowledge among components, thus offering limited convenience for industrial applications. To address these issues, we propose an end-to-end information extraction framework based on the Dynamically Tokenized Relation Transformer (DTRT), which can dynamically reduce received image tokens, filter redundant information, and efficiently extract structural knowledge to construct hyper-relational knowledge graphs. We practiced our model on piping and instrumentation diagrams (P&IDs) and electrical diagrams (EDs): the former are widely used in chemical engineering enterprises, while the latter are employed to describe circuit systems. DTRT achieves an R@1000 accuracy of 94.84% on PIDs and R@200 accuracy of 92.52% on EDs with a significantly reduced computational cost.

Text-to-image (T2I) models are increasingly popular, producing a large share of AI-generated images online. To compare their quality, voting-based leaderboards have become the standard, relying on anonymized model outputs for fairness. In this work, we show that such anonymity can be easily broken. We find that generations from each T2I model form distinctive clusters in the image embedding space, enabling accurate deanonymization without prompt control or training data. Using 22 models and 280 prompts (150K images), our centroid-based method achieves high accuracy and reveals systematic model-specific signatures. We further introduce a prompt-level distinguishability metric and conduct large-scale analyses showing how certain prompts amplify these signatures. Our findings expose fundamental security risks in T2I leaderboards and motivate stronger anonymization defenses.

Reward models (RMs) are inherently non-neutral value functions designed and trained to encode specific objectives, such as human preferences or text-image alignment. RMs have become crucial components of text-to-image (T2I) generation systems where they are used during pretraining, finetuning of models and test-time optimization and post-generation safety and quality filtering of T2I outputs. While specific problems with the integration of RMs into the T2I pipeline have been studied (e.g. reward hacking or mode collapse during training), their robustness and fairness as scoring functions remains largely unknown. We conduct a large-scale audit of RMs' robustness with respect to demographic biases during T2I model training and generation. We provide quantitative and qualitative evidence that while originally developed as quality measures, RMs encode demographic biases, which cause reward-guided optimization to sexualize female images (especially darker-skinned females), reinforce gender/racial stereotypes, and collapse demographic diversity. These findings highlight the shortcomings of current RMs, challenging their reliability as quality metrics and underscoring the critical need for alternative data collection, training, and optimization procedures to establish more robust scoring.

By 2026, up to 90% of online content may be synthetically generated, raising serious concerns about the spread of AI-generated disinformation. Policymakers and companies alike are responding California’s Bill AB 321 mandates watermarking of AI-generated media, while firms like Meta and Google are deploying watermarking systems to curb the misuse of generative models. Yet, watermarking techniques remain fragile. In this work, we introduce and analyze a novel vulnerability: the visual paraphrase attack, a generative method capable of stripping both visible and invisible watermarks from AI-generated images. The attack operates in two steps: first, a caption is generated for an image. Then, the image and its caption are passed to a diffusion-based text-to-image system, producing a visually similar but watermark free image. Our empirical evaluation demonstrates that visual paraphrasing reliably removes watermarks while preserving the original image’s semantic content, revealing a fundamental weakness in current watermarking systems. To address this, we introduce PECCAVI, the first watermarking method explicitly designed to withstand visual paraphrase attacks. PECCAVI embeds robust, distortion-free watermarks within semantically stable regions of the image, which we term Non-Melting Points (NMPs). The method uses multi-channel frequency domain watermarking and incorporates noisy burnishing to obfuscate watermark locations and resist reverse engineering. PECCAVI is model-agnostic and significantly more durable than existing approaches. We release the first visual paraphrase benchmark dataset and open-source all code and resources1, offering a foundation for future work on robust watermarking in the age of generative AI.

Large vision-language models (VLMs) are highly vulnerable to multimodal jailbreak attacks that exploit visual-textual interactions to bypass safety guardrails. In this paper, we present DTR, a novel inference-time defense that mitigates multimodal jailbreak attacks through optimizing the model’s key-value (KV) caches. Rather than relying on curated safety-specific data or costly image-to-text conversion, we introduce a new formulation of the safety-relevant distributional shift induced by the visual modality. This formulation enables DTR to dynamically adjust visual token weights, minimizing the impact of adversarial visual inputs while preserving the model’s general capabilities and inference efficiency. Extensive evaluation across diverse VLMs and attack benchmarks demonstrates that DTR outperforms existing defenses in both attack robustness and benign-task performance, marking the first successful application of KV cache optimization for safety enhancement in multimodal foundation models. (The code for replicating DTR is included in the supplementary materials.)

Recent advances in text-to-image (T2I) generation can produce highly resembling images of copyrighted characters, often indistinguishable from official depictions, raising serious concerns about intellectual property infringement. Consequently, robust detection of copyright character infringement is urgently needed. Yet, existing methods exhibit limited alignment with human judgments regarding the likelihood of infringement. To bridge this gap, we propose \textsc{CopyLens}, a novel prompt optimization framework that automatically refines textual prompts for vision-language model-based detectors to better match human infringement judgments. Our approach establishes a closed-loop refinement process between a large vision-language model (LVLM) and a large language model (LLM): the LVLM assesses generated images for copyright detection, while the LLM iteratively optimizes detection prompts via meta-prompting, guided by feedback signals derived from human annotation consistency. To facilitate the assessment of prompt-human alignment, we introduce \textsc{CopyChars}, a new large-scale dataset of over 7,000 AI-generated images spanning more than 100 popular copyrighted characters, along with detailed human annotations on potential infringement. Extensive experiments on \textsc{CopyChars} show that the proposed \textsc{CopyLens} can improve detection performance by 5\% to 10\% compared to recent state-of-the-art methods. This work offers a scalable and automated solution for visual copyright protection and highlights the critical role of prompt engineering.

Modern generative models, including diffusion-based architectures, produce highly creative outputs but also pose safety and ethical risks. These concerns have led to growing interest in concept erasure, the process of removing unwanted concepts from model representations. Existing approaches often achieve strong erasure performance but rely on iterative optimization and may inadvertently distort unrelated concepts. In this work, we present a simple yet principled alternative: a linear transformation framework that achieves concept erasure analytically, without any training. Our method adapts a pretrained model through two sequential, closed-form steps: first, computing a proxy projection of the target concept, and second, applying a constrained transformation within the left null space of known concept directions. This design yields a deterministic and geometrically interpretable procedure for safe, efficient, and theory-grounded concept removal. Across a wide range of experiments, including object and style erasure on multiple Stable Diffusion variants and the flow-matching model (FLUX), our approach matches or surpasses the performance of state-of-the-art methods while preserving non-target concepts more faithfully. Requiring only a few seconds to apply, it offers a lightweight and drop-in tool for controlled model editing, advancing the goal of safer and more responsible generative models.

Non-transferable learning (NTL) aims to enforce usage restrictions by limiting a model’s generalization on target-domain data while maintaining its utility on the source domain. Current approaches face three major challenges: (1) low training efficiency due to retraining of the backbone network, (2) low inference efficiency, and (3) a rigid reliance on a shared, non-adaptive backbone network spanning both source and target domains. This shared setup, which aims to maximize source-domain performance and minimize target-domain performance, often introduces optimization conflicts due to overlapping class categories across source and target domains.In this paper, we propose a novel and efficient NTL approach using a dynamic Early-Exit Network, named ENL-DEE, which leverages Bayesian theory and dynamic neural networks to address these limitations. Our custom loss function guides source-domain data to exit at later stages of the network, maximizing model utility, while target-domain data exits earlier with non-semantic features, ensuring limited transferability. ENL-DEE offers three key advantages: (1) it enhances training efficiency by optimizing only the parameters of dynamic exit classifiers, bypassing the need to retrain the backbone; (2) it improves inference efficiency as data exits at various exit classifiers in the network; and (3) it resolves optimization conflicts by using distinct parameter sets for source and target domains, achieving higher performance on the source domain and lower performance on the target domain, thereby strengthening NTL. Extensive experiments across diverse datasets and model architectures validate the scalability, efficiency, and effectiveness of our approach.

Machine unlearning (MU) addresses privacy risks in pretrained models. The main goal of MU is to remove the influence of designated data while preserving the utility of retained knowledge. Achieving this goal requires preserving semantic relations among retained instances, which existing studies often overlook. We observe that without such preservation, models suffer from progressive structural collapse, undermining both the deletion–retention balance. In this work, we propose a novel structure-faithful framework that introduces stakes, i.e., semantic anchors that serve as reference points to maintain the knowledge structure. By leveraging these anchors, our framework captures and stabilizes the semantic organization of knowledge. Specifically, we instantiate the anchors from language-driven attribute descriptions encoded by a semantic encoder (e.g., CLIP). We enforce preservation of the knowledge structure via structure-aware alignment and regularization: the former aligns the organization of retained knowledge before and after unlearning around anchors, while the latter regulates updates to structure-critical parameters. Results from image classification, retrieval, and face recognition show average gains of 32.9%, 22.5%, and 19.3% in performance, balancing the deletion–retention trade-off and enhancing generalization.

Federated active learning (FAL) seeks to reduce annotation cost under privacy constraints, yet its effectiveness degrades in realistic settings with severe global class imbalance and highly heterogeneous clients. We conduct a systematic study of query-model selection in FAL and uncover a central insight: the model that achieves more class-balanced sampling, especially for minority classes, consistently leads to better final performance. Moreover, global-model querying is beneficial only when the global distribution is highly imbalanced and client data are relatively homogeneous; otherwise, the local model is preferable.Based on these findings, we propose FairFAL, an adaptive class-fair FAL framework. FairFAL (1) infers global imbalance and local–global divergence via lightweight prediction discrepancy, enabling adaptive selection between global and local query models; (2) performs prototype-guided pseudo-labeling using global features to promote class-aware querying; and (3) applies a two-stage uncertainty–diversity balanced sampling strategy with $k$-center refinement. Experiments on five benchmarks show that FairFAL consistently outperforms state-of-the-art approaches under challenging long-tailed and non-IID settings.

Federated learning (FL) suffers from performance degradation due to the inevitable presence of noisy annotations in distributed scenarios. Existing approaches have advanced in distinguishing noisy samples from the dataset for label correction by leveraging loss values. However, noisy samples recognition relying on scalar loss lacks reliability for FL under heterogeneous scenarios. In this paper, we rethink this paradigm from a representation perspective and propose FedRG(Federated under Representation Gemometry), which follows ''the principle of ``representation geometry priority'' to recognize noisy labels. Firstly, FedRG creates label-agnostic spherical representations by using self-supervision. It then iteratively fits a spherical von Mises-Fisher (vMF) mixture model to this geometry using previously identified clean samples to capture semantic clusters. This geometric evidence is integrated with a semantic-label soft mapping mechanism to derive a distribution divergence between the label-free and annotated label-conditioned feature space, which robustly identifies noisy samples and updates the vMF mixture model with the newly separated clean dataset. Lastly, we employ an additional personalized noise absorption matrix on noisy labels to achieve robust optimization. Extensive experimental results demonstrate that FedRG significantly outperforms state-of-the-art methods for FL with data heterogeneity under diverse noisy client scenarios.

Growing privacy and security demands in the real world have spurred interest in adversarially robust Federated Learning (FL). While Adversarial Training (AT) is a well-established defense in centralized learning, its extension to the federated setting, known as Federated Adversarial Training (FAT), faces significant challenges due to data heterogeneity across clients. Existing FAT methods have made significant contributions, but they typically assume a balanced global data distribution, an assumption that rarely holds true in practice due to the prevalence of long-tailed distributions. This work first identifies and diagnoses the severe performance degradation of FAT under long-tailed data, attributing it to skewed feature representations and impaired classifier discriminability. To address this, we propose FedCART, a novel FAT framework that decouples the model into a shared feature extractor and a dual-classifier structure. On the client side, a representation-alignment loss enhances adversarial robustness, while gradient-based class prototypes are extracted for classifier calibration. On the server side, models and prototype sets are aggregated to synthesize balanced virtual features, enabling the re-training of an auxiliary classifier to mitigate long-tailed bias. Extensive experiments demonstrate that FedCART significantly improves both accuracy and robustness, outperforming state-of-the-art FAT methods. To the best of our knowledge, this is the first work to systematically investigate and address FAT under long-tailed distributions, representing a significant step toward practical adversarial robustness in FL. Our code will be publicly available upon acceptance.

Federated Learning (FL) facilitates decentralized model training while preserving data privacy. However, achieving both robust generalization and effective personalization simultaneously in heterogeneous (non-IID) environments remains a formidable challenge. Furthermore, the widespread adoption of proprietary Foundation Models (FMs) introduces a critical requirement for dual privacy: (a) protecting sensitive client data and (b) securing the server's valuable intellectual property. This mandates strictly black-box access to the FM. To address these multifaceted challenges, we introduce FedOT, a novel FL framework optimized for black-box FMs. FedOT employs a shared global task-dependent classifier while facilitating local adaptation through client-specific orthogonal transformations applied externally to the FM embeddings. This architecture inherently guarantees that the FM's internal parameters remain inaccessible and unmodified. By enforcing orthogonality, FedOT effectively mitigates gradient conflicts across diverse clients, which is theoretically bounded, preserves the semantic integrity of the FM representations, and achieves robust performance under significant data heterogeneity. The synergy of global and local parameters optimally balances generalization and personalization, markedly outperforming baseline FL methods across diverse benchmarks. Extensive empirical analysis, including rigorous multi-seed validation and scalability assessments, substantiates the robustness, efficiency, and superior performance of FedOT.

Machine unlearning (MU) seeks to remove the influence of specified data from a trained model in response to privacy requests or data poisoning. While certified unlearning has been analyzed in centralized and server-orchestrated federated settings (via guarantees analogous to differential privacy, DP), the decentralized setting—where peers communicate without a coordinator—remains underexplored. We study certified unlearning in decentralized networks with fixed topologies and propose \methodname, a random-walk procedure that performs one projected gradient ascent step on the forget set at the unlearning client and a geometrically distributed number of projected descent steps on the retained data elsewhere, combined with subsampled Gaussian noise and projection onto a trust region around the original model. We provide (i) convergence guarantees in the convex case and stationarity guarantees in the nonconvex case, (ii) $(\varepsilon,\delta)$ network-unlearning certificates on client views via subsampled Gaussian R\'enyi DP (RDP) with segment-level subsampling, and (iii) deletion-capacity bounds that scale with the forget-to-local data ratio and quantify the effect of decentralization (network mixing and randomized subsampling) on the privacy–utility trade-off. Empirically, on image benchmarks (MNIST, CIFAR-10), \methodname~ matches a given $(\varepsilon,\delta)$ while achieving higher test accuracy than decentralized DP baselines and reducing forget accuracy to random guessing (\(\approx 10\%\)).

Federated learning (FL) in post-deployment settings must adapt to non-stationary data streams across heterogeneous clients without access to ground-truth labels. A major challenge is learning rate selection under client-specific, time-varying distribution shifts, where fixed learning rates often lead to underfitting or divergence. We propose Fed-ADE (Federated Adaptation with Distribution Shift Estimation), an unsupervised federated adaptation framework that leverages lightweight estimators of distribution dynamics. Specifically, Fed-ADE employs uncertainty dynamics estimation to capture changes in predictive uncertainty and representation dynamics estimation to detect covariate-level feature drift, combining them into a per-client, per-timestep adaptive learning rate. We provide theoretical analyses showing that our dynamics estimation approximates the underlying distribution shift and yields dynamic regret and convergence guarantees. Experiments on image and text benchmarks under diverse distribution shifts (label, covariate, and concept) demonstrate consistent improvements over strong baselines. These results highlight that distribution shift-aware adaptation enables effective and robust federated post-adaptation under real-world non-stationarity.

Current referring video segmentation methods typically operate in an offline manner, where sparse frames are first selected for image-level referring segmentation, and the resulting masks are then propagated across the video. Although video sampling captures global context, its isolated processing steps not only complicate optimization but also restrict applicability to real-world streaming scenarios. In this paper, we propose a simple but efficient MLLM-based framework StreamingRVOS, which can extend image-level segmentation to video-level via a streaming pipeline without introducing extra parameters. Specifically, we employ a Semantic Embedding Recycling (SER) method to propagate temporal context across frames, enabling the model to perceive semantic representation in the video. Then, we propose an Online Mask Consistency Perception (OMCP) strategy to adaptively invoke the MLLM to re-perceive the current scene and regenerate the semantic embedding. We conduct extensive experiments on multiple downstream datasets to prove the effectiveness of StreamingRVOS. Compared to previous methods, our method achieves excellent performance in referring video segmentation (1B variant improves upon Sa2VA by 19.2 on the MeViS dataset), while operating at an average speed of 7 FPS under streaming inference on 1 $\times$ A800 GPU.

Understanding social interaction in video requires reasoning over a dynamic interplay of verbal and non-verbal cues: who is speaking, to whom, and with what gaze or gestures.While Multimodal Large Language Models (MLLMs) are natural candidates, simply adding visual inputs yields surprisingly inconsistent gains on social tasks. Our quantitative analysis of cross-modal attention inside state-of-the-art MLLMs reveals a core failure mode: in multi-speaker scenes, visual and textual tokens lack speaker-consistent alignment, exhibiting substantially weaker cross-modal attention than in object-centric images.To address this, we propose a multimodal multi-speaker attention alignment method that can be integrated into existing MLLMs. First, we introduce dynamic cross-modal head selection to identify attention heads most responsible for grounding. Then, an adaptive social-aware attention bias, computed from existing attention patterns and speaker locations, is injected into the attention mechanism. This bias reinforces alignment between a speaker’s visual representation and their utterances without introducing trainable parameters or architectural changes.We integrate our method into three distinct MLLMs (LLaVA-NeXT-Video, Qwen2.5-VL, and InternVL3) and evaluate on three benchmarks (TVQA+, MMSI, OnlineMMSI). Across four social tasks, results demonstrate that our approach improves the ability of MLLMs and achieves state-of-the-art results.Attention visualizations confirm our method successfully focuses the model on speaker-relevant regions, enabling more robust multi-party social reasoning.

Video Temporal Grounding (VTG), the task of localizing video segments from natural language queries, faces significant challenges in open-world applications. These challenges stem from the limited scale and semantic diversity of existing datasets, which lead to a performance gap between common and rare concepts. To overcome these limitations, we introduce OmniVTG, a new large-scale dataset for open-world VTG, coupled with a Self-Correction Chain-of-Thought (CoT) training paradigm designed to enhance the grounding capabilities of Multimodal Large Language Models (MLLMs). Our OmniVTG is constructed via a novel Semantic Coverage Iterative Expansion pipeline, which first identifies gaps in the vocabulary of existing datasets and collects videos that are highly likely to contain these target concepts. For high-quality annotation, we leverage the insight that modern MLLMs excel at dense captioning more than direct grounding and design a caption-centric data engine to prompt MLLMs to generate dense, timestamped descriptions. Beyond the dataset, we observe that simple supervised finetuning (SFT) is insufficient, as a performance gap between rare and common concepts still persists. We find that MLLMs' video understanding ability significantly surpasses their direct grounding ability. Based on this, we propose a Self-Correction Chain-of-Thought (CoT) training paradigm. We train the MLLM to first predict, then use its understanding capabilities to reflect on and refine its own predictions. This capability is instilled via a three-stage pipeline of SFT, CoT finetuning, and reinforcement learning. Extensive experiments show our approach not only excels at open-world grounding in our OmniVTG dataset but also achieves state-of-the-art zero-shot performance on four existing VTG benchmarks.

Spatio-Temporal Video Grounding (STVG) requires models to localize objects both spatially and temporally. Despite recent progress, existing methods struggle with complex and fine-grained spatial semantics in language descriptions, leading to error propagation from temporal to spatial grounding stages. We identify that this fundamental limitation arises from the absence of iterative refinement between temporal and spatial predictions. To address these challenges, we propose SARL-STG, the first RL-based framework for STVG. It progressively refines spatio-temporal grounding through multi-stage optimization, leveraging reinforcement learning to enable dynamic interaction between temporal and spatial modules, where spatial grounding quality serves as feedback to improve temporal localization. Specifically, SARL-STG contains: (1) a unified architecture that seamlessly integrates a pretrained MLLM for temporal reasoning with an open-vocabulary detector for spatial localization, (2) a hierarchical RL training strategy that progresses from coarse temporal to fine-grained spatio-temporal optimization, and (3) a spatial knowledge-injected reward mechanism that uses spatial grounding confidence as discriminative signals for temporal refinement. To facilitate training at scale, we also construct STVG-Wild, a large-scale dataset with diverse spatio-temporal annotations. Experiments demonstrate that our method achieves state-of-the-art performance on multiple benchmarks (HCSTVG, VidSTG, Charades-STA, etc.), significantly reducing error accumulation and enhances both temporal and spatial grounding accuracy.

While Video Large Language Models (Video-LLMs) have shown significant potential in multimodal understanding and reasoning tasks, how to efficiently select the most informative frames from videos remains a critical challenge. Existing methods attempt to optimize frame sampling by reducing inter-frame redundancy or employing unsupervised event localization. However, these approaches often fall short in handling complex instruction-following tasks and scenarios that demand precise temporal modeling, resulting in limited performance in both semantic alignment and temporal reasoning. To address the above challenges, we introduce Instructed Temporal Grounding for Videos (VideoITG), a framework aiming to adaptively customize frame sampling strategies based on user instructions. Specifically, we design the VidThinker pipeline, which automates annotation by generating instruction-conditioned captions, retrieving relevant video segments, and selecting key frames to enable efficient supervision. Using VidThinker, we build the VideoITG-40K dataset with 40K videos and 500K temporal grounding annotations. Our plug-and-play VideoITG model leverages Video-LLMs’ visual-language alignment and reasoning for discriminative frame selection. VideoITG consistently boosts the performance on multiple multimodal video understanding benchmarks, demonstrating its effectiveness and potential.

Referring video object segmentation (RVOS) aims to segment objects within a video according to natural language expressions. Unlike earlier works focusing on static single-object scenarios, recent studies address more complex motion scenes. Previous methods typically adopt a query-based, logically multi-stage pipeline to handle these scenarios. However, this paradigm learns trajectory consistency modeling and multimodal fusion from scratch, which often leads to trajectory inconsistencies and insufficient multimodal understanding. To address these limitations, we propose DeRVOS, a framework that decouples RVOS into two key branches: consistent trajectory generation and multimodal understanding. We extract temporally consistent object representations using a powerful pretrained instance trajectory generation model and perform cross-modal alignment via a unified multimodal encoder, enabling upstream modeling of trajectory consistency and vision-language understanding. This design reduces RVOS to the task of modeling the relationship between referring expressions and instance trajectories. To connect the two branches and enable efficient motion-aware semantic understanding, we introduce the Trajectory Alignment and Implicit Selection (TAIS) module, which progressively performs cross-frame multimodal alignment and motion-guided implicit trajectory selection. Extensive experiments demonstrate that DeRVOS achieves state-of-the-art results on both traditional RVOS benchmarks and the challenging MeViS dataset, surpassing LVLM-based methods by 4.7%.

Unified models aim to support both understanding and generation by encoding images into discrete tokens and processing them alongside text within a single autoregressive framework. This unified design offers architectural simplicity and cross-modal synergy, which facilitates shared parameterization, consistent training objectives, and seamless transfer between modalities. However, the large number of visual tokens required by such models introduces substantial computation and memory overhead, and this inefficiency directly hinders deployment in resource constrained scenarios such as embodied AI systems. In this work, we propose a unified token compression algorithm UniCompress that significantly reduces visual token count while preserving performance on both image understanding and generation tasks. Our method introduces a plug-in compression and decompression mechanism guided with learnable global meta tokens. The framework is lightweight and modular, enabling efficient integration into existing models without full retraining. Experimental results show that our approach reduces image tokens by up to (4\times), achieves substantial gains in inference latency and training cost, and incurs only minimal performance degradation, which demonstrates the promise of token-efficient unified modeling for real world multimodal applications.

Unlike offline processing, streaming video vision-language models face two fundamental constraints: causality and accumulation.Causality prevents access to future frames that offline methods exploit, while accumulation causes tokens to grow unbounded, creating efficiency bottlenecks.However, existing approaches only regulate post-LLM kv-cache, leaving costly pre-LLM prefill unchanged.We introduce StreamingTOM, a training-free, plug-and-play two-stage framework that addresses both pre-LLM and post-LLM bottlenecks with predictable latency.Causal Temporal Reduction imposes a fixed per-frame budget and selects tokens based on adjacent-frame changes and token saliency, drastically reducing per-frame prefill cost by processing only a compact subset of visual tokens per frame instead of all visual tokens.Online Quantized Memory stores tokens in 4-bit format, retrieves relevant groups on demand, and dequantizes them, keeping the active kv-cache bounded regardless of stream length.Experiments demonstrate our method achieves $15.7\times$ kv-cache compression, $1.2\times$ lower peak memory and $2\times$ faster TTFT compared to prior SOTA.StreamingTOM maintains state-of-the-art accuracy among training-free methods with an average of $63.8\%$ on offline benchmarks and $55.8\%/3.7$ on RVS.These results highlight the practical benefits of our two-stage approach for efficient streaming video understanding with bounded growth.

The heavy computational burden of Large Vision-Language Models (LVLMs) stems primarily from the lengthy visual token sequences generated by their vision encoders. To mitigate this, recent work has shifted towards pruning tokens within the vision encoder. However, we observe that these methods predominantly rely on a suboptimal decoupled heuristic method. This method is conceptually flawed: it is prone to sampling collapse, fails to fundamentally eliminate token redundancy, and tends to systematically discard secondary yet important semantic clusters.Addressing this limitation, this paper proposes to formalize visual token pruning as a unified Representativeness Optimization problem. We introduce SCoRe (Salience-Coverage Reduction), a unified optimization method theoretically grounded in the Weighted k-Center Problem. SCoRe constructs the final token set by greedily selecting tokens—at each iteration, choosing the token that maximizes the current set's unified representativeness score, thereby achieving the optimization of global representativeness. Extensive experiments demonstrate that SCoRe achieves State-of-the-Art (SOTA) performance across multiple benchmarks. Notably, with negligible computational overhead, our method reduces tokens by 94.4% while retaining 95% of the full performance.

Post-training quantization (PTQ) emerges as a vital technique for efficiently compressing large-scale models, with weight-compensation methods like GPTQ (symmetric calibration) and GPTAQ (asymmetric calibration) showing remarkable success. However, directly applying these methods to Vision-Language Models (VLMs) reveals two critical limitations: 1) their reliance on standard rounding is suboptimal for the asymmetric objective, failing to account for residual-induced shifts in the optimal quantization target; and 2) they uniformly process input channels across modalities, overlooking the distinct information densities of vision and language tokens. In this paper, we introduce VLM-PTQ, a new PTQ asymmetric framework for VLMs. First, we derive a closed-form correction term for the quantization point, which explicitly accounts for the propagated residual and the corresponding inverse Hessian column, yielding a better local optimum than standard rounding. Second, we propose a modality-aware quantization that differentiates channel importance between vision and language tokens, enabling the quantizer to prioritize salient channels through a lightweight fusion coefficient search. Our method extends GPTAQ with minimal overhead while achieving significant performance improvements in low-bit scenarios. Extensive experiments demonstrate that VLM-PTQ achieves state-of-the-art results, effectively compressing models from 1B to 72B parameters on a single GPU.

Vision-Language Models (VLMs) have demonstrated strong capability in a wide range of tasks such as visual recognition, document parsing, and visual grounding. Nevertheless, recent works show that while VLMs often manage to capture the correct image region corresponding to the question, they do not necessarily produce the correct answers. In this work, we demonstrate that this misalignment could be attributed to suboptimal information flow within VLMs, where text tokens distribute too much attention to irrelevant visual tokens, leading to incorrect answers. Based on the observation, we show that modulating the information flow during inference can improve the perception capability of VLMs. The idea is that text tokens should only be associated with important visual tokens during decoding, eliminating the interference of irrelevant regions. To achieve this, we propose a token dynamics-based method to determine the importance of visual tokens, where visual tokens that exhibit distinct activation patterns during different decoding stages are viewed as important. We apply our approach to representative open-source VLMs and evaluate on various datasets including visual question answering, visual grounding and counting, optical character recognition, and object hallucination. The results show that our approach significantly improves the performance of baselines. The code will be made available.

Post-Training Quantization (PTQ) has emerged as an effective technique for alleviating the substantial computational and memory overheads of Vision-Language Models (VLMs) by compressing both weights and activations without retraining the full model. Existing PTQ methods primarily rely on static identification and global compensation of sensitive or outlier channels, yet they often overlook the distributional differences of these important channels across inputs, leading to unsatisfactory quantization. In this work, we observe that the distributions and occurrence frequencies of important channels vary significantly both across modalities and among tokens, even within the same modality. Accordingly, we propose Quant Experts (QE), a token-aware adaptive error compensation with mixture-of-experts for VLMs quantization. QE divides the important channels into token-independent and token-dependent groups. For the former, a shared expert is designed for most tokens to compensate for global quantization error using a low-rank adapter. For the latter, routed experts including multiple routed low-rank adapters are elaborated to compensate for local quantization error related to specific tokens. Extensive experiments demonstrate that QE consistently enhances task accuracy across various quantization settings and model scales, ranging from 2B to 70B parameters, while maintaining performance comparable to full-precision models.

Large Vision-Language Models (LVLMs) excel in visual understanding and reasoning, but the excessive visual tokens lead to high inference costs. Although recent token reduction methods mitigate this issue, they mainly target single-turn Visual Question Answering (VQA), leaving the more practical multi-turn VQA (MT-VQA) scenario largely unexplored. MT-VQA introduces additional challenges, as subsequent questions are unknown beforehand and may refer to arbitrary image regions, making existing reduction strategies ineffective. Specifically, current approaches fall into two categories: prompt-dependent methods, which bias toward the initial text prompt and discard information useful for subsequent turns; prompt-agnostic ones, which, though technically applicable to multi-turn settings, rely on heuristic reduction metrics such as attention scores, leading to suboptimal performance. In this paper, we propose a learning-based prompt-agnostic method, termed MetaCompress, overcoming the limitations of heuristic designs. We begin by formulating token reduction as a learnable compression mapping, unifying existing formats such as pruning and merging into a single learning objective. Upon this formulation, we introduce a data-efficient training paradigm capable of learning optimal compression mappings with limited computational costs. Extensive experiments on MT-VQA benchmarks and across multiple LVLM architectures demonstrate that MetaCompress achieves superior efficiency–accuracy trade-offs while maintaining strong generalization across dialogue turns. Our code will be released to facilitate future research.

Modern multi-label image classification models suffer from a critical reliance on spurious correlations, failing to learn the underlying causal mechanisms.Many causality-inspired methods are impractical, demanding box-level supervision that is rarely available in real-world datasets.Others rely on static confounder dictionaries, which are inherently inflexible and fail to capture complex biases or adapt to feature space changes during training.To address this, we present prototype-based causal intervention (ProCI), a novel framework that approximates the backdoor adjustment using only image-level supervision. It models confounders as learnable contextual prototypes which, unlike traditional prototypes designed for discriminative features, are engineered to represent class-wise co-occurring bias.These prototypes are learned dynamically within a stable memory and leveraged to construct sample-specific bias vectors for an adaptive feature adjustment, effectively counteracting spurious correlations.Experiments on MS-COCO, Pascal VOC, and the challenging Sewer-ML dataset validate our approach. ProCI achieves competitive performance on standard benchmarks while setting a new state-of-the-art on the highly-confounded Sewer-ML. It outperforms the previous best model by a remarkable +5.44 points on the primary $F2_{CIW}$ metric. These results demonstrate the effectiveness of our approach in mitigating complex real-world biases using only image-level supervision.

Coreset selection compresses large datasets into compact, representative subsets, reducing the energy and computational burden of training deep neural networks. Existing methods are either: (i) DNN-based, which are inherently coupled with network-specific parameters, inevitably introducing architectural bias and compromising generalization; or (ii) DNN-free, which utilize heuristics that lack rigorous theoretical guarantees for stability and accuracy. Neither approach explicitly constrains distributional equivalence of the representative subsets, largely because continuous distribution matching is broadly considered inapplicable to discrete dataset sampling. Furthermore, prevalent distribution metrics (e.g., MSE, KL, MMD, and CE) are often incapable of accurately capturing higher-order moment differences. These deficiencies lead to suboptimal coreset performance, preventing the selected coreset from being truly equivalent to the original dataset.In this work, we propose FAST (Frequency-domain Aligned Sampling via Topology), the first DNN-free distribution-matching coreset selection framework that formulates coreset selection as a graph-constrained optimization problem grounded in spectral graph theory and employs the Characteristic Function Distance (CFD) to capture full distributional information (i.e., all moments and intrinsic correlations) in the frequency domain. We further discover that naive CFD suffers from a “vanishing phase gradient” issue in medium and high-frequency regions; to address this, we introduce an Attenuated Phase-Decoupled CFD. Furthermore, for better convergence, we design a Progressive Discrepancy-Aware Sampling strategy that progressively schedules frequency selection from low to high. This preserves global structures before refining local details, enabling accurate matching with few frequencies while preventing overfitting. Extensive experiments demonstrate that FAST significantly outperforms state-of-the-art coreset selection methods across all evaluated benchmarks, achieving an average accuracy gain of 9.12%. Compared to other baseline coreset methods, it reduces power consumption by 96.57% and achieves a 2.2$\times$ average speedup even on CPU with 1.7GB of memory, underscoring its high performance and energy efficiency.

Point-level weakly-supervised temporal sentiment localization (P-WTSL) aims to detect sentiment-relevant segments in untrimmed multimodal videos using timestamp sentiment annotations, which greatly reduces the costly frame-level labeling. To tackle the intrinsic challenges of imprecise sentiment boundaries in P-WTSL, we propose the Face-guided Sentiment Boundary Enhancement Network FSENet, a unified framework that leverages fine-grained facial features to guide sentiment localization. Specifically, our approach first introduces the Face-guided Sentiment Discovery (FSD) module, which integrates facial features into multimodal interaction via dual-branch modeling for effective sentiment stimuli clues; We then propose the Point-aware Sentiment Semantics Contrast (PSSC) strategy to discriminate sentiment semantics of candidate points (frame-level) near annotation points via contrastive learning, thereby enhancing the model's ability to recognize sentiment boundaries. At last, we design the Boundary-aware Sentiment Pseudo-label Generation (BSPG) approach to convert sparse point annotations into temporally smooth supervisory pseudo-labels. Extensive experiments and visualizations on the benchmark demonstrate the effectiveness of our framework, achieving state-of-the-art performance under full supervision, video-level, and point-level weak supervision, thereby showcasing the strong generalization ability of our FSENet across different annotation settings. The code will be open source to the public.

Partial label learning (PLL) is a weakly supervised learning, where each instance is assigned a set of candidate labels and only one is true. However, due to potentially inaccurate annotations, existing PLL algorithms disambiguate labeling by minimizing the prediction loss, which leaves the model unaware of its prediction credibility. To address this issue, this paper proposes the evidential deep partial label learning (ED-PLL) to quantify disambiguation uncertainty, aiming to achieve candidate label disambiguation and reliability prediction. Firstly, we extend the evidence modeling mechanism to PLL, treating the candidate label set as the source of evidence for the label hypothesis, and using belief and credibility to model classification uncertainty, thereby guiding a more reliable disambiguation process. Meanwhile, we propose the expectation calculation under the Dirichlet distribution of non-candidate labels, which suppresses the output of non-candidate labels by using consistency regularization to further improve the accuracy of disambiguation. Furthermore, a conflict-aware regularization is proposed to evaluate the degree of conflict, which measures the consistency between instances within the class by combining the differences in the distribution of prediction results and model uncertainty, and thus improves the robustness of the model. In addition, this paper theoretically analyzes our method from the perspective of the Expectation-Maximization (EM) algorithm, and the ED-PLL is compatible with any deep network or stochastic optimizer. Experiments on benchmark and real datasets verify the effectiveness of the proposed algorithm.

Current audio pre-training seeks to learn unified representations for broad audio understanding tasks, but it remains fragmented and is fundamentally bottlenecked by its reliance on weak, noisy, and scale-limited labels. Drawing lessons from vision's foundational pre-training blueprint, we argue that the audio field must first establish its own large-scale, strong supervision framework. We introduce a new data-centric pipeline that leverages a high-fidelity captioner to create SOTA-quality captions and the first Unified Tag System (UTS) that bridges speech, music, and environmental sounds. We then conduct a systematic comparative study of different pre-training objectives on these strong source data. Our experiments suggest that data quality and coverage are the primary drivers of performance, while the choice of objective dictates downstream task specialization.

Learning with noisy labels (LNL) has received growing attention, with most prior work following the paradigm of clean-sample reliance (e.g., sample selection). However, this reliance also imposes intrinsic limitations, as overfitting to even a few noisy samples is inevitable, creating a major bottleneck for further improvement. This limitation motivates us to go beyond mere clean-sample reliance and explore how to actively forget corrupted knowledge already internalized by models while suppressing further noise assimilation. To this end, we propose FINE, a fundamentally novel perspective for LNL that unifies active ForgettIng via machine unlearning (MU) and Noise supprEssion via negative learning (NL) within a cohesive framework. Specifically, we first reveal two key stages of noise fitting: early-stage generalized learning and later-stage noise overfitting. To actively forget early-stage noise accumulation, we introduce an MU-based module that employs a negative cross-entropy loss to erase corrupted knowledge, while an NL-based module leveraging complementary labels suppresses later-stage overfitting and mitigates reliance on noisy supervision. These modules act synergistically as plug-and-play regularizers, seamlessly integrating into existing baselines. Finally, extensive experiments on both synthetic and real-world noisy benchmarks demonstrate that our FINE consistently boosts robustness and generalization.

Open-set semi-supervised learning (OSSL) has achieved notable progress in exploiting unlabeled data, yet most existing methods overlook the distinct sensitivities of in-distribution (ID) and out-of-distribution (OOD) samples to semantic-preserving perturbations, resulting in unreliable OOD sample filtering. We address this gap by leveraging the behavioral difference between ID and OOD samples under perturbations and extend it into a representation-level signal for reliable OOD filtering. Specifically, we propose a novel filtering strategy, Perturbation-Aware Filtering (PAF), which identifies OOD samples by measuring the representation instability under semantic-preserving perturbations. We then integrate PAF into a carefully designed two-stage training framework, allowing the model to exploit abundant unlabeled data in the first stage and gradually adapt to the open-set setting with limited labels in the second stage. Extensive experimental results on widely-used OSSL benchmarks demonstrate that our proposed PAF approach achieves superior performance compared to state-of-the-art OSSL methods. The source code will be released publicly.

Recently, contrastive learning (CL) plays an important role in exploring complementary information for multi-view clustering (MVC) and has attracted increasing attention. Nevertheless, real-world multi-view data suffer from data incompleteness or noise, resulting in rare-paired samples or mis-paired samples which significantly challenges the effectiveness of CL-based MVC. That is, rare-paired issue prevents MVC from extracting sufficient multi-view complementary information, and mis-paired issue causes contrastive learning to optimize the model in the wrong direction. To address these issues, we propose a unified CL-based MVC framework for enhancing clustering effectiveness on incomplete and noise multi-view data. First, to overcome the rare-paired issue, we design a global-graph guided contrastive learning, where all view samples construct a global-view affinity graph to form new sample pairs for fully exploring complementary information. Second, to mitigate the mis-paired issue, we propose a local-graph weighted contrastive learning, which leverages local neighbors to generate pair-wise weights to adaptively strength or weaken the pair-wise contrastive learning. Our method is imputation-free and can be integrated into a unified global-local graph-guided contrastive learning framework. Extensive experiments on both incomplete and noise settings of multi-view data demonstrate that our method achieves superior performance compared with state-of-the-art approaches.

Out-of-distribution (OOD) detection is essential for the safe deployment of machine learning models. Extensive work has focused on devising various scoring functions for detecting OOD samples, while only a few studies focus on training neural networks using certain model calibration objectives, which often lead to a compromise in predictive accuracy and support only limited choices of scoring functions. In this work, we first identify the feature collapse phenomena in Logit Normalization (LogitNorm), then propose a novel hyperparameter-free formulation that significantly benefits a wide range of post-hoc detection methods. To be specific, we devise a feature distance-awareness loss term in addition to LogitNorm, termed $\textbf{ELogitNorm}$, which enables improved OOD detection and in-distribution (ID) confidence calibration. Extensive experiments across standard benchmarks demonstrate that our approach outperforms state-of-the-art training-time methods in OOD detection while maintaining strong ID classification accuracy.

Vision-Language-Action (VLA) models for autonomous driving often hit a performance plateau during Reinforcement Learning (RL) optimization. This stagnation arises from exploration capabilities constrained by previous Supervised Fine-Tuning (SFT), leading to "persistent failures" in long-tail scenarios. In these critical situations, all explored actions yield a zero-value driving score. This information-sparse reward signals a failure, yet fails to identify its root cause—whether it is due to incorrect planning, flawed reasoning, or poor trajectory execution. To address this limitation, we propose VLA with Explicit Learning from Failures (ELF-VLA), a framework that augments RL with structured diagnostic feedback. Instead of relying on a vague scalar reward, our method produces detailed, interpretable reports that identify the specific failure mode. The VLA policy then leverages this explicit feedback to generate a Feedback-Guided Refinement. By injecting these corrected, high-reward samples back into the RL training batch, our approach provides a targeted gradient, which enables the policy to solve critical scenarios that unguided exploration cannot. Extensive experiments demonstrate that our method unlocks the latent capabilities of VLA models, achieving state-of-the-art (SOTA) performance on the public Navsim benchmark for overall PDMS, EPDMS score and high-level planning accuracy.

Creating realistic and simulation-ready 3D assets is crucial for autonomous driving research and virtual environment construction. However, existing 3D vehicle generation methods are often trained on synthetic data with significant domain gaps from real-world distributions. The generated models often exhibit arbitrary poses and undefined scales, resulting in poor visual consistency when integrated into driving scenes. In this paper, we present Unposed-to-3D, a novel framework that learns to reconstruct 3D vehicles from real-world driving images using image-only supervision. Our approach consists of two stages. In the first stage, we train a image-to-3D reconstruction network using posed images with known camera parameters. In the second stage, we remove camera supervision and use a camera prediction head that directly estimates the camera parameter from unposed images. The predicted pose is then used for differentiable rendering to provide self-supervised photometric feedback, enabling the model to learn 3D geometry purely from unposed data. To ensure simulation readiness, we further introduce a scale-aware module to predict real-world size information, and a harmonization module that adapts the generated vehicles to the target driving scene with consistent lighting and appearance. Extensive experiments demonstrate that Unposed-to-3D effectively reconstructs realistic, pose-consistent, and harmonized 3D vehicle models from real-world images, providing a scalable path toward creating high-quality assets for driving scene simulation and digital twin environments.

The end-to-end (E2E) paradigm, which maps sensor inputs directly to driving decisions, has recently attracted significant attention due to its unified modeling capability and scalability. However, ensuring safety in this unified framework remains one of the most critical challenges. In this work, we propose SafeDrive, an E2E planning framework designed to perform explicit and interpretable safety reasoning through a trajectory-conditioned Sparse World Model. SafeDrive comprises two complementary networks: the Sparse World Network (SWNet) and the Fine-grained Reasoning Network (FRNet). SWNet constructs trajectory-conditioned sparse worlds that simulate the future behaviors of critical dynamic agents and road entities, providing interaction-centric representations for downstream reasoning. FRNet then evaluates agent-specific collision risks and temporal adherence to drivable regions, enabling precise identification of safety-critical events across future timesteps. SafeDrive achieves state-of-the-art performance on both open-loop and closed-loop benchmarks. On NAVSIM, it records a PDMS of 91.6 and an EPDMS of 87.5, with only 61 collisions out of 12,146 scenarios (0.5%). On Bench2Drive, SafeDrive attains a 66.6% driving score.

Vehicle trajectory prediction is a critical technology for safe and efficient autonomous driving.However, its generalization and scalability have long been hindered by a heavy reliance on real-time, online priors.To break this bottleneck, we introduce RAG-TP, a general framework that reframes the problem from relying on uncertain online perception to retrieving from a large-scale, structured, offline knowledge base.The core of RAG-TP is to enhance predictions at inference time by dynamically querying a pre-built, heterogeneous knowledge base rich with scene topologies and motion patterns, using the retrieved historical experiences as priors.We further design a dynamic fusion module based on a learnable Mixture-of-Experts (MoE), which intelligently weights and integrates the multi-source retrieved knowledge via cross-attention to generate a high-density context for the final multi-modal trajectory decoding.By decoupling online inference from offline knowledge, this retrieval-augmented approach grounds predictions in a vast structured database, thereby mitigating model hallucination and compensating for unreliable priors to significantly enhance robustness and domain adaptation.Extensive experiments demonstrate that RAG-TP achieves excellent performance in both map-based and map-free settings, surpassing existing map-free methods while achieving performance comparable to state-of-the-art (SOTA) map-based models.It demonstrates significant advantages, particularly in cross-domain and zero-shot generalization tasks.Our work provides a promising and effective technical pathway toward building more scalable and robust prediction systems for autonomous driving.

Reliable long-horizon trajectory prediction requires both high positional accuracy and physically plausible temporal motion consistency. However, existing methods suffer from two fundamental limitations. First, they overlook the inherent difference in prediction logic: near-future trajectories are primarily governed by historical dynamics, whereas distant-future behaviors are driven by high-level semantic context. Yet, most methods employ a unified decoding pathway that blurs the temporal distinction.Second, although the near future is relatively easier to predict, existing methods lack mechanisms for coherent trajectory propagation across time horizons, often resulting in kinematically implausible predictions with inconsistent heading evolution and degraded long-horizon performance. To address these challenges, we propose NDPNet, a dual-stage architecture that decouples near- and distant-horizon modeling into specialized pathways, with a dedicated transition module ensuring smooth temporal bridging. Furthermore, we introduce a novel motion-aware coherence loss that explicitly embeds kinematic priors to enforce trajectory consistency. Extensive experiments show that NDPNet achieves SOTA performance on Argoverse 2 and WOMD. Notably, on WOMD, it ranks 1$^{\text{st}}$ in both minFDE${}_6$ and minADE${}_6$ across all standard horizons (3s, 5s, 8s) without ensemble learning or NMS post-processing, and is the first to achieve sub-1.75 minFDE${}_6$ for 8s prediction, surpassing prior methods by a large margin. The code will be released subsequently.

Visual-Inertial Odometry (VIO) is a critical component for robust ego-motion estimation, enabling foundational capabilities such as autonomous navigation in robotics and real-time 6-DoF tracking for augmented reality.Existing methods face a well-known trade-off: filter-based approaches are efficient but prone to drift, while optimization-based methods, though accurate, rely on computationally prohibitive Visual-Inertial Bundle Adjustment (VIBA) that is difficult to run on resource-constrained platforms.Rather than removing VIBA altogether, we aim to reduce how often and how heavily it must be invoked. To this end, we cast two key design choices in modern VIO, when to run the visual frontend and how strongly to trust its output, as sequential decision problems, and solve them with lightweight reinforcement learning (RL) agents. Our framework introduces a lightweight, dual-pronged RL policy that serves as our core contribution: (1) a Select Agent intelligently gates the entire VO pipeline based only on high-frequency IMU data; and (2) a composite Fusion Agent that first estimates a robust velocity state via a supervised network , before an RL policy adaptively fuses the full (p, v, q) state.Experiments on the EuRoC MAV and TUM-VI datasets show that, in our unified evaluation, the proposed method achieves a more favorable accuracy–efficiency–memory trade-off than prior GPU-based VO/VIO systems: it attains the best average ATE while running up to $1.77\times$ faster and using less GPU memory. Compared to classical optimization-based VIO systems, our approach maintains competitive trajectory accuracy while substantially reducing computational load.

Controllable driving scene generation is critical for realistic and scalable autonomous driving simulation, yet existing approaches struggle to jointly achieve photorealism and precise control. We introduce \textbf{HorizonForge}, a unified framework that reconstructs scenes as editable Gaussian Splats and Meshes, enabling fine-grained 3D manipulation and language-driven vehicle insertion. Edits are rendered through a noise-aware video diffusion process that enforces spatial and temporal consistency, producing diverse scene variations in a single feed-forward pass without per-trajectory optimization. To standardize evaluation, we further propose \textbf{HorizonSuite}, a comprehensive benchmark spanning ego- and agent-level editing tasks such as trajectory modifications and object manipulation. Extensive experiments show that Gaussian Splatting delivers substantially higher fidelity than alternative 3D representations, and that temporal priors from video diffusion are essential for coherent synthesis. Combining these finding, \textbf{HorizonSuite} establishes a simple yet powerful paradigm for photorealistic, controllable driving simulation. achieving an 83.4\% user-preference gain and a 25.19\% FID improvement over the second best state-of-the-art method.

Online High-Definition (HD) map construction is pivotal for autonomous driving. While recent approaches leverage historical temporal fusion to improve performance, we identify a critical safety flaw in this paradigm: it is inherently "spatially backward-looking." These methods predominantly enhance map reconstruction in traversed areas, offering minimal improvement for the unseen road ahead. Crucially, our analysis of downstream planning tasks reveals a severe asymmetry: while rearward perception errors are often tolerable, inaccuracies in the forward region directly precipitate hazardous driving maneuvers. To bridge this safety gap, we propose AMap, a novel framework for Ahead-aware online HD Mapping. We pioneer a "distill-from-future" paradigm, where a teacher model with privileged access to future temporal contexts guides a lightweight student model restricted to the current frame. This process implicitly compresses prospective knowledge into the student model, endowing it with "look-ahead" capabilities at zero inference-time cost. Technically, we introduce a Multi-Level BEV Distillation strategy with spatial masking and an Asymmetric Query Adaptation module to effectively transfer future-aware representations to the student's static queries. Extensive experiments on the nuScenes and Argoverse 2 benchmark demonstrate that AMap significantly enhances current-frame perception. Most notably, it outperforms state-of-the-art temporal models in critical forward regions while maintaining the efficiency of current-frame inference.

We introduce DFM-Drive, a vision–language–action (VLA) model that casts ego-trajectory planning as discrete flow matching over a structured token space. In contrast to autoregressive decoders, DFM-Drive performs fully parallel, bidirectional denoising, enabling coarse-to-fine refinement with a tunable compute–accuracy trade-off. Specifically, the approach combines a metric-aligned numerical tokenizer that preserves scalar geometry via triplet-margin learning, a geometry-aware flow objective and a simulator-guided GRPO alignment that integrates safety, ego progress, and comfort rewards while retaining parallel generation. A multi-stage adaptation converts a pre-trained auto-regressive backbone (Janus-1.5B) from causal decoding to non-causal flow model and strengthens road-scene competence through continued multimodal pretraining. Thanks to the inherent nature of consistency model training and parallel decoding inference, DFM-Drive achieves superior closed-loop performance against autoregressive and diffusion-based VLA baselines, with 1-step inference attaining 88.7 PDMS and 5-step inference reaching 90.3 PDMS on NAVSIM v1 benchmark. These results establish discrete flow matching as a new promising paradigm for end-to-end autonomous driving.

Diffusion-based planners have emerged as a promising approach for human-like trajectory generation in autonomous driving. Recent works incorporate reinforcement fine-tuning to enhance the robustness of diffusion planners through reward-oriented optimization in a generation–evaluation loop. However, they struggle to generate multi-modal, scenario-adaptive trajectories, hindering the exploitation efficiency of informative rewards during fine-tuning. To resolve this, we propose PlannerRFT, a sample-efficient reinforcement fine-tuning framework for diffusion-based planners. PlannerRFT adopts a dual-branch optimization that simultaneously refines the trajectory distribution and adaptively guides the denoising process toward more promising exploration, without altering the original inference pipeline. To support parallel learning at scale, we develop nuMax, an optimized simulator that achieves 10 times faster rollout compared to native nuPlan. Extensive experiments shows that PlannerRFT yields state-of-the-art performance with distinct behaviors emerging during the learning process. Code and simulator would be released.

Most existing underwater instance segmentation approaches are constrained by close-vocabulary prediction, limiting their ability to recognize novel marine categories. To support evaluation, we introduce MARIS (Marine Open-Vocabulary Instance Segmentation), the first large-scale fine-grained benchmark for underwater Open-Vocabulary (OV) Instance segmentation (UOVIS), featuring a limited set of seen categories and diverse unseen categories. Although OV instance segmentation has shown promise on natural images, our analysis reveals that transfer to underwater scenes suffers from severe visual degradation (e.g., color attenuation) and semantic misalignment caused by lack underwater class definitions. To address these issues, we propose a unified framework with two complementary components. The Geometric Prior Enhancement Module (GPEM) leverages stable part-level and structural cues to maintain object consistency under degraded visual conditions. The Semantic Alignment Injection Mechanism (SAIM) enriches language embeddings with domain-specific priors, mitigating semantic ambiguity and improving recognition of unseen categories. Experiments show that our framework consistently outperforms existing OV baselines both In-Domain and Cross-Domain setting on MARIS, establishing a strong foundation for future underwater perception research. The code of this paper can be found in the supplementary materials.

X-ray contraband detection is critical for public safety. However, current methods primarily rely on bounding box annotations, which limit model generalization and performance due to the lack of pixel-level supervision and real-world data. To address these limitations, we introduce XSeg. To the best of our knowledge, XSeg is the largest X-ray contraband segmentation dataset to date, including 98,644 images and 295,932 instance masks, and contains the latest 30 common contraband categories. The images are sourced from public datasets and our synthesized data, filtered through a custom data cleaning pipeline to remove low-quality samples. To enable accurate and efficient annotation and reduce manual labeling effort, we propose Adaptive Point SAM (APSAM), a specialized mask annotation model built upon the Segment Anything Model (SAM). We address SAM’s poor cross-domain generalization and limited capability in detecting stacked objects by introducing an Energy-Aware Encoder that enhances the initialization of the mask decoder, significantly improving sensitivity to overlapping items. Additionally, we design an Adaptive Point Generator that allows users to obtain precise mask labels with only a single coarse point prompt. Extensive experiments on XSeg demonstrate the superior performance of APSAM.

Vision-Language models (e.g., CLIP) facilitate the development of open-vocabulary camouflaged object segmentation (OVCOS), but existing methods still rely on mask annotations for fully-supervised training. In contrast, the training-free paradigm can rapidly process unseen data, representing a highly promising solution. However, in camouflage scenarios, existing training-free methods utilize sparse textual prompts and ignore the category similarity between visual patches, leading to inadequate object binding capability. To alleviate these issues, we propose a fine-grained object binding and adaptive hybrid prompt framework for training-free OVCOS. The framework first employs multimodal large language models (MLLMs) to explicitly model fine-grained textual descriptions of camouflaged objects and background. Building on this, we construct a semantic probe to decouple object and background features and explicitly model category similarity between visual patches via semantic consistency ranking, thereby achieving accurate object binding. Subsequently, we propose an entropy-guided text embedding adjustment strategy to adjust textual embeddings, aiming to further enhance fine-grained object binding. Finally, we utilize an adaptive hybrid prompt generation strategy to generate hybrid prompts, assisting SAM in accurately segmenting camouflaged objects. Experimental results on the OVCamo benchmark demonstrate that our method achieves excellent performance, significantly surpassing the advanced training-free ResCLIP.

The Segment Anything Model 2 (SAM2) has emerged as a foundation model for universal segmentation. Owing to its generalizable visual representations, SAM2 has been successfully applied to various downstream tasks. However, extending SAM2 to RGB-D video salient object detection (RGB-D VSOD) task encounters three challenges including limited spatial modeling of linear LoRA, insufficient employment of SAM's multi-scale features, and dependence of initialization on explicit prompts. To address the issues, we present **M**ulti-**M**odal **M**ixture-of-Experts with **M**emory-Augmented **SAM** (**M**$^4$-**SAM**), which equips SAM2 with modality-related PEFT, hierarchical feature fusion, and prompt-free memory initialization. Firstly, we inject **Modality-Aware MoE-LoRA**, which employs convolutional experts to encode local spatial priors and introduces a modality dispatcher for efficient multi-modal fine-tuning, into SAM2's encoder. Secondly, we deploy **Gated Multi-Level Feature Fusion**, which hierarchically aggregates multi-scale encoder features with an adaptive gating mechanism, to balance spatial details and semantic context. Finally, to conduct zero-shot VSOD without manual prompts, we utilize a **Pseudo-Guided Initialization**, where a coarse mask is regarded as a pseudo prior and used to bootstrap the memory bank. Extensive experiments demonstrate that M$^4$-SAM achieves the state-of-the-art performance across all evaluation metrics on three public RGB-D VSOD datasets.

Remote sensing image segmentation is critical for a range of applications, including natural disaster monitoring and precision agriculture. Open-vocabulary segmentation enhances flexibility by removing fixed category constraints, enabling more fine-grained and adaptive scene understanding. Unlike CLIP’s original pretraining objective, which emphasizes global image-text alignment, segmentation tasks require accurate and discriminative patch-level representations to support precise pixel-wise predictions. As a result, the quality of attention maps—particularly those generated in the final transformer layers—plays a pivotal role in guiding inter-region interactions. However, current methods generate suboptimal representations when capturing the complex spatial hierarchies in remote sensing. We address this gap by optimizing CLIP's 197×197 attention matrix through three key modifications: (1) substituting the 196×196 patch-to-patch submatrix with intermediate-layer feature similarities to preserve spatial structures; (2) prioritizing intermediate-layer attention for global-to-local (class-to-patch) token alignment to reduce classification interference; (3) disabling the \texttt{[CLS]} token's self-attention to mitigate bias. Extensive experiments on eight remote sensing benchmarks and two building/road extraction datasets demonstrate that our method achieves state-of-the-art performance among existing training-free approaches.

Test-Time Adaptive Segmentation (TTA-Seg) aims to adapt a trained segmentation model to test data under distribution shift in an unsupervised manner. Existing approaches typically utilize class-wise prototypes to capture and transfer the source distribution, but inevitably neglecting the diversity within source samples. In this paper, we propose a new test-time adaptation paradigm based on the mixture-of-experts (MoE), where domain experts are designed to 1) better capture the source distribution, and 2) dynamically adjust their contribution in test case prediction. Specifically, during source training, prototypes are derived as the class-wise average for source pixel features. We then generate multiple experts through clustering these prototypes, providing each class with several experts with enhanced representativeness. At test time, each pixel's prediction is drawn from all experts' knowledge in an adaptive manner, \ie, a gating network assigns weight according to pixel-expert correlation. To optimize the system, we devise a min-max entropy optimization scheme for the gating network but keeping the rest frozen, minimizing the entropy of model prediction but maximizing the entropy in expert selection. Consequently, the model is urged to derived confident predictions with effective utilization of domain experts, hence promoting the adaptation. Experiments on two scenarios, Test-time Adaptation (TTA) and the more challenging continual TTA, demonstrate that our approach achieves the new state-of-the-art.

Video Object‑Centric Learning seeks to decompose raw videos into a small set of object slots, but existing slot‑attention models often suffer from severe over‑fragmentation.This is because the model is implicitly encouraged to occupy all slots to minimize the reconstruction objective, thereby representing a single object with multiple redundant slots.We tackle this limitation with a reconstruction‑guided slot curriculum (SlotCurri). Training starts with only a few coarse slots and progressively allocates new slots where reconstruction error remains high, thus expanding capacity only where it is needed and preventing fragmentation from the outset. Yet, during slot expansion, meaningful sub‑parts can emerge only if coarse‑level semantics are already well separated; however, with a small initial slot budget and an MSE objective, semantic boundaries remain blurry.Therefore, we augment MSE with a structure‑aware loss that preserves local contrast and edge information to encourage each slot to sharpen its semantic boundaries.Lastly, we propose a cyclic inference that rolls slots forward and then backward through the frame sequence, producing temporally consistent object representations even in the earliest frames.All combined, SlotCurri addresses object over-fragmentation by allocating representational capacity where reconstruction fails, further enhanced by structural cues and cyclic inference.Notable FG-ARI gains of +6.8 on YouTube-VIS and +8.3 on MOVi-C validate the effectiveness of SlotCurri.

Video instance segmentation (VIS) for low-light content remains highly challenging for both humans and machines alike, due to adverse imaging conditions including noise, blur and low-contrast. The lack of large-scale annotated datasets and the limitations of current synthetic pipelines, particularly in modeling temporal degradations, further hinder progress. Moreover, existing VIS methods are not robust to the degradations found in low-light videos and, as a result, perform poorly even when finetuned on low-light data. In this paper, we introduce ELVIS (Enhance Low-light for Video Instance Segmentation), a novel framework that enables effective domain adaptation of state-of-the-art VIS models to low-light scenarios. ELVIS comprises an unsupervised synthetic low-light video pipeline that models both spatial and temporal degradations, a calibration-free degradation profile synthesis network (VDP-Net) and an enhancement decoder head that disentangles degradations from content features. ELVIS improves performances by up to +3.7AP on the synthetic low-light YouTube-VIS 2019 dataset. Code will be released upon acceptance.

Continual Generalized Category Discovery (C-GCD) seeks to incrementally discover new categories from unlabeled data and memorize old categories’ knowledge, fostering model adaptability in real-world scenarios. Especially, the unlabeled data is from both old and new classes, requiring the model to recognize previously learned classes while discovering. In response, recent efforts focus on devising specific frameworks and various anti-forgetting strategies, striving for a typical stability-plasticity trade-off. Unlike previous studies, in this work, we first revisit these methods and identify that most of these methods over-protect old classes, hampering the accurate discovery of novel ones. To address this challenge, we introduce the Decouple Your Discovery and Memory (DYDM), a dual-branch architecture that decouples the discovery of new classes and the memorization of old classes. The discovery branch is focused on accurately recognizing new classes, while the memory branch consolidates all identified categories in a recursive manner and functions as the inference branch. Importantly, benefiting from the strong knowledge retention ability of the memory branch, the discovery branch can facilitate the recognition of novel classes from the unlabeled data, achieving a win-win outcome between plasticity and stability. Extensive experiments on various datasets and settings demonstrate the superiority of our approach, achieving leads of up to 9.87%, 7.30%, 3.18%, and 8.25%. Furthermore, our framework can integrate with existing approaches, consistently enhancing their performance.

Generalized Category Discovery (GCD) aims to identify both known and novel classes from unlabeled data with the aid of labeled examples. While promising, most existing GCD methods rely on simultaneous access to labeled and unlabeled datasets—an assumption often impractical in real-world deployments. Continual Category Discovery (CCD) relaxes this requirement by adapting a pre-trained model to streaming unlabeled data, yet it typically assumes domain-consistent data distributions. This places a strong limitation on its applicability. In this work, we study Open Continual Category Discovery (OCCD), where the model must robustly discover previously unseen concepts from real-world data streams that may originate from heterogeneous and shifting domains. To address this, we propose an adaptive framework built on three key ideas. First, we propose a weight-aware separation module, which leverages partial unbalanced optimal transport for instance probability modeling and employs binary response spectrum quantization to generate cues for distinguishing known and unknown categories, enabling automatic sample separation. Second, for known categories, we introduce a cross-domain semantic alignment module that incorporates adversarial learning to perform adaptive prototype matching, thereby enhancing robustness against domain shifts. Finally, for unknown categories, we design a category topology consistency constraint that preserves semantic relationships between known and novel classes during distribution shifts. Experiments show our approach excels at discovering new categories while maintaining strong performance on known ones in evolving domains.

Memory-efficient transfer learning (METL) approaches have recently achieved promising performance in adapting pre-trained models to downstream tasks. They avoid applying gradient backpropagation in large backbones, thus significantly reducing the number of trainable parameters and high memory consumption during fine-tuning. However, since they typically employ a lightweight and learnable side network, these methods inevitably introduce additional memory and time overhead during inference, which contradicts the ultimate goal of efficient transfer learning. To address the above issue, we propose a novel approach dubbed Masked Dual Path Distillation (MDPD) to accelerate inference while retaining parameter and memory efficiency in fine-tuning with fading side networks. Specifically, MDPD develops a framework that enhances the performance by mutually distilling the frozen backbones and learnable side networks in fine-tuning, and discard the side network during inference without sacrificing accuracy. Moreover, we design a novel feature-based knowledge distillation method for the encoder structure with multiple layers. Extensive experiments on distinct backbones across vision/language-only and vision-and-language tasks demonstrate that our method not only accelerates inference by at least 25.2\% while keeping parameter and memory consumption comparable, but also remarkably promotes the accuracy compared to SOTA approaches. The source code will be released upon acceptance.

Existing Incremental Learning (IL) methods are primarily evaluated under either a single-domain class-incremental setting, or a multi-domain task-incremental setting with known task identifiers. However, these assumptions often fail to hold in real-world applications. To bridge this gap, we introduce Heterogeneous Incremental Learning (HIL), a new setting for evaluating IL methods under realistic and challenging conditions, where task boundaries are ambiguous or unknown, class distributions shift dynamically across environments, and training data is limited. Model editing is inherently well-suited for this challenging HIL, as it allows for the efficient integration of new knowledge while preserving model capabilities. Thus, we propose a novel Sparse and Anchored Model Editing (SAME) for addressing HIL. Specifically, SAME sparsely and selectively updates task-relevant model parameters to extract compact, task-specific key–value knowledge pairs from limited data. Using these task knowledge pairs, the model performs knowledge injection for new tasks under double-anchor constraints. The knowledge anchor aligns the updated and original model features, while the parameter anchor constrains parameter magnitudes, ensuring stable and consistent knowledge injection. Our method can efficiently solve HIL using only a few labeled examples, without introducing additional model parameters. Extensive experiments on 11 diverse visual-language datasets across 22 sequential tasks show that our method outperforms existing continual learning approaches by 6.8% in average accuracy, while retaining 95.8% of the oracle model performance, demonstrating strong stability and cross-domain generalization.

To effectively manage the complexities of real-world dynamic environments, continual learning must incrementally acquire, update, and accumulate knowledge from a stream of tasks of different nature—without suffering from catastrophic forgetting of prior knowledge. While this capability is innate to human cognition, it remains a significant challenge for modern deep learning systems.At the heart of this challenge lies the stability-plasticity dilemma: the need to balance leveraging prior knowledge, integrating novel information, and allocating model capacity adaptively based on task complexity and synergy. In this paper, we propose a novel exemplar-free class-incremental continual learning (ExfCCL) framework that addresses these issues through a Hierarchical Exploration-Exploitation (HEE) approach. The core of our method is a HEE-guided efficient neural architecture search (HEE-NAS) that enables a learning-to-adapt backbone via four primitive operations - reuse, new, adapt, and skip—thereby serving as an internal memory that dynamically updates selected components across streaming tasks. To address the task ID inference problem in ExfCCL, we exploit an external memory of task centroids proposed in the prior art. We term our method CHEEM (Continual Hierarchical-Exploration-Exploitation Memory). CHEEM is evaluated on the challenging MTIL and VDD benchmarks using both Tiny and Base Vision Transformers and a proposed holistic Figure-of-Merit (FoM) metric. It significantly outperforms state-of-the-art prompting-based continual learning methods, closely approaching full fine-tuning upper bounds. Furthermore, it learns adaptive model structures tailored to individual tasks in a semantically meaningful way.

State-Space Models (SSMs) excel at capturing long-range dependencies with structured recurrence, making them well-suited for sequence modeling. However, their evolving internal states pose unique challenges in Continual Learning (CL). Without access to the full distribution of previous tasks, updates to the state-space dynamics become unconstrained, leading to catastrophic forgetting. To address this, we propose $\textbf{Inf-SSM}$, a geometry-aware regularization framework for CL in SSMs. It constrains state evolution via the infinite-dimensional Grassmannian of SSM observability subspaces, without requiring any exemplars from past tasks. Unlike classical CL methods that restrict weight updates, Inf-SSM directly regularizes the infinite-horizon state evolution encoded by the extended observability subspace of the SSM. We show that enforcing this regularization requires solving a matrix equation known as the Sylvester equation, which typically incurs $\mathcal{O}(n^3)$ complexity. Thus, we develop a $\mathcal{O}(n^2)$ solution by exploiting the structure and properties of SSMs. This leads to an efficient regularization mechanism that can be seamlessly integrated into existing CL methods. Comprehensive experiments on challenging benchmarks of ImageNet-R, CIFAR-100, and Caltech-256 demonstrate a significant reduction in forgetting while improving accuracy across sequential tasks.

Continual Learning (CL) aims to train neural networks on a dynamic stream of tasks without forgetting previously learned knowledge. Among optimization-based approaches, C-Flat has emerged as a promising solution due to its plug-and-play nature and its ability to encourage uniformly low-loss regions for both new and old tasks. However, C-Flat requires three additional gradient computations per iteration, imposing substantial overhead on the optimization process. In this work, we propose C-Flat Turbo, a faster yet stronger optimizer that significantly reduces the training cost. We show that the gradients associated with first-order flatness contain direction-invariant components relative to the proxy-model gradients, enabling us to skip redundant gradient computations in the perturbed ascent steps. Moreover, we observe that these flatness-promoting gradients progressively stabilize across tasks, which motivates a linear scheduling strategy with an adaptive trigger to allocate larger turbo steps for later tasks. Experiments demonstrate that C-Flat Turbo accelerates a wide range of CL methods by at least $1\times$ (and up to $1.25\times$) compared to C-Flat, while achieving comparable or even improved accuracy.

Functional magnetic resonance imaging (fMRI) is widely used for studying and diagnosing brain disorders, with functional connectivity (FC) matrices providing powerful representations of large-scale neural interactions. However, existing diagnostic models are trained either on a single site or under full multi-site access, making them unsuitable for real-world scenarios where clinical data arrive sequentially from different institutions. This results in limited generalization and severe catastrophic forgetting.This paper presents the first continual learning framework specifically designed for fMRI-based diagnosis across heterogeneous clinical sites. Our framework introduces a structure-aware variational autoencoder that synthesizes realistic FC matrices for both patient and control groups. Built on this generative backbone, we develop a multi-level knowledge distillation strategy that aligns predictions and graph representations between new-site data and replayed samples. To further enhance efficiency, we incorporate a hierarchical contextual bandit scheme for adaptive replay sampling.Experiments on multi-site datasets for major depressive disorder (MDD), schizophrenia (SZ), and autism spectrum disorder (ASD) show that the proposed generative model enhances data augmentation quality, and the overall continual learning framework substantially outperforms existing methods in mitigating catastrophic forgetting.

Face retouching requires removing subtle imperfections while preserving unique facial identity features, in order to enhance overall aesthetic appeal. However, existing methods suffer from a fundamental trade-off. Supervised learning on labeled data is constrained to pixel-level label mimicry, failing to capture complex subjective human aesthetic preferences. Conversely, while online reinforcement learning (RL) excels at preference alignment, its stochastic exploration paradigm conflicts with the high-fidelity demands of face retouching and often introduces noticeable noise artifacts due to accumulated stochastic drift. To address these limitations, we propose BeautyGRPO, a reinforcement learning framework that aligns face retouching with human aesthetic preferences. We construct FRPref-10K, a fine-grained preference dataset covering five key retouching dimensions, and train a specialized reward model capable of evaluating subtle perceptual differences. To reconcile exploration and fidelity, we introduce Dynamic Path Guidance (DPG). DPG stabilizes the stochastic sampling trajectory by dynamically computing an anchor-based ODE path and replanning a guided trajectory at each sampling timestep, effectively correcting stochastic drift while maintaining controlled exploration. Extensive experiments show that BeautyGRPO outperforms both specialized face retouching methods and general image editing models, achieving superior texture quality, more accurate blemish removal, and overall results that better align with human aesthetic preferences. Our code will be made publicly available.

Generating realistic and expressive audio-driven talking avatars remains a central challenge in digital human synthesis. Existing methods often depend on intermediate representations such as pose estimations for natural body motion, which restricts flexibility and adds visual distortions. Moreover, most audio-driven approaches rely on discrete emotion classifiers or text labels to regulate facial expression, reducing complex affective dynamics to coarse categories such as happy, sad, or angry. Such categorical supervision fails to capture the continuous and fine-grained speech dynamics (rhythm, energy, intensity) resulting in limited synchronization and emotionally shallow motion. To overcome these limitations, we present SyncDreamer, a unified Diffusion Transformer framework that generates identity-preserving and emotionally expressive talking avatars from only a single image, speech audio, and text prompt.We propose a visual adapter with Attention Localization Loss to maintain identity fidelity, further incorporating an audio dynamics encoder for rhythm- and emotion-aware motion, and an RL-based Cross-Modal Prompt Enhancer grounding textual cues in visual context for fine-grained motion control. Extensive experiments on portrait and full-body benchmarks demonstrate state-of-the-art performance in realism, synchronization accuracy, and semantic controllability, establishing a scalable foundation for expressive digital avatars in interactive and creative applications.

This paper primarily investigates the task of editing facial expression in an input portrait video based on a driving video, which plays a crucial role in animation and film industries. Most existing research mainly focuses on portrait animation, which aims to animate a static portrait image according to the facial motion from the driving video. As a consequence, it remains challenging for them to disentangle the facial expression from head pose rotation and thus lack the ability to edit facial expression independently. In this paper, we propose PerformRecast, a versatile portrait video expression editing method which is dedicated to recast the performance in existing film and animation. The key insight of our method comes from the characteristics of 3D Morphable Face Model (3DMM), which models the face identity, facial expression and head pose of 3D face mesh with separate parameters. Therefore, we modify the keypoints transformation formula in previous methods to make it more consistent with 3DMM model, which achieves a better disentanglement and provides users with much more fine-grained control. Furthermore, to avoid the misalignment around the boundary of face in generated results, we decouple the facial and non-facial regions of input portrait images and pre-train a teacher model to provide separate supervision for them. Extensive experiments show that our method produces high-quality results which are more faithful to the driving video, outperforming existing methods in both controllability and efficiency. We will release our code and trained models to facilitate future research.

Generating lifelike conversational avatars requires modeling not just isolated speakers, but the dynamic, reciprocal interaction of speaking and listening.However, modeling the listener is exceptionally challenging: direct audio-driven training fails, producing stiff, static listening motions. This failure stems from a fundamental imbalance: the speaker's motion is strongly driven by speech audio, while the listener's motion primarily follows an internal motion prior and is only loosely guided by external speech. This challenge has led most methods to focus on speak-only generation. The only prior attempt at joint generation relies on extra speaker's motion to produce the listener.This design is not end-to-end, thereby hindering the real-time applicability.To address this limitation, we present UniLS, the first end-to-end framework for generating unified speak-listen expressions, driven by only dual-track audio.Our method introduces a novel two-stage training paradigm.Stage 1 first learns the internal motion prior by training an audio-free autoregressive generator, capturing the spontaneous dynamics of natural facial motion. Stage 2 then introduces the dual-track audio, fine-tuning the generator to modulate the learned motion prior based on external speech cues.Extensive evaluations show UniLS achieves state-of-the-art speaking accuracy. More importantly, it delivers up to 44.1% improvement in listening metrics, generating significantly more diverse and natural listening expressions. This effectively mitigates the stiffness problem and provides a practical, high-fidelity audio-driven solution for interactive digital humans.

Recent advancements in audio-driven talking face generation have made great progress in lip synchronization. However, current methods often lack sufficient control over talking face, such as speaking style and emotional expression, resulting in uniform facial motion. In this paper, we focus on improving two key factors: lip-audio alignment control(LAC) and emotion control(EMC), to enhance the diversity and user-friendliness of talking videos. Lip-audio alignment control ensures accurate lip-sync across varied speaking styles to simulate different talking habits, whereas emotion control aims to generate realistic emotional expressions with varying intensities and mixed emotional states. To achieve precise facial animation control, we propose a novel and efficient framework, PC-Talk, which enables lip-audio alignment control and emotion control through implicit keypoint deformations. First, our LAC module generates lip-synced talking faces with a specific speaking style, derived from either a video reference or preset options. It also supports lip movement scale adjustment and fine-grained editing of speaking styles for specific articulations. Second, our EMC module produces vivid emotional facial expressions through pure emotional deformation. It further enables precise control over emotion intensity and the compound emotions across different facial regions. Our method demonstrates outstanding control capabilities and achieves state-of-the-art performance on HDTF and MEAD datasets in experiments. The code will be publicly available.

Current diffusion-based acceleration methods for long-portrait animation struggle to ensure identity (ID) consistency. This paper presents FlashPortrait, an end-to-end video diffusion transformer capable of synthesizing ID-preserving, infinite-length videos while achieving up to 6$\times$ acceleration in inference speed. In particular, FlashPortrait begins by computing the identity-agnostic facial expression features with an off-the-shelf extractor. It then introduces a Normalized Facial Expression Block to align facial features with diffusion latents by normalizing them with their respective means and variances, thereby improving identity stability in facial modeling. During inference, FlashPortrait adopts a dynamic sliding-window scheme with weighted blending in overlapping areas, ensuring smooth transitions and ID consistency in long animations. In each context window, based on the latent variation rate at particular timesteps and the derivative magnitude ratio among diffusion layers, FlashPortrait utilizes higher-order latent derivatives at the current timestep to directly predict latents at future timesteps, thereby skipping several denoising steps and achieving 6$\times$ speed acceleration. Experiments on benchmarks show the effectiveness of FlashPortrait both qualitatively and quantitatively.

Autonomous driving has made substantial progress recently, achieving reliable performance in most real-world environments. However, existing algorithms still depend heavily on high-definition maps, making them ineffective in mapless scenarios such as indoor parking lots. These limitations hinder seamless point-to-point navigation and restrict the broader deployment of the autonomous driving system.To address this challenge, we propose DriveVLN, a new task that extends Vision-and-Language Navigation (VLN) to autonomous driving. DriveVLN employs visual and linguistic priors to guide vehicles toward destinations based solely on concise natural-language descriptions, without access to predefined maps or routes. Unlike conventional VLN, which relies on detailed step-wise instructions in indoor environments, DriveVLN requires models to produce navigation information based on diverse visual cues and history, including signs, landmarks, and textual indicators.We further develop a CARLA-based simulation engine comprising over 200 realistic scenes reconstructed from real road scans, enabling large-scale training and closed-loop evaluation. A baseline model is established through supervised fine-tuning on real data, followed by reinforcement learning in simulation.Comprehensive experiments show that DriveVLN effectively bridges map-based and mapless driving, providing a new foundation for unified, language-driven autonomous navigation in complex real-world environments.

Vision-Language Navigation (VLN) aims to enable agents to navigate to a target location based on language instructions. Traditional VLN often follows a close-set assumption, i.e., training and test data share the same style of the input images and instructions. However, the real world is open and filled with various unseen environments, posing enormous difficulties for close-set methods.To this end, we focus on the General Scene Adaptation (GSA-VLN) task, aiming to learn generalized navigation ability by introducing diverse environments and inconsistent intructions.Towards this task, when facing unseen environments and instructions, the challenge mainly lies in how to enable the agent to dynamically produce generalized strategies during the navigation process. Recent research indicates that by means of fast and slow cognition systems, human beings could generate stable policies, which strengthen their adaptation for open world. Inspired by this idea, we propose the slow4fast-VLN, establishing a dynamic interactive fast-slow reasoning framework.The fast-reasoning module, an end-to-end strategy network, outputs actions via real-time input. It accumulates execution records in a history repository to build memory.The slow-reasoning module analyze the memories generated by the fast-reasoning module. Through deep reflection, it extracts experiences that enhance the generalization ability of decision-making. These experiences are structurally stored and used to continuously optimize the fast-reasoning module. Unlike traditional methods that treat fast-slow reasoning as independent mechanisms, our framework enables fast-slow interaction. By leveraging the experiences from slow reasoning, it continually improves the accuracy and generalization ability of fast decisions. This interaction allows the system to continuously adapt and efficiently execute navigation tasks when facing unseen scenarios. Extensive experiments demonstrate the superiorities of our method.

Vision-Language-Action (VLA) models are emerging as a promising paradigm for end-to-end autonomous driving, valued for their potential to leverage world knowledge and reason about complex driving scenes. However, existing methods suffer from two critical limitations: a persistent misalignment between language instructions and action outputs, and the inherent inefficiency of typical auto-regressive action generation. In this paper, we introduce LinkVLA, a novel architecture that directly addresses these challenges to enhance both alignment and efficiency. First, we establish a structural link by unifying language and action tokens into a shared discrete codebook, processed within a single multi-modal model. This structurally enforces cross-modal consistency from the ground up. Second, to create a deep semantic link, we introduce an auxiliary action understanding objective that trains the model to generate descriptive captions from trajectories, fostering a bidirectional language–action mapping. Finally, we replace the slow, step-by-step generation with a two-step coarse-to-fine generation method (C2F) that efficiently decodes the action sequence, saving 86% inference time. Experiments on closed-loop driving benchmarks show consistent gains in instruction following accuracy and driving performance, alongside reduced inference latency.

Human driving behavior is inherently personal, which is shaped by long-term habits and influenced by short-term intentions. Individuals differ in how they accelerate, brake, merge, yield, and overtake across diverse situations. However, existing end-to-end autonomous driving systems either optimize for generic objectives or rely on fixed driving modes, lacking the ability to adapt to individual preferences or interpret natural language intent. To address this gap, we propose Drive My Way (DMW), a personalized Vision-Language-Action (VLA) driving framework that aligns with users' long-term driving habits and adapts to real-time user instructions. DMW learns a user embedding from our personalized driving dataset collected across multiple real drivers and conditions the policy on this embedding during planning, while natural language instructions provide additional short-term guidance. Closed-loop evaluation on the Bench2Drive benchmark demonstrates that DMW improves style instruction adaptation, and user studies show that its generated behaviors are recognizable as each driver’s own style, highlighting personalization as a key capability for human-centered autonomous driving.

Vision-Language Models (VLMs) have emerged as a promising paradigm in autonomous driving (AD), offering a unified framework for perception, reasoning, and decision-making by jointly modeling visual inputs and natural language instructions. However, their real-world deployment is hindered by the significant computational overhead incurred when processing high-resolution, multi-view images—a standard setup in AD systems that utilize six or even more synchronized cameras to perceive the environment comprehensively. This overhead stems from the large number of visual tokens generated during encoding, which significantly increases inference latency and memory consumption when passed to large language models, owing to the quadratic complexity of self-attention. To address these challenges, we propose Prune2Drive, a plug-and-play visual token pruning framework specifically designed for multi-view VLMs in autonomous driving. Prune2Drive introduces two core innovations: (i) a diversity-aware token selection mechanism inspired by farthest point sampling, which prioritizes semantic and spatial coverage across views rather than relying solely on attention scores; and (ii) a view-adaptive pruning controller that automatically learns optimal pruning ratios for each camera view based on their importance to downstream driving tasks. Unlike prior methods, Prune2Drive does not require model retraining or access to attention maps, making it compatible with modern efficient attention implementations. Extensive experiments on two large-scale multi-view driving benchmarks, DriveLM and DriveLMM-o1, demonstrate that Prune2Drive achieves significant speedups and memory savings while maintaining—and in some cases improving—task performance. Our results establish Prune2Drive as a practical and generalizable solution for efficient vision-language reasoning in autonomous driving. When retaining only 10% of the visual tokens, our method achieves a 6.40$\times$ speedup in the prefilling phase and consumes 13.4% of the original FLOPs, with only a 3% average performance drop compared to the original model on the DriveLM benchmark.

We present a compact pipeline for high-fidelity hair reconstruction from multi-view images. While recent 3D Gaussian Splatting (3DGS) methods achieve realistic results, they often require millions of primitives, leading to high storage and rendering costs. Observing that hair exhibits structural and visual similarities across a hairstyle, we cluster strands into representative hair cards and group these into shared texture codebooks. Our approach integrates this structure with 3DGS rendering, significantly reducing reconstruction time and storage while maintaining comparable visual quality. In addition, we propose a generative prior accelerated method to reconstruct the initial strand geometry from a set of images. Our experiments demonstrate a 4-fold reduction in strand reconstruction time and achieve comparable rendering performance with over 200× lower memory footprint.

We introduce HyperGaussians, a novel extension of 3D Gaussian Splatting for high-quality animatable face avatars. While tremendous successes have been achieved for static faces, animatable avatars from dynamic videos still fall in the uncanny valley. The de facto standard, 3D Gaussian Splatting (3DGS), represents a face through a collection of 3D Gaussian primitives. 3DGS excels at rendering static faces, but the state-of-the-art still struggles with nonlinear deformations, complex lighting effects, and fine details. While most related works focus on predicting better Gaussian parameters from expression codes, we rethink the 3D Gaussian representation itself and how to make it more expressive. Our insights lead to a novel extension of 3D Gaussians to high-dimensional multivariate Gaussians, dubbed 'HyperGaussians'. The higher dimensionality increases expressivity through conditioning on a learnable local embedding. However, splatting HyperGaussians is computationally expensive because it requires inverting a high-dimensional covariance matrix. We solve this by reparameterizing the covariance matrix, dubbed the 'inverse covariance trick'. This trick boosts the efficiency so that HyperGaussians can be seamlessly integrated into existing models. To demonstrate this, we plug in HyperGaussians into two state-of-the-art methods for face avatars: FlashAvatar and GaussianHeadAvatar. Our evaluation on 29 subjects from 6 face datasets shows that HyperGaussians outperform 3DGS numerically and visually, particularly for high-frequency details like eyes, teeth, wrinkles, and specular reflections.

Reconstructing high-fidelity 3D head geometry from images is critical for a wide range of applications, yet existing methods face fundamental limitations. Traditional photogrammetry achieves exceptional detail but requires extensive camera arrays (25-200+ views), substantial computation, and manual cleanup in challenging areas like facial hair. Recent alternatives present a fundamental trade-off: foundation models enable efficient single-image reconstruction but lack fine geometric detail, while optimization-based methods achieve higher fidelity but require dense views and expensive computation. We bridge this gap with a hybrid approach that combines the strengths of both paradigms.Our method introduces a multi-view surface normal prediction model that extends monocular foundation models with cross-view attention to produce geometrically consistent normals in a feed-forward pass. We then leverage these predictions as strong geometric priors within an inverse rendering optimization framework to recover high-frequency surface details. Our approach outperforms state-of-the-art single-image and multi-view methods, achieving high-fidelity reconstruction on par with dense-view photogrammetry while reducing camera requirements and computational cost. The code and model will be released.

3D Gaussian Splatting (3DGS) has become a standard approach to reconstruct and render photorealistic 3D head avatars. A major challenge is to relight the avatars to match any scene illumination. For high quality relighting, existing methods require subjects to be captured under complex time-multiplexed illumination, such as one-light-at-a-time (OLAT). We propose a new generalized relightable 3D Gaussian head model that can relight any subject observed in a single- or multi-view images without requiring OLAT data for that subject. Our core idea is to learn a mapping from flat-lit 3DGS avatars to corresponding relightable Gaussian parameters for that avatar. Our model consists of two stages: a first stage that models flat-lit 3DGS avatars without OLAT lighting, and a second stage that learns the mapping to physically-based reflectance parameters for high-quality relighting. This two-stage design allows us to train the first stage across diverse existing multi-view datasets without OLAT lighting ensuring cross-subject generalization, where we learn a dataset-specific lighting code for self-supervised lighting alignment. Subsequently, the second stage can be trained on a significantly smaller dataset of subjects captured under OLAT illumination. Together, this allows our method to generalize well and relight any subject from the first stage as if we had captured them under OLAT lighting. Furthermore, we can fit our model to unseen subjects from as little as a single image, allowing several applications in novel view synthesis and relighting for digital avatars.

We present MATCH (Multi-view Avatars from Topologically Corresponding Heads), a multi-view Gaussian registration method for high-quality head avatar creation and editing. State-of-the-art multi-view head avatars require time-consuming head tracking, which is followed by an expensive avatar optimization, often resulting in a total creation time that exceeds one day. MATCH instead directly predicts Gaussian splat textures in correspondence from calibrated multi-view images in 0.5 seconds per frame. While the learned intra-subject correspondence across frames allows us to quickly build personalized head avatars, correspondence across subjects enables various applications such as expression transfer, optimization-free tracking, semantic editing, and identity interpolation. We learn to establish such correspondences end-to-end, with a transformer-based model that predicts textures of Gaussian splats in the fixed UV layout of a template mesh. To this end, we introduce a novel registration-guided attention block, in which each UV map token attends exclusively to image tokens depicting its corresponding mesh region. MATCH outperforms existing methods for novel-view synthesis, geometry registration, and head avatar generation, the latter being $10\times$ faster than the qualitatively closest baseline. Code and model weights will be published upon acceptance.

Large Vision-Language Models (LVLMs) can reason effectively from image-text inputs and perform well in various multimodal tasks. Despite this success, they are affected by language priors and often produce hallucinations. Hallucinations denote generated content that is grammatically and syntactically coherent, yet bears no match or direct relevance to actual visual input.To address this problem, we propose Residual Decoding (\textit{ResDec}). It is a novel training-free method that uses historical information to aid decoding. The method relies on the internal implicit reasoning mechanism and token logits evolution mechanism of LVLMs to correct biases. Extensive experiments demonstrate that \textit{ResDec} effectively suppresses hallucinations induced by language priors, significantly improves visual grounding, and reduces object hallucinations. In addition to mitigating hallucinations, \textit{ResDec} also performs exceptionally well on comprehensive LVLM benchmarks, highlighting its broad applicability.

Large Vision-Language Models (LVLMs) have achieved remarkable progress in visual-textual understanding, yet their reliability is critically undermined by hallucinations, i.e., the generation of factually incorrect or inconsistent responses.While recent studies using steering vectors demonstrated promise in reducing hallucinations, a notable challenge remains: they inadvertently amplify the severity of residual hallucinations. We attribute this to their exclusive focus on the decoding stage, where errors accumulate autoregressively and progressively worsen subsequent hallucinatory outputs.To address this, we propose $\textbf{P}$refill-$\textbf{T}$ime $\textbf{I}$ntervention ($\textbf{PTI}$), a novel steering paradigm that intervenes only once during the prefill stage, enhancing the initial Key-Value (KV) cache before error accumulation occurs.Specifically, PTI is modality-aware, deriving distinct directions for visual and textual representations. This intervention is decoupled to steer keys toward visually-grounded objects and values to filter background noise, correcting hallucination-prone representations at their source.Extensive experiments demonstrate PTI's significant performance in mitigating hallucinations and its generalizability across diverse decoding strategies, LVLMs, and benchmarks. Moreover, PTI is orthogonal to existing decoding-stage methods, enabling plug-and-play integration and further boosting performance.

Unlike environmental sounds that mainly indicate event occurrence (e.g., dog barking), human speech carries rich semantics and temporal structures. Despite the advancement of audio-visual large-language models (LLMs) in video understanding, it remains unexplored whether current models can accurately align speech contents with corresponding visual signals. In this work, we show that speech content can induce hallucinations in audio-visual LLMs, where models generate inaccurate or misleading outputs. To systematically study this, we introduce SVHalluc, the first comprehensive benchmark for evaluating speech–vision hallucination in audio-visual LLMs.Our benchmark diagnoses speech–vision hallucinations from two complementary perspectives: semantic and temporal. Experimental results demonstrate that most advanced audio-visual LLMs struggle with aligning speech content with corresponding visual signals. Our work uncovers a fundamental limitation of current audio-visual LLMs and highlights the need for speech-aware and grounded speech-video perception and comprehension. Code will be released upon acceptance.

Large Vision-Language Models (LVLMs) often suffer from object hallucination, generating objects that are absent from the image. Prior work largely attributes this to insufficient visual attention. However, in this work, we are surprised to find that both real and hallucinated objects receive equally strong visual attention in the model’s mid-to-late layers. This indicates that the key issue may not be how much the model attends, but what it attends to and why. To this end, we decode the visual features of high-attention regions using Logit Lens, and observe that high-attention regions corresponding to real objects can be correctly decoded to the target object token, whereas those for hallucinated objects cannot. Building on this, we identify two distinct hallucination mechanisms: (i) visual uncertainty, triggered by semantically similar or confusable regions, masking these regions eliminates the hallucination. (ii) contextual prior, triggered by strong co-occurrence priors, even when the initially attended region is masked, the hallucination persists and attention drifts to other regions. Based on these findings, we propose a simple yet effective training-free Detect–Mitigate framework comprising a Logit-Lens Consistency Check to detect hallucination and targeted remedies: High-Attention Regions Masking (HARM) for visual uncertainty hallucination, and Visual Evidence Enhanced Decoding (VEED) for contextual prior hallucination, which leverages genuine visual evidence to suppress erroneous priors. Our approach achieves state-of-the-art results on multiple hallucination benchmarks.

The recent success of reinforcement learning (RL) in large reasoning models has inspired the growing adoption of RL for post-training Multimodal Large Language Models (MLLMs) to enhance their visual reasoning capabilities. Although many studies have reported improved performance, it remains unclear whether RL training truly enables models to learn from and reason with visual information. In this work, we propose the Hallucination-as-Cue Framework, an analytical framework designed to investigate the effects of RL-based post-training on multimodal reasoning models from the perspective of model hallucination. Specifically, we introduce hallucination-inductive, modality-specific corruptions that remove or replace essential information required to derive correct answers, thereby forcing the model to reason by hallucination. By applying these corruptions during both training and evaluation, our framework provides a unique perspective for diagnosing RL training dynamics and understanding the intrinsic properties of reasoning datasets. Through extensive experiments and analyses across multiple multimodal reasoning benchmarks, we reveal that the role of model hallucination is more significant than previously recognized. For instance, we find that RL post-training under purely hallucination-inductive settings can achieve comparable or even better performance to standard training. These findings challenge prevailing assumptions about MLLM reasoning and motivate the development of more modality-aware RL-based post-training strategies.

We address hallucination detection in Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) by framing the problem through the lens of dynamical systems stability theory. Rather than treating hallucination as a straightforward classification task, we conceptualize (M)LLMs as dynamical systems, where factual knowledge is represented by stable equilibrium points within the representation space. Our main insight is that hallucinations tend to arise at the boundaries of knowledge—transition regions separating stable and unstable zones. To capture this phenomenon, we propose Lyapunov Probes: lightweight networks trained with derivative-based stability constraints that enforce a monotonic decay in confidence under input perturbations. By performing systematic perturbation analysis and applying a two-stage training process, these probes reliably distinguish between stable factual regions and unstable, hallucination-prone areas. Experiments on diverse datasets and models demonstrate consistent improvements over existing baselines.

World engines aim to synthesize long, 3D-consistent videos that support interactive exploration of a scene under user-controlled camera motion. However, existing systems struggle under aggressive 6-DoF trajectories and complex outdoor layouts: they lose long-range geometric coherence, deviate from the target path, or collapse into overly conservative motion. To this end, we introduce Captain Safari, a pose-conditioned world engine that generates videos by retrieving from a persistent world memory. Given a camera path, our method maintains a dynamic local memory and uses a retriever to fetch pose-aligned world tokens, which then condition video generation along the trajectory. This design enables the model to maintain stable 3D structure while accurately executing challenging camera maneuvers.To evaluate this setting, we curate OpenSafari, a new in-the-wild FPV dataset containing high-dynamic drone videos with verified camera trajectories, constructed through a multi-stage geometric and kinematic validation pipeline. Across video quality, 3D consistency, and trajectory following, Captain Safari substantially outperforms state-of-the-art camera-controlled generators. It reduces MEt3R from 0.3703 to 0.3690, improves AUC@30 from 0.181 to 0.200, and yields substantially lower FVD than all camera-controlled baselines. In a 50-participant human study, \textbf{67.6\%} of preferences favor our method across all axes. Our results demonstrate that pose-conditioned world memory is a powerful mechanism for long-horizon, controllable video generation and provide \emph{OpenSafari} as a challenging new benchmark for future world-engine research.

We present Gen3R, a method that bridges the strong priors of foundational reconstruction models and video diffusion models for scene-level 3D generation. We repurpose the VGGT reconstruction model to produce geometric latents by training an adapter on its tokens, which are regularized to align with the appearance latents of pre-trained video diffusion models. By jointly generating these disentangled yet aligned latents, \method{} produces both RGB videos and corresponding 3D geometry, including camera poses, depth maps, and global point clouds. Experiments demonstrate that our approach achieves state-of-the-art results in single- and multi-image conditioned 3D scene generation. Additionally, our method can enhance the robustness of reconstruction by leveraging generative priors, demonstrating the mutual benefit of tightly coupling reconstruction and generative models.

We introduce PerpetualWonder, a hybrid generative simulator that enables long-horizon, action-conditioned 4D scene generation from a single image. Current works fail at this task because their physical state is decoupled from their visual representation, which prevents generative refinements to update the underlying physics for subsequent interactions. PerpetualWonder solves this by introducing the first true closed-loop system. It features a novel unified representation that creates a bidirectional link between the physical state and visual primitives, allowing generative refinements to correct both the dynamics and appearance. It also introduces a robust update mechanism that gathers supervision from multiple viewpoints to resolve optimization ambiguity. Experiments demonstrate that from a single image, PerpetualWonder can successfully simulate complex, multi-step interactions from long-horizon actions, maintaining physical plausibility and visual consistency.

Cinematic video production requires control over scene-subject composition and camera movement, but live-action shooting remains costly due to the need for constructing physical sets. To address this, we introduce the task of cinematic video generation with decoupled scene context: given multiple images of a static environment, the goal is to synthesize high-quality videos featuring dynamic subject while preserving the underlying scene consistency and following a user-specified camera trajectory. We present CineScene, a framework that leverages implicit 3D-aware scene representation for cinematic video generation. Our key innovation is a novel context conditioning mechanism that injects 3D-aware features in an implicit way: By encoding scene images into visual representations through VGGT, CineScene injects spatial priors into a pretrained text-to-video generation model by additional context concatenation, enabling camera-controlled video synthesis with consistent scenes and dynamic subjects. To further enhance the model's robustness, we introduce a simple yet effective random-shuffling strategy for the input scene images during training. To address the lack of training data, we construct a scene-decoupled dataset with Unreal Engine 5, containing paired videos of scenes with and without dynamic subjects, panoramic images representing the underlying static scene, along with their camera trajectories. Experiments show that CineScene achieves state-of-the-art performance in scene-consistent cinematic video generation, handling large camera movements and demonstrating generalization across diverse environments.

Stereo video inpainting, which aims to fill the occluded regions of warped videos with visually coherent content while maintaining temporal consistency, remains a challenging open problem. The regions to be filled are scattered along object boundaries and occupy only a small fraction of each frame, leading to two key challenges. First, existing approaches perform poorly on such tasks due to the scarcity of high-quality stereo inpainting datasets, which limits their ability to learn effective inpainting priors. Second, these methods apply equal processing to all regions of the frame, even though most pixels require no modification, resulting in substantial redundant computation. To address these issues, we introduce three interconnected components. We first propose Gradient-Aware Parallax Warping (GAPW), which leverages backward warping and the gradient of the coordinate mapping function to obtain continuous edges and smooth occlusion regions. Then, a Parallax-Based Dual Projection (PBDP) strategy is introduced, which incorporates GAPW to produce geometrically consistent stereo inpainting pairs and accurate occlusion masks without requiring stereo videos. Finally, we present Sparsity-Aware Stereo Inpainting (SASI), which reduces over 70\% of redundant tokens, achieving a 10.7$\times$ speedup during diffusion inference and delivering results comparable to its full-computation counterpart, enabling real-time processing of HD ($768\times1280$) videos at 25\,FPS on a single A100 GPU.

We identify occlusion reasoning as a fundamental yet overlooked aspect for 3D layout–conditioned generation. It is essential for synthesizing partially occluded objects with depth-consistent geometry and scale. While existing methods can generate realistic scenes that follow input layouts, they often fail to model precise inter-object occlusions. We propose \textbf{SeeThrough3D}, a model for 3D layout conditioned generation that explicitly models occlusions. We introduce an occlusion-aware 3D scene representation (OSCR), where objects are depicted as translucent 3D boxes placed within a virtual environment and rendered from desired camera viewpoint. The transparency encodes hidden object regions, enabling the model to reason about occlusions, while the rendered viewpoint provides explicit camera control during generation. We condition a pretrained flow based text-to-image image generation model by introducing a set of visual tokens derived from our rendered 3D representation. Furthermore, we apply masked self-attention to accurately bind each object bounding box to its corresponding textual description, enabling accurate generation of multiple objects without object attribute mixing. To train the model, we construct a synthetic dataset with diverse multi-object scenes with strong inter-object occlusions. SeeThrough3D generalizes effectively to unseen object categories and enables precise 3D layout control with realistic occlusions and consistent camera control.

High-quality video editing and processing are crucial in domains such as filmmaking and autonomous driving, where accurate visual refinement and data preparation are essential. However, it is challenging to achieve precise control over dynamic objects while maintaining spatiotemporal consistency. Current approaches typically utilize text prompts or 2D structural priors for video editing to ensure consistency, yet they struggle to effectively constrain the spatial variations of dynamic 3D objects. In this paper, we introduce $\textbf{RecEdit-Drive}$, a framework that integrates $\textbf{Spatial Feature Warping}$ and $\textbf{Spatiotemporal Collaborative Modeling}$ to effectively control 3D object variations and enhance video consistency. The spatial feature warping enhances precise control over the edited foreground 3D objects, enhancing spatial consistency in the generated videos; and the spatiotemporal collaborative modeling seamlessly integrates edited foreground objects into the background, yielding realistic and consistent edited videos. Besides, we design an inference strategy to reconstruct an accurate background structure through noise manipulation, providing a reliable reference for foreground instance editing at early denoising stages. We perform extensive qualitative and quantitative comparisons regarding general video editing and downstream tasks on the public datasets, demonstrating the state-of-the-art performance of our proposed method.

World foundation models aim to simulate the evolution of the real world with physically plausible behavior. Unlike prior methods that handle spatial and temporal correlations separately, we propose RayNova, a geometry-free world model that employs a dual-causal autoregressive framework. It follows both scale-wise and temporal topological orders in the autoregressive process, and leverages global attention for unified 4D spatio-temporal reasoning. Different from existing works that impose strong 3D geometric priors, RayNova constructs an isotropic spatio-temporal representation across views, frames, and scales based on relative Plücker-ray positional encoding, enabling robust generalization to diverse camera setups and ego motions. We further introduce a recurrent training paradigm to alleviate distribution drift in long-horizon video generation. RayNova achieves state-of-the-art multi-view video generation results on nuScenes, while offering higher throughput and strong controllability under diverse input conditions, generalizing to novel views and camera configurations without explicit 3D scene representation.

Recent progress in 4D generation has advanced the reconstruction of dynamic geometry, yet the modeling of rig and motion, the two core elements of animation, remains disconnected. Existing approaches typically treat rigging and motion generation as independent tasks: auto-rigging methods rely on human-annotated skeletons and skinning weights, while motion-generation models predict dense vertex trajectories without any explicit structure. This separation contradicts the nature of animation itself, which is the coupled outcome of both structure and motion, and it limits scalability, interpretability, and control.We present RigMo, a unified generative framework that jointly learns rig and motion directly from raw mesh sequences without any rig annotations or human priors. RigMo encodes per-vertex deformations into a compact latent space and decodes a set of implicit Gaussian bones, skinning weights, and time-varying transformations that together define an animatable mesh. This design makes the model animatable by construction: a single latent representation yields both an explicit rig structure and temporally coherent motion parameters. Unlike optimization-based auto-rigging methods that overfit to a specific sequence, RigMo generalizes across object categories and motion styles, offering feed-forward inference for arbitrary deformable objects. Experiments on DeformingThings4D, Objaverse-XL, and diverse human and animal datasets demonstrate that RigMo generates smooth, interpretable, and physically consistent rigs, achieving superior reconstruction and generalization compared to existing 4D generative baselines. RigMo establishes a new paradigm for structure-aware, controllable, and scalable 4D generation.

Given a monocular video, the goal of video re-rendering is to generate views of the scene from a novel camera trajectory. Existing methods face two distinct challenges. Geometrically unconditioned models lack spatial awareness, leading to drift and deformation under viewpoint changes. On the other hand, geometrically-conditioned models depend on estimated depth and explicit reconstruction, making them susceptible to depth inaccuracies and calibration errors.We propose to address these challenges by using the implicit geometric knowledge embedded in the latent space of a large 4D reconstruction model to condition the video generation process. These latents capture scene structure in a continuous space without explicit reconstruction. Therefore, they provide a flexible representation that allows the pretrained diffusion prior to regularize errors more effectively. By jointly conditioning on these latents and source camera poses, we demonstrate that our model achieves state-of-the-art results on the video re-rendering task.

Recent person re-identification (ReID) leverages heterogeneous sensing with multiple modalities and viewpoints to improve robustness across diverse conditions. However, most approaches target predefined scenario pairs (e.g., visible-infrared or aerial-ground) and train separate task-specific models. In contrast, real-world applications require retrieving identities from galleries that cover all scenarios, making such designs inefficient and complex to deploy. To bridge this gap, we introduce Any-Scenario ReID (AS-ReID): given a query from any (modality, viewpoint) scenario, a single model retrieves the same identity from a heterogeneous gallery spanning all scenarios. Progress toward AS-ReID is limited by two factors: (i) the lack of a real-world-aligned benchmark with broad scenario coverage, and (ii) the challenge of learning representations that are cohesive within identities and strongly discriminative across identities under diverse scenarios. To this end, we construct MSAG, a Multispectral Aerial-Ground benchmark with 2,337 identities and 434,620 images captured by RGB, near-infrared, and thermal infrared cameras on both ground and UAV platforms. MSAG spans day-night, multiple seasons, and varied weather conditions, and supports AS-ReID as well as conventional ReID tasks. We further propose the Unified Alignment and Discrimination (UAD) framework. Progressive Center Alignment (ProCA) aggregates multi-view features into modality centers and then aligns them toward identity centers to reduce scenario bias. Global Prototype Discrimination (GPD) contrasts samples against global identity prototypes to enforce large-margin discrimination. Extensive experiments highlight the challenges of MSAG and demonstrate the effectiveness of UAD on AS-ReID. The dataset and code will be released.

Object detection has long been dominated by traditional coordinate regression-based models, such as YOLO, DETR, and Grounding DINO. Although recent efforts have attempted to leverage MLLMs to tackle this task, they face challenges like low recall rate, duplicate predictions, coordinate misalignment, etc. In this work, we bridge this gap and propose \textbf{Rex-Omni}, a 3B-scale MLLM that achieves state-of-the-art object perception performance. On benchmarks like COCO and LVIS, Rex-Omni attains performance comparable to or exceeding regression-based models (e.g., DINO, Grounding DINO) in a zero-shot setting. This is enabled by three key designs: \textbf{1) Task Formulation}: we use special tokens to represent quantized coordinates from 0 to 999, reducing the model's learning difficulty and improving token efficiency for coordinate prediction; \textbf{2) Data Engines}: we construct multiple data engines to generate high-quality grounding, referring, and pointing data, providing semantically rich supervision for training; \textbf{3) Training Pipelines}: we employ a two-stage training process, combining supervised fine-tuning on 22 million data with GRPO-based reinforcement post-training. This RL post-training leverages geometry-aware rewards to effectively bridge the discrete-to-continuous coordinate prediction gap, improve box accuracy, and mitigate undesirable behaviors like duplicate predictions that stem from the teacher-guided nature of the initial SFT stage. Beyond conventional detection, Rex-Omni's inherent language understanding enables versatile capabilities such as object referring, pointing, visual prompting, GUI grounding, spatial referring, OCR and key-pointing, all systematically evaluated on dedicated benchmarks. We believe that Rex-Omni paves the way for more versatile and language-aware visual perception systems.

Gait recognition is a biometric technique that identifies individuals based on their walking patterns, offering advantages in long-range, non-intrusive scenarios. However, real-world scenarios often involve heterogeneous sensing modalities such as LiDAR and RGB cameras, making LiDAR-Camera Cross-modal Gait recognition (LCCGR) a critical yet challenging task due to the substantial modality gap between 2D videos and 3D point cloud sequences. To address this challenge, we propose TCFDNet, a Text-guided Cross-modal Feature Disentanglement Network, which leverages modality-aware textual priors as semantic anchors to guide the learning of disentangled modality-shared representations. Specifically, we construct a Gait Modality Text Dictionary (GMTD) using large language models to generate rich semantic descriptions of gait across modalities and viewpoints. A CLIP-based Multi-grained Feature Encoder then aligns visual and textual features within a unified vision-language space. Furthermore, the Text-guided Feature Disentanglement (TFD) module selects the $top\text{-}k$ matched textual descriptions to reconstruct modality-specific representations and derive modality-shared features via residual decomposition and orthogonality constraints. To mitigate the fragility of the disentangled shared features, we propose a Feature Stability Enhancement (FSE) module, which models spatial and channel-wise correlations to improve feature robustness. In addition, a cross-modal patch exchange strategy is introduced to further improve generalization. Extensive experiments on SUSTech1K and FreeGait datasets demonstrate that TCFDNet achieves new state-of-the-art results and validate the effectiveness of the proposed modules.

Cross-degradation generalization remains a critical challenge for RGB-infrared multimodal object detection, especially when training data covers limited degradation types. This paper presents a distribution alignment framework with a key insight: aligning fused features to the pretrained distribution where the frozen detector performs optimally, rather than adapting to training-specific degradations. By freezing the pretrained detector and training only a lightweight fusion module (15\% of total parameters), our approach leverages complementary infrared information to reduce distribution shift while maintaining computational efficiency. The method achieves state-of-the-art results on three benchmarks (LLVIP, FLIR, DroneVehicle) with 4× faster training. Critically, we demonstrate that aligning to the pretrained distribution substantially outperforms aligning to training degradations when generalizing to unseen scenarios.

Detection Transformer (DETR) has redefined object detection by casting it as a set prediction task within an end-to-end framework. Despite its elegance, DETR and its variants still rely on fixed learnable queries and suffer from severe query utilization imbalance, which limits adaptability and leaves the model capacity underused. We propose PaQ-DETR (Pattern and Quality-Aware DETR), a unified framework that enhances both query adaptivity and supervision balance. It learns a compact set of shared latent patterns capturing global semantics and dynamically generates image-specific queries through content-conditioned weighting. In parallel, a quality-aware one-to-many assignment strategy adaptively selects positive samples based on localization–classification consistency, enriching supervision and promoting balanced query optimization. Experiments on COCO, CityScapes, and other benchmarks show consistent gains of 1.5\%–4.2\% mAP across DETR backbones, including ResNet and Swin-Transformer. Beyond accuracy improvement, our method provides interpretable insights into how dynamic patterns cluster semantically across object categories.

Annotating bounding boxes is costly and limits the scalability of object detection. This challenge is compounded by the need to preserve high accuracy while minimizing manual effort in real-world applications. Prior active learning (AL) methods often depend on model features or modify detector internals and training schedules, increasing integration overhead. Moreover, they rarely jointly exploit the benefits of image-level signals, class-imbalance cues, and instance-level uncertainty for comprehensive selection. We present Portable Active Learning (PAL), a detector-agnostic, easily portable framework that operates solely on inference outputs. PAL combines class-wise instance uncertainty with image-level diversity to guide data selection. At each round, PAL trains lightweight class-specific logistic classifiers to distinguish true from false positives, producing entropy-based uncertainty scores for proposals. Candidate images are then refined using global image entropy, class diversity, and image similarity, yielding batches that are both informative and diverse. PAL requires no changes to model internals or training pipelines, ensuring broad compatibility across detectors. Extensive experiments on COCO, PASCAL VOC, and BDD100K demonstrate that PAL consistently improves label efficiency and detection accuracy compared to existing active learning baselines, making it a practical solution for scalable and cost-effective deployment of object detection in real-world settings.

Modern vision models must capture image-level context without sacrificing local detail while remaining computationally affordable. We revisit this tradeoff and advance a simple principle: decouple the roles of global reasoning and local representation. To operationalize this principle, we introduce ConvNeur, a two-branch architecture in which a lightweight neural memory branch aggregates global context on a compact set of tokens, and a locality-preserving branch extracts fine structure. A learned gate lets global cues modulate local features without entangling their objectives. This separation yields subquadratic scaling with image size, retains inductive priors associated with local processing, and reduces overhead relative to fully global attention. On standard classification, detection, and segmentation benchmarks, ConvNeur matches or surpasses comparable alternatives at similar or lower compute and offers favorable accuracy versus latency trade-offs at similar budgets. These results support the view that efficiency follows global-local decoupling.

Person re-identification (ReID) is critical for public safety, yet the performance of RGB-based methods is limited under challenging lighting and occlusion conditions. In contrast, low-frequency radio frequency (RF) signals, with their superior penetration capability and illumination invariance, provide ideal complementary information. However, a key challenge in fusing these heterogeneous modalities lies in the conventional approach that relies heavily on cross‑modal distribution matching, which often over‑regularizes and weakens the discriminative capacity within each modality. Rather than enforcing direct distribution alignment, canonical correlation analysis (CCA) constructs a shared subspace that maximizes cross‑modal correlation, inherently balancing modality specificity and shared semantics. Inspired by this, we reformulate cross-modal alignment as a correlation maximization problem, avoiding direct constraints on feature distributions and guiding the model to harmonize intra‑modal discriminative learning with cross‑modal alignment. Specifically, VRCLIP first refines CLIP’s visual encoder with illumination‑disentangling objectives, then aligns RGB and RF embeddings in a canonical correlation subspace, and finally employs an RF‑anchored reliability gate for adaptive fusion. To advance the area, we will release {VRR}, the first large‑scale vision–radio ReID dataset with over 650K paired image–radar samples and position annotations for 31 participants. Extensive experiments show state‑of‑the‑art 93.9\% mAP and robust generalization across diverse lighting and occlusion conditions.

Unsupervised Camouflaged Object Detection (UCOD) remains a challenging task due to the high intrinsic similarity between target objects and their surroundings, as well as the reliance on noisy pseudo-labels that hinder fine-grained texture learning. While existing refinement strategies aim to alleviate label noise, they often overlook intrinsic perceptual cues, leading to boundary overflow and structural ambiguity. In contrast, learning without pseudo-label guidance yields coarse features with significant detail loss. To address these issues, we propose a unified UCOD framework that enhances both the reliability of pseudo-labels and the fidelity of features. Our approach introduces the Multi-Cue Native Perception module, which extracts intrinsic visual priors by integrating low-level texture cues with mid-level semantics, enabling precise alignment between masks and native object information. Additionally, Pseudo-Label Evolution Fusion intelligently refines labels through teacher-student interaction and utilizes depthwise separable convolution for efficient semantic denoising. It also incorporates Spectral Tensor Attention Fusion to effectively balance semantic and structural information through compact spectral aggregation across multi-layer attention maps. Finally, Local Pseudo-Label Refinement plays a pivotal role in local detail optimization by leveraging attention diversity to restore fine textures and enhance boundary fidelity. Extensive experiments on multiple UCOD datasets demonstrate that our method achieves state-of-the-art performance, characterized by superior detail perception, robust boundary alignment, and strong generalization under complex camouflage scenarios.

Domain adaptive object detection (DAOD) aims to generalize an object detector trained on a source domain to a target domain, where the domain gap degrades the adaptability. Recently, large-scale vision foundation models (VFMs), pretrained on web-scale datasets, exhibit such powerful generalization capabilities that many approaches leverage them to bridge the domain gap. However, their generalized knowledge is not tailored to the specific domain, which makes it difficult to offer precise guidance in the target domain. In this paper, we propose an Expert-Teacher-Student collaborative learning (ETS) framework to synergize the generalized knowledge from VFMs with the domain-specific knowledge from the teacher model. Concretely, we first design an Expert-Teacher Collaborative Teaching (ETCT) module, which leverages the complementary knowledge of expert and teacher models to collaboratively generate high-quality pseudo labels for supervising student model learning. Second, we devise an Expert-Teacher Joint Consolidating (ETJC) module, which introduces class-wise prototype alignment among expert, teacher, and student models, to jointly consolidate generalized and domain-specific knowledge within the student model. ETS leverages VFMs as the expert model in a free lunch manner, thus avoiding significant additional training costs. Extensive experiments exhibit that our method outperforms the existing SOTA methods on three benchmarks.Our code is available in the supplementary materials.

Text-to-video (T2V) generation has recently attracted considerable attention, resulting in the development of numerous high-quality datasets that have propelled progress in this area. However, existing public datasets are primarily composed of isolated text–video (T–V) pairs and thus fail to model inter-clip relationships. To address this limitation, we introduce CI-VID, a dataset that moves beyond isolated T2V generation toward text-and-video-to-video (T&V2V) generation. CI-VID contains over 340,000 samples, each comprising a semantically coherent video sequence with interleaved text captions that capture both clip-level content and inter-clip relationships. To validate its effectiveness, we design a comprehensive, multi-dimensional benchmark incorporating human evaluation, VLM-based assessment, and similarity-based metrics. Experimental results demonstrate that models trained on CI-VID significantly improve both accuracy and content consistency in multi-clip video generation. This enables the creation of story-driven content with smooth transitions and strong semantic coherence, underscoring the value of CI-VID as a foundation for advancing controllable and coherent video generation.

Video quality assessment (VQA) seeks to predict the perceptual quality of a video in alignment with human visual perception, serving as a fundamental tool for quantifying quality degradation across video processing workflows. The dominant VQA paradigm relies on supervised training with human-labeled datasets, which, despite substantial progress, still suffers from poor generalization to unseen video content. In this work, we explore \textbf{weak-to-strong (W2S) learning} as a new paradigm for advancing VQA without reliance on human-labeled datasets. We first provide empirical evidence that a straightforward W2S strategy allows a strong student model to not only match its weak teacher on in-domain benchmarks but also surpass it on out-of-distribution (OOD) benchmarks, revealing a \textbf{distinct weak-to-strong effect in VQA}. Building on this insight, we propose a novel framework that enhances W2S learning from two aspects: (1) \textbf{integrating homogeneous and heterogeneous supervision signals} from diverse VQA teachers---including off-the-shelf VQA models and synthetic distortion simulators---via a learn-to-rank formulation, and (2) \textbf{iterative W2S training}, where each strong student is recycled as the teacher in subsequent cycles, progressively focusing on challenging cases. Extensive experiments show that our method achieves state-of-the-art results across both in-domain and OOD benchmarks, with especially strong gains in OOD scenarios. Our findings highlight W2S learning as a principled route to break annotation barriers and achieve scalable generalization in video quality assessment.

Multimodal Large Language Models (MLLMs) have recently achieved remarkable progress in visual-language understanding. Yet, human perception is inherently multisensory, integrating sight, sound, and motion to reason about the world. Among these modalities, sound provides indispensable cues about spatial layout, off-screen events, and causal interactions, particularly in egocentric settings where auditory and visual signals are tightly coupled. To this end, we introduce EgoSound, the first benchmark designed to systematically evaluate egocentric sound understanding in MLLMs. EgoSound unifies data from Ego4D and EgoBlind, encompassing both sighted and sound-dependent experiences. It defines a seven-task taxonomy spanning intrinsic sound perception, spatial localization, causal inference, and cross-modal reasoning. Constructed through a multi-stage auto-generative pipeline, EgoSound contains 7315 validated QA pairs across 900 videos. Comprehensive experiments on nine state-of-the-art MLLMs reveal that current models exhibit emerging auditory reasoning abilities but remain limited in fine-grained spatial and causal understanding. EgoSound establishes a challenging foundation for advancing multisensory egocentric intelligence, bridging the gap between seeing and truly hearing the world.

Recent advances in video large language models have demonstrated strong capabilities in understanding short clips. However, scaling them to hours- or days-long videos remains highly challenging due to limited context capacity and the loss of critical visual details during abstraction. Existing memory-augmented methods mitigate this by leveraging textual summaries of video segments, yet they heavily rely on text and fail to utilize visual evidence when reasoning over complex scenes. Moreover, retrieving from fixed temporal scales further limits their flexibility in capturing events that span variable durations. To address this, we introduce WorldMM, a novel multimodal memory agent that constructs and retrieves from multiple complementary memories, encompassing both textual and visual representations. WorldMM comprises three types of memory: episodic memory indexes factual events across multiple temporal scales, semantic memory continuously updates high-level conceptual knowledge, and visual memory preserves detailed information about scenes. During inference, an adaptive retrieval agent iteratively selects the most relevant memory source and leverages multiple temporal granularities based on the query, continuing until it determines that sufficient information has been gathered. WorldMM significantly outperforms existing baselines across five long video question-answering benchmarks, achieving an average 8.4% performance gain over previous state-of-the-art methods, showing its effectiveness on long video reasoning.

Video Large Language Models (VLMs) have achieved remarkable success in video understanding, but the significant computational cost from processing dense frames severely limits their practical application. Existing methods alleviate this by selecting keyframes, but their greedy decision-making, combined with a decoupled evaluation of relevance and diversity, often falls into local optima and results in erroneously selecting irrelevant noise frames. To address these challenges, we propose GIFT: Global Irreplaceability Frame Targeting, a novel training-free framework that selects frames by assessing their intrinsic irreplaceability. Specifically, we first introduce Directed Diversity to quantify a frame's uniqueness conditioned on relevance, which allows us to formulate a unified irreplaceability score. Subsequently, our Budget-Aware Refinement strategy employs a adaptive iterative process that first secures a core set of frames with the highest irreplaceability, and then shifts its priority to building crucial temporal context around these selections as the budget expands. Extensive experiments demonstrate that GIFT achieves a maximum average improvement of 12.5\% across long-form video benchmarks on LLaVA-Video-7B compared to uniform sampling. Code will be released soon.

Long-form video reasoning remains a major challenge for Video Large Language Models (Video LLMs), as static uniform frame sampling leads to information dilution and obscures critical evidence. Furthermore, existing pixel-space video reasoning agents, which are designed to actively interact with the video to acquire new visual information, remain suboptimal due to their lack of rigorous reward mechanisms to enforce evidence purity and their inability to perform temporal information supplementation beyond pre-sampled frames. To address this critical gap, we propose a novel evidence-prioritized adaptive framework built upon our core philosophy: “Select Less, Reason More.” Our core contribution is the evidence-aware reinforcement learning (EARL) framework, which transforms the model into an active interrogator of evidence. EARL is precisely engineered to dynamically select the most relevant frames and, crucially, to perform localized re-sampling around the selected key frames to access fine-grained temporal detail. Extensive experiments on five demanding video reasoning benchmarks demonstrate that our EARL-trained model achieves new state-of-the-art among open-source Video LLMs, simultaneously learning an effective and high-purity visual evidence selection policy. Impressively, our 7B model achieves 59.8\% on LongVideoBench, 69.0\% on MVBench and 64.9\% on VideoMME. These results highlight the importance of prioritizing evidence purity and the effectiveness of our framework.

Multimodal AI agents are increasingly automating complex real-world workflows that involve online web execution. However, current web-agent benchmarks suffer from a critical limitation: they focus entirely on web-based interaction and perception, lacking grounding in the user's real-world physical surroundings. This limitation prevents evaluation in crucial scenarios, such as when an agent must use egocentric visual perception (e.g., via AR glasses) to recognize an object in the user's surroundingsand then complete a related task online (e.g., making a purchase).To address this gap, we introduce Ego2Web, the first benchmark designed to bridge egocentric video perception and multimodal web agent execution. Ego2Web pairs real-world first-person video recordings with web tasks that require visual understanding, web task planning, and interaction in an online environment for successful completion. We utilize an automatic data-generation pipeline combined with human verification to curate well-constructed, high-quality video-task pairs across diverse web task types, including e-commercial, navigation, media search, and so on. To facilitate a more accurate and scalable evaluation for our novel benchmark, we also develop a novel LLM-as-a-Judge automatic evaluation method Ego2WebJudge, and demonstrate around 85\% agreement with human judgment, substantially higher than existing evaluation methods.Experiments with diverse SoTA multimodal agents show that they perform significantly below the human level, revealing a major gap in capability. We also conduct a comprehensive ablation study on task design, highlighting the necessity of video perception in the proposed task and the limitations of current agents.We hope Ego2Web can be a critical new resource for developing truly capable AI assistants that can seamlessly see, understand, and act across the physical and digital worlds.

Composed Video Retrieval (CoVR) aims to retrieve a target video given a reference video and a textual modification describing the desired change. The core challenge lies in modeling compositional multimodal transformations, i.e., how objects, actions, and scenes evolve across video and language modalities in response to fine‑grained textual edits. Existing methods address this issue by training on large‑scale video–text–video triplets or by generating dense textual descriptions to capture subtle visual differences. However, these supervised approaches often rely on noisy web‑scale data and dataset‑specific correspondences, leading to overfitting and limited generalization in diverse or fine‑grained scenarios, while also failing to effectively model compositional and temporal transformations. To overcome these limitations, we propose a zero‑shot, fine‑grained transformation reasoning framework based on Multimodal Large Language Models (MLLMs). Our method decomposes the compositional transformation into three complementary reasoning dimensions, i.e., \emph{entity}, \emph{action}, and \emph{scene}, and performs pairwise candidate reasoning to explicitly capture semantic evolution over time. Furthermore, we introduce a recall‑oriented multi‑objective candidate selection module that identifies high‑quality retrieval targets by jointly balancing visual, textual, and multimodal similarities before transformation reasoning. Experiments on EgoCVR and WebVid‑CoVR demonstrate the effectiveness of our method over state‑of‑the‑art approaches under the zero‑shot setting, with R@1 improvements of +5.8 and +10.8, respectively.

Video foundation models aim to integrate video understanding, generation, editing, and instruction following within a single framework, making them a central direction for next-generation multimodal systems. However, existing evaluation benchmarks remain fragmented and limited in scope, as they each target a single task, rely on task-specific metrics, and typically use short or simple video clips. As a result, they do not capture the unified capabilities that these models are designed to deliver. To address this gap, we introduce UniVBench, a benchmark purpose-built for evaluating video foundation models across four core abilities: video understanding, video generation, video editing, and a newly proposed task, video reconstruction, which assesses how faithfully a model can reproduce video content it has encountered. Our benchmark substantially expands the complexity of evaluation by incorporating 200 high-quality, diverse and multi-shot videos, each paired with detailed captions, multi-format editing instructions, and reference images. All videos are entirely human-created and carefully validated, offering richer cinematic information than prior benchmarks. In addition, we develop a unified agentic evaluation system (UniV-Eval) that standardizes prompting, instruction parsing, and scoring across all tasks, enabling fair, scalable, and reproducible comparisons of unified video models. By grounding evaluation in instruction-based multi-shot video tasks, UniVBench provides the first framework for measuring the integrated capabilities that video foundation models aim to achieve. Extensive human annotations ensure our evaluation aligns with human judgment, enabling rigorous assessment and accelerating progress toward robust video intelligence.

Flow-based Transformer models have achieved state-of-the-art image generation performance, but often suffer from high inference latency and computational cost due to their large parameter sizes. To improve inference efficiency without compromising quality, we propose Bridged Progressive Rectified Flow Transformers (NAMI), which decompose the generation process across temporal, spatial, and architectural demensions. We divide the rectified flow into different stages according to resolution, and use a BridgeFlow module to connect them. Fewer Transformer layers are used at low-resolution stages to generate image layouts and concept contours, and more layers are progressively added as the resolution increases. Experiments demonstrate that our approach achieves fast convergence and reduces inference time while ensuring generation quality. The main contributions of this paper are summarized as follows: (1) We introduce Bridged Progressive Rectified Flow Transformers that enable multi-resolution training, accelerating model convergence; (2) NAMI leverages piecewise flow and spatial cascading of Diffusion Transformer (DiT) to rapidly generate images, reducing inference time by 64% for generating 1024×1024 resolution images; (3) We propose a BridgeFlow module to align flows between different stages; (4) We propose the NAMI-1K benchmark to evaluate human preference performance, aiming to mitigate distributional bias and comprehensively assess model effectiveness. The results show that our model is competitive with state-of-the-art models.

Recent diffusion-based models achieve photorealism in image inpainting but require many sampling steps, limiting practical use. Few-step text-to-image models offer faster generation, but naively applying them to inpainting yields poor harmonization and artifacts between the background and inpainted region. We trace this cause to random Gaussian noise initialization, which under low function evaluations causes semantic misalignment and reduced fidelity. To overcome this, we propose InverFill, a one-step inversion method tailored for inpainting that injects semantic information from the target masked image into the initial noise, enabling high-fidelity few-step inpainting. Instead of training inpainting models, InverFill leverages few-step text-to-image models in a blended sampling pipeline with semantically aligned noise as inputs, significantly improving vanilla blended sampling and even matching specialized inpainting models at low NFEs. Moreover, InverFill requires no real-image supervision and only adds minimal inference overhead. Extensive experiments show that InverFill consistently boosts baseline few-step models, improving image quality and text coherence without costly retraining or heavy iterative optimization.

The recent surge in video generation has shown the growing demand for high-quality video synthesis using large vision models. Existing video generation models are predominantly based on the video diffusion transformer (vDiT), however, they suffer from substantial inference delay due to self-attention. While prior studies have focused on reducing redundant computations in self-attention, they often overlook the inherent spatio-temporal correlations in video streams and directly leverage sparsity patterns from large language models to reduce attention computations.In this work, we take a principled approach to accelerate self-attention in vDiTs by leveraging the spatio-temporal correlations in the latent space. We show that the attention patterns within vDiT are primarily due to the dominant spatial and temporal correlations at the token channel level. Based on this insight, we propose a lightweight and adaptive reuse strategy that approximates attention computations by reusing partial attention scores of spatially or temporally correlated tokens along individual channels. We demonstrate that our method achieves significantly higher computational savings (85\%) compared to state-of-the-art techniques over 4 vDiTs, while preserving almost identical video quality ($<$0.06\% loss on VBench).

Procedural sequence generation aims to create intermediate images through multi-step processes, which is applied in industrial design, educational tutorial, and creative content inspiration. However, existing methods often focus on a specific domain or initialize several expert networks for different domains, which face three challenges. First, the poor generalization to unseen domains. Second, the parameter redundancy due to multiple expert networks.Third, the difficulty in adaptively determining the number of generation steps for different processes.To address these challenges, we propose ProcessMaker, a novel framework that harnesses the inherent generalization capabilities in Diffusion Transformers (DiTs) for procedural sequence generation. Concretely, we introduce three key innovations: (1) Self-supervised Representation Alignment to explore the generalized ability for unseen processes. (2) Sparse Masks for different domains without additional expert networks. (3) A sliding window strategy, which dynamically accommodates the generation steps based on the process complexity. Extensive experiments validate that our ProcessMaker achieves procedural sequence generation with generalization ability and adaptive steps, while using only 7.3% trainable parameters compared with the state-of-the-art method.

MeanFlow (MF) is a diffusion-motivated generative model that enables efficient few-step generation by learning long jumps directly from noise to data. In practice, it is often used as a latent MF by leveraging the pre-trained Stable Diffusion variational autoencoder (SD-VAE) for high-dimensional data modeling. However, MF training remains computationally demanding and is often unstable. During inference, the SD-VAE decoder dominates the generation cost, and MF depends on complex guidance hyperparameters for class-conditional generation. In this work, we develop an efficient training and sampling scheme for MF in the latent space of a Representation Autoencoder (RAE), where a pre-trained vision encoder (e.g., DINO) provides semantically rich latents paired with a lightweight decoder. We observe that naive MF training in the RAE latent space suffers from severe gradient explosion. To stabilize and accelerate training, we adopt Consistency Mid-Training for trajectory-aware initialization and use a two-stage scheme: distillation from a pre-trained flow matching teacher to speed convergence and reduce variance, followed by an optional bootstrapping stage with a one-point velocity estimator to further reduce deviation from the oracle mean flow. This design removes the need for guidance, simplifies training configurations, and reduces computation in both training and sampling. Empirically, our method achieves a 1-step FID of 2.03, outperforming vanilla MF’s 3.43, while reducing sampling GFLOPS by 38% and total training cost by 83% on ImageNet 256. We further scale our approach to ImageNet 512, achieving a competitive one-step FID of 3.23 with the lowest GFLOPS among all baselines.Code and proofs are available in the supplementary material.

Diffusion-based image compression has recently shown outstanding perceptual fidelity, yet its practicality is hindered by prohibitive sampling overhead and high memory usage.Most existing diffusion codecs employ UNet architectures, where hierarchical downsampling forces diffusion to operate in shallow latent spaces (typically with only $8\times$ spatial downscaling), resulting in excessive computation.In contrast, conventional VAE-based codecs work in much deeper latent domains ($16\times$–$64\times$ downscaled), motivating a key question:Can diffusion operate effectively in such compact latent spaces without compromising reconstruction quality?To address this, we introduce DiT-IC—an Aligned Diffusion Transformer for Image Compression—which replaces the UNet with a Diffusion Transformer capable of performing diffusion in latent space entirely at $32\times$ downscaled resolution.DiT-IC adapts a pretrained text-to-image multi-step DiT into a single-step reconstruction model through three key alignment mechanisms:(1) a variance-guided reconstruction flow that adapts denoising strength to latent uncertainty for efficient reconstruction;(2) a self-distillation alignment that enforces consistency with encoder-defined latent geometry to enable one-step diffusion; and(3) a latent-conditioned guidance that replaces text prompts with semantically aligned latent conditions, enabling text-free inference.With these designs, DiT-IC achieves state-of-the-art perceptual quality while offering up to 30× faster decoding and drastically lower memory usage than existing diffusion-based codecs. Remarkably, it can reconstruct $2048\times2048$ images on a 16 GB laptop GPU. Code will be released.

Directly modeling the explicit likelihood of the raw data distribution is a key topic in the machine learning area, which achieves the scaling success in Large Language Models by autoregressive modeling. However, continuous AR modeling over visual pixel data suffers from extremely long sequences and high-dimensional spaces. In this paper, we present FARMER, a novel end-to-end generative framework that unifies Normalizing Flows (NFs) and Autoregressive (AR) models for tractable likelihood estimation and high-quality image synthesis directly from raw pixels. FARMER employs an invertible autoregressive flow to transform images into latent sequences, whose distribution is implicitly modeled by an autoregressive model. To address the redundancy and complexity in pixel-level modeling, we propose a self-supervised dimension reduction scheme that partitions NF latent channels into informative and redundant groups, enabling more effective and efficient AR modeling. Furthermore, we design a one-step distillation scheme to significantly accelerate inference speed and introduce a resampling-based classifier-free guidance algorithm to boost image generation quality. Extensive experiments demonstrate that FARMER achieves competitive performance compared to existing pixel-based generative models while providing exact likelihoods and scalable training.

Accurate weather forecasts are essential across various domains and are safety-critical in extreme weather conditions.Compared to simulation-based forecasting, data-driven approaches show greater efficiency, enabling short-term, high-resolution nowcasting.In particular, diffusion models proved effective in weather nowcasting due to their strong probabilistic foundation.However, existing methods rely on deterministic compression to reduce the complexity of high-dimensional weather data, limiting their ability to capture uncertainty in the decoding process.In this work, we introduce FREUD, a Frame-wise Encoder and United Decoder model based on rectified flow transformers for efficient compression of spatio-temporal weather data. Frame-wise encoding enables continuous forecast updates, while the unified video decoder ensures temporal consistency. Our uncertainty-preserving first stage allows us to capture aleatoric uncertainty through ensembling, which is particularly beneficial for extreme weather events with high decoding variability. We achieve state-of-the-art performance in precipitation nowcasting with a compact latent-space rectified flow transformer on the SEVIR benchmark and show further performance gains by scaling model size. With FREUD and the latent rectified flow model, we aim to push the boundaries of data-driven weather nowcasting.

With the surge of pre-trained text-to-image flow matching models, text-based image editing performance has gained remarkable improvement, especially for simple editing that only contains a single editing target. However, to satisfy the exploding editing requirements, the complex editing that contains multiple editing targets is posed as a more challenging task. However, current complex editing solutions: single-round and multi-round editing are limited by long text following and cumulative inconsistency, respectively. Thus, they struggle to strike a balance between semantic alignment and source consistency.In this paper, we propose FlowDC, which decouples the complex editing into multiple sub-editing effects and superposes them in parallel during the editing process. Meanwhile, we observed that the velocity quantity that is orthogonal to the editing displacement harms the source structure preserving. Thus, we decompose the velocity and decay the orthogonal part for better source consistency.To evaluate the effectiveness of complex editing settings, we construct a complex editing benchmark: Complex-PIE-Bench. On two benchmarks, FlowDC shows superior results compared with existing methods. We also detail the ablations of our module designs.

Face swapping aims to seamlessly transfer a source facial identity onto a target while preserving target attributes such as pose and expression. Diffusion models, known for their superior generative capabilities, have recently shown promise in advancing face-swapping quality. This paper addresses two key challenges in diffusion-based face swapping: the prioritized preservation of identity over target attributes and the inherent conflict between identity and attribute conditioning. To tackle these issues, we introduce an identity-constrained attribute-tuning framework for face swapping that first ensures identity preservation and then fine-tunes for attribute alignment, achieved through a decoupled condition injection. We further enhance fidelity by incorporating identity and adversarial losses in a post-training refinement stage. Our proposed identity-constrained diffusion-based face-swapping model outperforms existing methods in both qualitative and quantitative evaluations, demonstrating superior identity similarity and attribute consistency, achieving a new state-of-the-art performance in high-fidelity face swapping.

Current approaches to zero-shot 3D-object perception typically rely on ensembles of frozen foundation models.This limits deep object understanding and cross-domain generalization, making performance inadequate for real-world deployment. The 3D-Object Perception Transformer (3PT) addresses this limitation by unifying detection, segmentation, and 6DoF pose estimation in a single framework, directly trained for 3D-object perception. Based on two large-scale trained Transformers that specialize in 2D and 3D object-centric scene understanding respectively, 3PT continuously refines its object representations without depth input, enhancing 3D understanding by incorporating multi-view information. 3PT surpasses task-specialized models for detection and pose estimation, often achieving double-digit percentage improvements on the diverse BOP-benchmarks. Achieving high accuracy and robustness, \algshort{} is well-suited for practical industrial robotics applications such as bin picking and precise insertion.

Camera-only 3D object detection has emerged as a cost-effective and scalable alternative to LiDAR for autonomous driving, yet existing methods primarily prioritize overall performance while overlooking the severe long-tail imbalance inherent in real-world datasets. In practice, many rare but safety-critical categories such as children, strollers, or emergency vehicles are heavily underrepresented, leading to biased learning and degraded performance. This challenge is further exacerbated by pronounced inter-class ambiguity (e.g., visually similar subclasses) and substantial intra-class diversity (e.g., objects varying widely in appearance, scale, pose, or context), which together hinder reliable long-tail recognition. In this work, we introduce SemLT3D, a Semantic-Guided Expert Distillation framework designed to enrich the representation space for underrepresented classes through semantic priors. SemLT3D consists of: (1) a language-guided mixture-of-experts module that routes 3D queries to specialized experts according to their semantic affinity, enabling the model to better disentangle confusing classes and specialize on tail distributions; and (2) a semantic projection distillation pipeline that aligns 3D queries with CLIP-informed 2D semantics, producing more coherent and discriminative features across diverse visual manifestations. Although motivated by long-tail imbalance, the semantically structured learning in SemLT3D also improves robustness under broader appearance variations and challenging corner cases, offering a principled step toward more reliable camera-only 3D perception.

Bird’s-Eye-View (BEV) detection has become a dominant paradigm for 3D object detection in autonomous driving, due to its strong perception capability. However, most existing methods mainly focus on constructing high-quality BEV feature representations, while neglecting the design of task-specific detection heads. In practice, they directly adopt the center-based head originally developed for 2D detection, without any specific optimization. This leads to three inherent limitations: (i) a geometric mismatch between the Gaussian kernel used for classification and the real BEV object, (ii) degraded end-to-end performance without Non-Maximum Suppression(NMS), and (iii) sparse supervisory signals. To address these issues, we propose Spe-BEVHead, a detection head specifically tailored for BEV 3D object detection. Spe-BEVHead introduces three BEV-specific adaptations: (1) a Rotated Box Kernel that generates geometry-aligned classification weights, (2) a Local Response Refinement Module (LRRM) that suppresses non-peak responses and improves end-to-end performance, and (3) a dual-branch architecture that provides richer supervisory signals to promote more robust learning while inherently preserving the performance for end-to-end inference. Extensive experiments show that Spe-BEVHead can be seamlessly integrated into existing BEV backbones, delivering direct performance gains while retaining competitive performance under the challenging end-to-end setting.

The LiDAR sensor based multi-agent and single-agent perception has shown promising performance in the environmental understanding for robots and automated vehicles. However, there is no existing method that simultaneously solves both multi-agent and single-agent perception in an unsupervised way. By sharing sensor data between multiple agents via communication, this paper discovers two key insights: 1) Improved point cloud density after the data sharing from cooperative views could benefit unsupervised object classification, 2) Cooperative view of multiple agents can be used as unsupervised guidance to the 3D object detection in the single view. Based on these two discovered insights, we propose an Unsupervised Multi-agent and Single-agent (UMS) perception framework that leverages multi-agent cooperation without human annotations to simultaneously solve multi-agent and single-agent perception. UMS combines a learning-based Proposal Purifying Filter to better classify the candidate proposals after multi-agent point cloud density cooperation, followed by a Progressive Proposal Stabilizing module to yield reliable pseudo labels by the easy-to-hard curriculum learning. Furthermore, we design a Cross-View Consensus Learning to use multi-agent cooperative view to guide detection in single-agent view. Experimental results on two public datasets V2V4Real and OPV2V show that our UMS method achieved significantly higher 3D detection performance than the state-of-the-art methods on both multi-agent and single-agent perception tasks in an unsupervised way.

3D object detection is fundamental for spatial understanding. Real-world environments demand models capable of recognizing diverse, previously unseen objects, which remains a major limitation of closed-set methods. Existing open-vocabulary 3D detectors relax annotation requirements but still depend on training scenes, either as point clouds or images. We take this a step further by introducing $Zoo3D$, the first training-free 3D object detection framework. Our method constructs 3D bounding boxes via graph clustering of 2D instance masks, then assigns semantic labels using a novel open-vocabulary module with best-view selection and view-consensus mask generation. $Zoo3D$ operates in two modes: the zero-shot $Zoo3D_{0}$, which requires no training at all, and the self-supervised $Zoo3D_{1}$, which refines 3D box prediction by training a class-agnostic detector on $Zoo3D_{0}$-generated pseudo labels. Furthermore, we extend $Zoo3D$ beyond point clouds to work directly with posed and even unposed images. Across ScanNet200 and ARKitScenes benchmarks, both $Zoo3D_{0}$ and $Zoo3D_{1}$ achieve state-of-the-art results in open-vocabulary 3D detection. Remarkably, our zero-shot $Zoo3D_{0}$ outperforms all existing self-supervised methods, hence demonstrating the power and adaptability of training-free, off-the-shelf approaches for real-world 3D understanding.

Depth priors provide salient geometric structure that benefits camouflaged object detection (COD), but directly using Monocular Depth Estimation (MDE) causes a task misalignment that still fails to identify camouflaged objects.To address this issue, we propose the Depth Segment Anything Model (DepthSAM), a MDE-adapted method specifically designed to mitigate this misalignment.DepthSAM incorporates two core innovations: (1) a Sparse Mixture-of-Experts Adapter (SMEA) that enables DEM to learn semantic information unique to camouflaged scenes, and (2) a Geometric–Semantic Fusion Module (GSFM) that efficiently integrates geometric cues with high-level semantics. With these components, DepthSAM achieves both robust semantic understanding in camouflaged environments and accurate segmentation of camouflaged objects.Extensive experiments show that DepthSAM achieves new SOTA performance on three major benchmarks. For example, on COD10K, its $S_{\alpha}$ and $F_{\beta}^{\omega}$ metrics surpass the best competing methods by 3.0\% and 4.3\%, respectively.

Adversarial robustness of BEV 3D object detectors is critical for autonomous driving (AD). Existing invasive attacks require altering the target vehicle itself (e.g. attaching patches), making them unrealistic and impractical for real-world evaluation. While non-invasive attacks that place adversarial objects in the environment are more practical, current methods still lack the multi-view and temporal consistency needed for physically plausible threats. In this paper, we present the first framework for generating universal, non-invasive, and 3D consistent adversarial objects that expose fundamental vulnerabilities for BEV 3D object detectors. Instead of modifying target vehicles, our method inserts rendered objects into scenes with an occlusion-aware module that enforces physical plausibility across views and time. To maintain attack effectiveness across views and frames, we optimize adversarial object appearance using a BEV spatial feature-guided optimization strategy that attacks the detector's internal representations. Extensive experiments demonstrate that our learned universal adversarial objects can consistently degrade multiple BEV detectors from various viewpoints and distances.More importantly, the new environment-manipulation attack paradigm exposes models' over-reliance on contextual cues and provides a practical pipeline for robustness evaluation in AD systems.

Color affects how we interpret image style and emotion. Previous color grading methods rely on patch-wise recoloring or fixed filter banks, struggling to generalize across creative intents or align with human aesthetic preferences. In this study, we propose **AceTone**, the first approach that supports multimodal conditioned color grading within a unified framework. AceTone formulates grading as a generative color transformation task, where a model directly produces 3D-LUTs conditioned on text prompts or reference images. We develop a VQ-VAE-based tokenizer which compresses a $3\times32^3$ LUT vector to 64 discrete tokens with $\Delta \text{E}<2$ fidelity. We further build a large-scale dataset, AceTone-800K, and train a vision-language model to predict LUT tokens, followed by reinforcement learning to align outputs with perceptual fidelity and aesthetics. Experiments show that AceTone achieves state-of-the-art performance on both text-guided and reference-guided grading tasks, improving LPIPS by up to **50%** over existing methods. Human evaluations confirm that AceTone’s results are visually pleasing and stylistically coherent, demonstrating a new pathway toward language-driven, aesthetic-aligned color grading. The models and datasets will be publicly available.