Superpixels partition an image into perceptually coherent regions, reducing the cost of downstream vision tasks. Modern deep learning methods excel at superpixel generation but often yield irregular boundaries and isolated pixels, necessitating non-differentiable post-processing to enforce connectivity. This undermines the end-to-end learning capabilities. We propose a simple, fully differentiable graph-Laplacian loss that encourages spatial regularity and connectivity during training. The loss is model-agnostic and can be seamlessly integrated into the training of existing architectures to improve the quality of superpixels. In addition, we introduce two novel metrics, the average stray pixel count and excess component count, to measure the quality of superpixels. We demonstrate both qualitative and quantitative improvements over state-of-the-art methods with and without enforced connectivity. Our approach represents a significant step toward eliminating non-differentiable post-processing.

Recent advances in pretraining 3D point cloud encoders (e.g., Point-BERT, Point-MAE) have produced powerful models, whose abilities are typically evaluated on geometric or semantic tasks. At the same time, topological descriptors have been shown to provide informative summaries of a shape's multiscale structure. In this paper we pose the question whether topological information can be derived from features produced by 3D encoders. To address this question, we first introduce DONUT, a synthetic benchmark with controlled topological complexity, and propose FILTR (Filtration Transformer), a learnable framework to predict persistence diagrams directly from frozen encoders. FILTR adapts a transformer decoder to treat diagram generation as a set prediction task. Our analysis on DONUT reveals that existing encoders retain only limited global topological signals, yet FILTR successfully leverages information produced by these encoders to approximate persistence diagrams. Our approach enables, for the first time, data-driven extraction of persistence diagrams from raw point clouds through an efficient learnable feed-forward mechanism.

This work proposes a new formulation to the long-standing problem of convex decomposition through learning feature fields, enabling the first feed-forward model for open-world learning of convex decomposition. Our method produces high-quality decompositions of 3D shapes into a union of convex bodies, which are essential to accelerate collision detection in physical simulation, amongst many other applications.The key insight is to adopt a feature learning approach and learn a continuous feature field that can later be clustered to yield a good convex decomposition via our self-supervised, purely-geometric objective derived from the classical definition of convexity.Our formulation can be used for single shape optimization, but more importantly, feature prediction unlocks scalable, self-supervised learning on large datasets resulting in the first learned open-world for convex decomposition.Experiments show that our decompositions are higher-quality than alternatives and generalize across open-world objects as well as across representations to meshes, CAD models, and even Gaussian splats.

We introduce a novel framework that directly learns a spectral basis for shape and manifold analysis from unstructured data, eliminating the need for traditional operator selection, discretization, and eigensolvers.Grounded in optimal-approximation theory, we train a network to decompose an implicit approximation operator by minimizing the reconstruction error in the learned basis over a chosen distribution of probe functions. For suitable distributions, they can be seen as an approximation of the Laplacian operator and its eigendecomposition, which are fundamental in geometry processing. Furthermore, our method recovers in a unified manner not only the spectral basis, but also the implicit metric's sampling density and the eigenvalues of the underlying operator. Notably, our unsupervised method makes no assumption on the data manifold, such as meshing or manifold dimensionality, allowing it to scale to arbitrary datasets of any dimension.On point clouds lying on surfaces in 3D and high-dimensional image manifolds, our approach yields meaningful spectral bases, that can resemble those of the Laplacian, without explicit construction of an operator. By replacing the traditional operator selection, construction, and eigendecomposition with a learning-based approach, our framework offers a principled, data-driven alternative to conventional pipelines. This opens new possibilities in geometry processing for unstructured data, particularly in high-dimensional spaces.

The escalating parameter counts in modern deep learningmodels pose a fundamental challenge to efficient trainingand resolution of overfitting. We address this by introducingthe Mapping Networks which replace the high dimensionalweight space by a compact, trainable latent vector based onthe hypothesis that the trained parameters of large networksreside on smooth, low-dimensional manifolds. Henceforth,the Mapping Theorem enforced by a dedicated MappingLoss, shows the existence of a mapping from this latentspace to the target weight space both theoretically and inpractice. Mapping Networks significantly reduce overfittingand achieve comparable to better performance than target network across complex vision and sequence tasks, including Image Classification, Deepfake Detection etc., with99.5%, i.e., around 500× reduction in trainable parameters.

Most research decoding brain signals into images, often using them as priors for generative models, has focused only on visual content. This overlooks the brain's natural ability to integrate auditory and visual information, for instance, sound strongly influences how we perceive visual scenes. To investigate this,we propose a new task of reconstructing continuous video stimuli from multimodal brain signals recorded during audiovisual stimulation. To enable this, we introduce CineBrain, the first large-scale dataset that synchronizes fMRI and EEG during audiovisual viewing, featuring six hours of The Big Bang Theory episodes for cross-modal alignment. We also conduct the first systematic exploration of combining fMRI and EEG for video reconstruction and present CineSync, a framework for reconstructing dynamic video using a Multi-Modal Fusion Encoder and a Neural Latent Decoder. CineSync achieves state-of-the-art performance in dynamic reconstruction, leveraging the complementary strengths of fMRI and EEG to improve visual fidelity. Our analysis shows that auditory cortical activations enhance decoding accuracy, highlighting the role of auditory input in visual perception.

Hyperspectral cameras rely on spectral filters, dispersive optics, or coded apertures, which reduce light throughput and increase hardware complexity. These systems face harsh trade-offs between spatial, spectral, and temporal resolution in inherently low-photon conditions. Computational imaging systems break through these trade-offs with compressive sensing, but have typically required complex optics and/or extensive computation. We present Spectrum from Defocus (SfD), a chromatic focal sweep method that achieves state-of-the-art hyperspectral imaging using only two off-the-shelf lenses, a grayscale sensor, and less than one second of reconstruction time. By capturing a chromatically-aberrated focal stack that preserves nearly all incident light, and reconstructing it with a fast physics-based iterative algorithm, SfD delivers sharp, accurate hyperspectral images. The combination of photon efficiency, optical simplicity, and physical interpretability makes SfD a promising solution for fast, compact, and interpretable hyperspectral imaging.

Event cameras provide superior temporal resolution, dynamic range, energy efficiency, and pixel bandwidth. Spiking Neural Networks (SNNs) naturally complement event data through discrete spike signals, making them ideal for event-based tracking. However, current approaches combining Artificial Neural Networks (ANNs) and SNNs suffer from suboptimal architectures that compromise energy efficiency and limit tracking performance. To address these limitations, we propose the first Transformer-based \textbf{S}pike-\textbf{D}riven \textbf{T}racking (SDTrack) pipeline. It incorporates a novel event frame aggregation method called Global Trajectory Prompt (GTP) and a Transformer-based tracker. The GTP method effectively captures global trajectory information and aggregates it with event streams into event frames to enhance spatiotemporal representation. The Transformer-based tracker comprises a fully spike-driven SNN backbone and a simple tracking head. The SDTrack pipeline operates end-to-end without data augmentation or post-processing. Extensive experiments demonstrate that our SDTrack-Tiny pipeline achieves competitive accuracy with only 19.61$M$ parameters and 8.16$mJ$ energy consumption, while our Base version achieves state-of-the-art accuracy across three datasets. Our work establishes a solid foundation for future neuromorphic vision research.

Long video understanding is essential for human-like intelligence, enabling coherent perception and reasoning over extended temporal contexts. While the emerging thinking-with-frames paradigm—which alternates between global temporal reasoning and local frame examination—has advanced the reasoning capabilities of video multi-modal large language models (MLLMs), it suffers from a significant efficiency bottleneck due to the progressively growing and redundant multi-modal context. To address this, we propose SpecTemp, a reinforcement learning-based Speculative Temporal reasoning framework that decouples temporal perception from reasoning via a cooperative dual-model design. In SpecTemp, a lightweight draft MLLM rapidly explores and proposes salient frames from densely sampled temporal regions, while a powerful target MLLM focuses on temporal reasoning and verifies the draft’s proposals, iteratively refining its attention until convergence. This design mirrors the collaborative pathways of the human brain, balancing efficiency with accuracy. To support training, we construct the SpecTemp-80K dataset, featuring synchronized dual-level annotations for coarse evidence spans and fine-grained frame-level evidence. Experiments across multiple video understanding benchmarks demonstrate that SpecTemp not only maintains competitive accuracy but also significantly accelerates inference compared with existing thinking-with-frames methods.

Recent unified multi-modal models have made unprecedented progress in understanding and generation, yet they largely support multi-modal inputs with single-modality outputs, struggling to produce complex interleaved text–image content due to data scarcity and the difficulty of modeling long-range cross-modal context. We introduce Weaver, which frames interleaved generation as an autoregressive planning–visualization process within a unified multi-modal architecture. A planner, i.e., understanding expert, digests rich text–image context to produce visualization triggers and their dense textual guidance except for plain text, while a visualizer, i.e., generation expert, produces images conditioned on the planner’s textual guidance and visual references. This design enables decoupled learning: we train the two experts on large collections of textual planning and reference-guided image data in parallel, yielding powerful interleaved multi-modal generation capability at inference. Moreover, training the planner with datasets from diverse understanding and generation tasks equips the model with automatic task inference. To analyze and evaluate the model from multiple dimensions, we further introduce a benchmark that covers a range of everyday use cases. Extensive experiments show that, even without or with only very limited real interleaved data training, Weaver achieves superior performance on interleaved multi-modal generation.

Medical vision–language models can automate the generation of radiology reports but struggle with accurate visual grounding and factual consistency. Existing models often misalign textual findings with visual evidence, leading to unreliable or weakly grounded predictions. We present "CURE", an error-aware curriculum learning framework that improves grounding and report quality without any additional data. CURE tunes a multimodal instructional model on phrase grounding, grounded report generation, and anatomy-grounded report generation using public datasets. The method dynamically adjusts sampling based on model performance emphasizing on harder samples to improve spatial and textual alignment. CURE improves grounding accuracy by +0.37 IoU, boosts report quality by +0.188 CXRFEScore, and reduces hallucinations by 18.6%. CURE is a data-efficient framework that enhances both grounding accuracy and report reliability.

Large-scale foundation models pretrained on massive datasets have demonstrated strong generalization capabilities in medical image analysis. However, they are typically trained on static datasets and struggle to cope with the continuously evolving nature of clinical data, where new imaging devices, institutions, and disease subtypes constantly emerge. While domain-incremental learning (DIL) provides a solution for sequential adaptation without revisiting historical data, existing methods typically assume fixed label spaces and limited domain heterogeneity, restricting their applicability to real-world clinical scenarios. To address these challenges, we propose DK-DDIL, a rehearsal-free framework for dynamic DIL that integrates two synergistic modules: a Dynamic Adaptation Module (DAM) employing dynamic rank selection and adaptive regularization to flexibly allocate model capacity under domain shifts, and a Knowledge Inheritance and Refinement (KIR) module that stabilizes cross-domain knowledge transfer through selective adapter fusion and prototype-level contrastive refinement. Experiments on the Skin Pathology Diagnosis dataset, the Cyst-X 3D MRI cohort, and the OfficeHome benchmark demonstrate that DK-DDIL consistently outperforms state-of-the-art DIL approaches, highlighting its effectiveness and versatility across dynamic 2D medical, 3D medical, and natural image domains.

Curvilinear structure segmentation is essential in domains such as medical imaging, remote sensing, and materials science. Existing methods often require extensive domain-specific training and lack generalization to novel domains. To overcome these limitations, we propose the Segment Anything Curve Model (SACM) — a universal, curvilinear segmentation framework built upon the pretrained Segment Anything Model (SAM). SACM introduces a dual-level adapter architecture that enables both fine-grained and domain-adaptive enhancement: block-level internal adapters refine local structural representations, while external adapters facilitate cross-domain feature alignment. Specifically, the internal adapters are embedded within each Transformer block to locally adapt and refine features for thin and intricate curvilinear patterns, while the external adapters operate across blocks to capture global, multi-layer contextual information and facilitate domain adaptation. Furthermore, SACM introduces a feature fusion mechanism that aggregates multi-layer features from all external adapters and fuses them via a Feed-Forward Network (FFN) module, and a dual-stage refinement process in the mask decoder to enhance topology and connectivity. This design enables prompt-free, data-efficient fine-tuning and achieves robust cross-domain generalization when trained with only 18 annotated images. Extensive experiments across twelve diverse curvilinear datasets validate that SACM achieves state-of-the-art performance.

Medical vision-language models (VLMs) are strong zero-shot recognizers for medical imaging, but their reliability under domain shift hinges on calibrated uncertainty with guarantees. Split conformal prediction (SCP) offers finite-sample coverage, yet prediction sets often become large (low efficiency) and class-wise coverage unbalanced—high class-conditioned coverage gap (CCV), especially in few-shot, imbalanced regimes; moreover, naively adapting to calibration labels breaks exchangeability and voids guarantees. We propose $\texttt{\textbf{LATA}}$ (Laplacian-Assisted Transductive Adaptation), a $\textit{training- and label-free}$ refinement that operates on the joint calibration and test pool by smoothing zero-shot probabilities over an image–image $k$NN graph using a small number of CCCP mean-field updates, preserving SCP validity via a deterministic transform. We further introduce a $\textit{failure-aware}$ conformal score that plugs into the vision-language uncertainty (ViLU) framework, providing instance-level difficulty and label plausibility to improve prediction set efficiency and class-wise balance at fixed coverage. $\texttt{\textbf{LATA}}$ is black-box (no VLM updates), compute-light (windowed transduction, no backprop), and includes an optional prior knob that can run strictly label-free or, if desired, in a label-informed variant using calibration marginals once. Across $\textbf{three}$ medical VLMs and $\textbf{nine}$ downstream tasks, $\texttt{\textbf{LATA}}$ consistently reduces set size and CCV while matching or tightening target coverage, outperforming prior transductive baselines and narrowing the gap to label-using methods, while using far less compute. Comprehensive ablations and qualitative analyses show that $\texttt{\textbf{LATA}}$ sharpens zero-shot predictions without compromising exchangeability.

Lesion detection, symptom tracking, and visual explainability are central to real-world medical image analysis, yet current medical Vision-Language Models (VLMs) still lack mechanisms that translate their broad knowledge into clinically actionable outputs. To bridge this gap, we present Medic-AD, a clinically oriented VLM that strengthens these three capabilities through a stage-wise framework. First, learnable anomaly-aware tokens (Ano) encourage the model to focus on abnormal regions and build more discriminative lesion centered representations. Second, inter-image difference tokens (Diff) explicitly encode temporal changes between studies, allowing the model to distinguish worsening, improvement, and stability in disease burden. Finally, a dedicated explainability stage trains the model to generate heatmaps that highlight lesion-related regions, offering clear visual evidence that is consistent with the model's reasoning. Through our staged design, Medic-AD steadily boosts performance across anomaly detection, symptom tracking, and anomaly segmentation, achieving state-of-the-art results compared with both closed source and medical-specialized baselines. Evaluations on real longitudinal clinical data collected from real hospital workflows further show that Medic-AD delivers stable predictions and clinically faithful explanations in practical patient-monitoring and decision-support workflows.

Medical image segmentation is vital for clinical diagnosis and quantitative analysis, yet remains challenging due to the heterogeneity of imaging modalities and the high cost of pixel-level annotations. Although general interactive segmentation models like SAM have achieved remarkable progress, their transfer to medical imaging still faces two key bottlenecks: (i) the lack of adaptive mechanisms for modality- and anatomy-specific tasks, which limits generalization in out-of-distribution medical scenarios; and (ii) current medical adaptation methods fine-tune on large, heterogeneous datasets without selection, leading to noisy supervision, higher cost, and negative transfer. To address these issues, we propose SegMoTE, an efficient and adaptive framework for medical image segmentation. SegMoTE preserves SAM’s original prompt interface, efficient inference, and zero-shot generalization while introducing only a small number of learnable parameters to dynamically adapt across modalities and tasks. In addition, we design a progressive prompt tokenization mechanism that enables fully automatic segmentation, significantly reducing annotation dependence. Trained on MedSeg-HQ, a curated dataset less than 1% of existing large-scale datasets, SegMoTE achieves SOTA performance across diverse imaging modalities and anatomical tasks. It represents the first efficient, robust, and scalable adaptation of general segmentation models to the medical domain under extremely low annotation cost, advancing the practical deployment of foundation vision models in clinical applications.

Deep learning-based methods have revolutionized the field of imaging inverse problems, yielding state-of-the-art performance across various imaging domains. The best performing networks incorporate the imaging operator within the network architecture, typically in the form of deep unrolling. However, in large-scale problems, such as 3D imaging, most existing methods fail to incorporate the operator in the architecture due to the prohibitive amount of memory required by global forward operators, which hinder typical patching strategies. In this work, we present a domain partitioning strategy and normal operator approximations that enable the training of end-to-end reconstruction models incorporating forward operators of arbitrarily large problems into their architecture. The proposed method achieves state-of-the-art performance on 3D X-ray cone-beam tomography and 3D multi-coil accelerated MRI, while requiring only a single GPU for both training and inference.

Federated learning (FL) has emerged as a widely adopted training paradigm for privacy-preserving machine learning. Despite the past success of SGD-based methods, they still suffer from severe data heterogeneity and the lack of adaptivity in practical applications. While several adaptive federated optimization methods (such as FedAdam) have been proposed and demonstrated to achieve faster convergence, they fail to show significant improvements in generalization performance under highly heterogeneous data distributions, and their optimization and generalization mechanisms remain insufficiently understood. To fill this gap, we introduce diffusion theory into the adaptive federated optimization framework and analyze the distinct effects of adaptive learning rate and global momentum from the perspectives of saddle-point escaping and flat-minima selection. Theoretical results show that although FedAdam outperforms FedAvg/FedAvgM in escaping saddle points, the latter escapes sharp minima more efficiently. The root cause lies in that adaptive learning rates, while enhancing saddle-point escape, weaken the preference for flat minima. Motivated by these insights, we propose FedAdamom, a new adaptive federated optimization algorithm that adapts the momentum hyperparameter rather than the learning rate. FedAdamom maintains strong saddle-point escaping capability while enhancing flat-minima selection. We further establish its convergence guarantees under non-convex objectives. Extensive experiments demonstrate that FedAdamom significantly outperforms existing adaptive federated optimization methods in terms of convergence speed, generalization performance, and preference for flat minima.

Achieving fully autonomous driving systems requires learning rational decisions in a wide span of scenarios, including safety-critical and out-of-distribution ones. However, such cases are underrepresented in real-world corpus collected by human experts. To complement for the lack of data diversity, we introduce a novel and scalable simulation framework capable of synthesizing these crucial massive unseen states upon existing driving logs. Our pipeline utilizes advanced neural rendering with a reactive environment to generate high-fidelity multi-view observations controlled by ego trajectory perturbations. Furthermore, we develop a pseudo-expert trajectory generation mechanism to provide feasible action supervision for these newly simulated states to provide action supervision.Upon the synthesized data, we find that a simple co-training strategy on both real-world and simulated samples can lead to significant improvements in both robustness and generalization for various planning methods on challenging real-world benchmarks, up to +6.8 EPDMS on navhard and +2.9 on navtest. More importantly, such policy improvement scales smoothly by increasing simulation data only, even without extra real-world data streaming in. We further reveal crucial findings of such a sim-real paradigm, includingthe design of pseudo-experts and the scaling properties for different policy architectures. Simulation data and code would be released.

Current event simulation methods focus on employing videos to synthesize new event data, suffering from costly video capture and limited scalability across viewpoints, motions, and lighting. To this end, we propose a Text-to-event simulation framework (Texvent) that can directly generate asynchronous event data from simple text prompts. Texvent first renders prompt-driven videos via multimodal large language models and subsequently applies a new physical simulator to generate event streams. Specifically, an adaptive brightness-aware frame interpolation approach is proposed to enhance the temporal resolution of the rendered videos. A balanced logarithmic intensity comparison strategy and a cache–based voltage refreshment mechanism are introduced into the simulator to generate event data.To narrow the sim-to-real gap, we also introduce background activity noise injection and dense time stamp reconstruction operations. Extensive experiments demonstrate Texvent’s superior computational efficiency and its ability to generate more realistic event data than existing simulators.

Generative world models are reshaping embodied AI, enabling agents to synthesize realistic 4D driving environments that look convincing but often fail physically or behaviorally. Despite rapid progress, the field still lacks a unified way to assess whether generated worlds preserve geometry, obey physics, or support reliable control. We introduce WorldLens, a full-spectrum benchmark evaluating how well a model builds, understands, and behaves within its generated world. It spans five aspects - Generation, Reconstruction, Action-Following, Downstream Task, and Human Preference - jointly covering visual realism, geometric consistency, physical plausibility, and functional reliability. Across these dimensions, no existing world model excels universally: those with strong textures often violate physics, while geometry-stable ones lack behavioral fidelity. To align objective metrics with human judgment, we further construct WorldLens-26K, a large-scale dataset of human-annotated videos with numerical scores and textual rationales, and develop WorldLens-Agent, an evaluation model distilled from these annotations to enable scalable, explainable scoring. Together, the benchmark, dataset, and agent form a unified ecosystem for measuring world fidelity - standardizing how future models are judged not only by how real they look, but by how real they behave.

In this paper, we propose BuildingGPT, a novel auto-regressive model for building wireframe reconstruction from point clouds with reinforcement learning.Unlike prior works based on detection or diffusion models, BuildingGPT reformulates the building wireframe reconstruction task into a sequence prediction problem.Based on the hierarchical building wireframe tokenization, the wireframe sequences are organized in a structurally- and semantically-aware order for the next-token prediction.The point cloud encoder first transforms the input point cloud into a fixed-length latent code that serves as the starting of the sequence.Then, BuildingGPT auto-regressively predicts tokens conditioned on the latent code and previously generated tokens.With token sequence predicted, the building wireframe is obtained through detokenization.To enhance the model performance, we adopt a two-stage training paradigm including the pre-training and post-training.After the auto-regressive pre-training, Direct Preference Optimization (DPO) is employed as a post-training strategy to align reconstruction results with human preferences.Extensive experiments on the large-scale MunichWF dataset show that BuildingGPT outperforms existing state-of-the-art methods.We commit to release the code and dataset.

3D foundation models (3DFMs) have recently transformed 3D vision, enabling joint prediction of depths, poses, and point maps directly from images. Yet their ability to reason under extreme, non-overlapping views remains largely unexplored. In this work, we study their internal representations and find that 3DFMs exhibit an emergent understanding of extreme-view geometry, despite never being trained for such conditions. To further enhance these capabilities, we introduce a lightweight alignment scheme that refines their internal 3D representation by tuning only a small subset of backbone bias terms, leaving all decoder heads frozen. This targeted adaptation substantially improves relative pose estimation under extreme viewpoints without degrading per-image depth or point quality. Additionally, we contribute MegaUnScene, a new benchmark of Internet scenes unseen by existing 3DFMs, with dedicated test splits for both relative pose estimation and dense 3D reconstruction. All code and data will be released.

3D vision foundation models like Visual Geometry Grounded Transformer (VGGT) have advanced greatly in geometric perception. However it is time-consuming and memory-intensive for long sequences, limiting application to large-scale scenes beyond hundreds of images. To address this, we propose LiteVGGT, achieving up to 10× speedup and substantial memory reduction, enabling efficient processing of 1000-image scenes. We derive two key insights for 3D reconstruction: 1) tokens from local image regions have inherent geometric correlations, leading to high similarity and computational redundancy; 2) token similarity acroses adjacent network layers remains stable, allowing for reusable merge decisions. Guided by these, we design a simple yet efficient strategy, dubbed geometry-aware cached token merging . We analyze each token’s geometric importance, optimizing anchor token selection to better preserve key information for reconstruction. We also cache and reuse merge indices across layers, substantially reducing latency with minimal accuracy impact. This strategy retains VGGT’s core performance, enabling efficient fine-tuning and FP8 quantization for further gains. Extensive experiments validate LiteVGGT’s effectiveness, scalability, and robustness.

Recent feed-forward reconstruction models like VGGT and $\pi^3$ achieve impressive reconstruction quality but cannot process streaming videos due to quadratic memory complexity, limiting their practical deployment. While existing streaming methods address this through learned memory mechanisms or causal attention, they require extensive retraining and may not fully leverage the strong geometric priors of state-of-the-art offline models. We propose LASER, a training-free framework that converts an offline reconstruction model into a streaming system byaligning predictions across consecutive temporal windows. We observe that simple similarity transformation ($Sim(3)$) alignment fails due to layer depth misalignment: monocular scale ambiguity causes relative depth scales of different scene layers to vary inconsistently between windows. To address this, we introduce layer-wise scale alignment, which segments depth predictions into discrete layers, computes per-layer scale factors, and propagates them across both adjacent windows and timestamps.Extensive experiments show that LASER achieves state-of-the-art performance on camera pose estimation and point map reconstruction while operating at 14 FPS with 6 GB peak memory on a RTX A6000 GPU, enabling practical deployment for kilometer-scale streaming videos.Our code will be released publicly.

Panoramic imagery offers a full $360^\circ$ field of view and is increasingly common in consumer devices. However, it introduces non-pinhole distortions that challenge joint pose estimation and 3D reconstruction. Existing feed-forward models, built for perspective cameras, generalize poorly to this setting.We propose PanoVGGT, a permutation-equivariant Transformer framework that jointly predicts camera poses, depth maps, and 3D point clouds from one or multiple panoramas in a single forward pass. The model incorporates spherical-aware positional embeddings and a panorama-specific three-axis SO(3) rotation augmentation, enabling effective geometric reasoning in the spherical domain. To resolve inherent global-frame ambiguity, we further introduce a stochastic anchoring strategy during training. In addition, we contribute PanoCity, a large-scale outdoor panoramic dataset with dense depth and 6-DoF pose annotations. Extensive experiments on PanoCity and standard benchmarks demonstrate that PanoVGGT achieves competitive accuracy, strong robustness, and improved cross-domain generalization. Code and dataset will be released.

Robust 3D environmental perception is critical for applications like autonomous navigation and robotics, yet existing optical sensors like cameras and LiDAR fail in adverse conditions such as smoke, fog, and non-ideal lighting. While specialized radar systems can operate in these conditions, their reliance on bespoke, ultra-wideband hardware and licensed spectrum limits their scalability and cost-effectiveness. This paper introduces Rascene, a novel framework that enables high-fidelity 3D imaging by repurposing ubiquitous mmWave OFDM communication signals. Recognizing that a single-frame RF signal is inherently sparse, noisy, and highly ambiguous, the key innovation of Rascene is a multi-frame 3D imaging framework designed to fuse information from signals captured across multiple, arbitrary poses. This framework leverages a spatially adaptive fusion mechanism to find geometric consensus across the multiple views, effectively suppressing multipath artifacts while preserving sparse geometric details. Experiments demonstrate that our method reconstructs 3D scenes with high precision, providing a new pathway for low-cost, scalable, and robust 3D perception.

We present a scalable 3D reconstruction model that addresses a critical limitation in offline feed-forward methods: their computational and memory requirements grow quadratically w.r.t. the number of input images. Our approach is built on the key insight that this bottleneck stems from the varying-length Key-Value (KV) space representation of scene geometry, which we distill into a fixed-size Multi-Layer Perceptron (MLP) via test-time training.VGG-T$^3$ ($\mathbf{V}$isual $\mathbf{G}$eometry $\mathbf{G}$rounded $\mathbf{T}$est $\mathbf{T}$ime $\mathbf{T}$raining) scales linearly w.r.t. the number of input views, similar to online models, and achieves a $11.6\times$ speed-up over baselines that rely on softmax attention for reconstructing a $1k$ image collection in just $54$ seconds. Because our method retains global scene aggregation capability, our resulting point map reconstruction error is comparable to VGGT.

Although depth-assisted scene flow estimation has advanced rapidly, mainstream dense frameworks (e.g., RAFT-3D) still rely primarily on 2D feature correlations to optimize 3D motion fields, which hinders their ability to exploit 3D structural priors effectively and consequently limits robustness in complex scenes. We present SEA-Flow3D, a simple, efficient, and accurate framework for dense scene flow estimation.At its core lies a Spatial Vector Sampling (SVS) module that jointly samples 3D coordinates and correlation volumes within the local neighborhood of matched points, producing a direction-aware correlation representation with explicit spatial vectors and providing strong geometric guidance for subsequent optimization. Following the simplicity-and-efficiency principle, SEA-Flow3D adopts a RAFT-style multi-scale recurrent refinement architecture, integrating an RNN-based optimizer with context-guided upsampling to achieve higher accuracy with fewer iterations. Extensive experiments on KITTI and Sintel demonstrate that SEA-Flow3D achieves state-of-the-art performance while maintaining remarkable efficiency and a lightweight design.

General 3D foundation models have started to lead the trend of unifying diverse vision tasks, yet most assume RGB-only inputs and ignore readily available geometric cues (e.g., camera intrinsics, poses, and depth maps). To address this issue, we introduce OmniVGGT, a novel framework that can effectively benefit from an arbitrary number of auxiliary geometric modalities during both training and inference. In our framework, a GeoAdapter is proposed to encode depth and camera intrinsics/extrinsics into a spatial foundation model. It employs zero-initialized convolutions to progressively inject geometric information without disrupting the foundation model's representation space. This design ensures stable optimization with negligible overhead, maintaining inference speed comparable to VGGT even with multiple additional inputs. Additionally, a stochastic multimodal fusion regimen is proposed, which randomly samples modality subsets per instance during training. This enables an arbitrary number of modality inputs during testing and promotes learning robust spatial representations instead of overfitting to auxiliary cues. Comprehensive experiments on monocular/multi-view depth estimation, multi-view stereo, and camera pose estimation demonstrate that OmniVGGT outperforms prior methods with auxiliary inputs and achieves state-of-the-art results even with RGB-only input. To further highlight its practical utility, we integrated OmniVGGT into vision-language-action (VLA) models. The enhanced VLA model by OmniVGGT not only outperforms the vanilla point-cloud-based baseline on mainstream benchmarks, but also effectively leverages accessible auxiliary inputs to achieve consistent gains on robotic tasks.

We present a robust, real-time RGB SLAM system that handles dynamic environments by leveraging differentiable Uncertainty-aware Bundle Adjustment. Traditional SLAM methods typically assume static scenes, leading to tracking failures in the presence of motion. Recent dynamic SLAM approaches attempt to address this challenge using predefined dynamic priors or uncertainty-aware mapping, but they remain limited when confronted with unknown dynamic objects or highly cluttered scenes where geometric mapping becomes unreliable. In contrast, our method estimates per-pixel uncertainty by exploiting multi-view visual feature inconsistency, enabling robust tracking and reconstruction even in real-world environments. The proposed system achieves state-of-the-art camera poses and scene geometry in cluttered dynamic scenarios while running in real time at around 8 FPS. The source code will be publicly released.

Visual Geometry Grounded Transformer (VGGT) has shown significant progress in 3D vision tasks. However, its global attention layers incur quadratic computational cost with respect to the number of input views, becoming a critical bottleneck for scalability. Several sparsification-based acceleration techniques have been proposed to alleviate this issue, but they often suffer from substantial accuracy degradation. We hypothesize that the accuracy degradation stems from the heterogeneity in head-wise sparsification sensitivity, as the existing methods apply a uniform sparsity pattern across all heads. Motivated by this hypothesis, we present a two-stage sparsification pipeline that effectively quantifies and exploits head-wise sparsification sensitivity. In the first stage, we measure head-wise sparsification sensitivity using a novel metric, the Head Sensitivity Score (HeSS), which approximates the Hessian with respect to two distinct error terms on a small calibration set. In the inference stage, we perform HeSS-Guided Sparsification, leveraging the pre-computed HeSS to reallocate the total attention budget—assigning denser attention to sensitive heads and sparser attention to more robust ones. We demonstrate that HeSS effectively captures head-wise sparsification sensitivity and empirically confirm that attention heads in the global attention layers exhibit heterogeneous sensitivity characteristics. Extensive experiments further show that our method effectively mitigates performance degradation under high sparsity, demonstrating strong robustness across varying sparsification levels.

Direct Time-of-Flight (dToF) sensors provide highly accurate metric depth and are more robust than indirect ToF systems in challenging real-world conditions. However, their high manufacturing cost and limited photodiode array size produce depth maps that are extremely sparse, low-resolution, and noisy, making them unsuitable for VR/XR, robotics, and 3D perception tasks that require dense metric depth. Existing monocular and depth completion methods struggle to handle the unique sampling patterns and hardware artifacts of dToF devices, and their performance often deteriorates significantly under severe sparsity or noise. We present a generalizable framework for dense metric depth completion from sparse dToF measurements, capable of operating across diverse sensor types, sparsity levels, and noise conditions. Our model employs a depth-guided dual-branch Vision Transformer encoder that processes RGB images and sparse dToF measurements separately, while a masked joint attention module allows depth tokens to reliably guide image features without being overwritten by them. A lightweight decoder reconstructs dense metric depth efficiently, without diffusion-based or refinement-heavy post-processing. To address the scarcity of paired training data, we introduce a comprehensive dToF simulation pipeline that reproduces the characteristics of flash, sub-VGA flash, and rotating sensors, including hardware-induced degradation, irregular sparsity, and realistic noise distributions. Trained entirely on synthetic data, our model achieves strong zero-shot generalization across 6 datasets and 2 real dToF devices, outperforming state-of-the-art approaches in both accuracy and computational efficiency. This establishes a robust and practical solution for dense metric depth completion from sparse direct ToF sensors.

We present Online3R, a new sequential reconstruction framework that is capable of adapting to new scenes through online learning, effectively resolving inconsistency issues. Specifically, we introduce a set of learnable lightweight visual prompts into a pretrained, frozen geometry foundation model to capture the knowledge of new environments while preserving the fundamental capability of the foundation model for geometry prediction. To solve the problems of missing groundtruth and the requirement of high efficiency when updating these visual prompts at test time, we introduce a local-global self-supervised learning strategy by enforcing the local and global consistency constraints on predictions. The local consistency constraints are conducted on intermediate and previously local fused results, enabling the model to be trained with high-quality pseudo groundtruth signals; the global consistency constraints are operated on sparse keyframes spanning long distances rather than per frame, allowing the model to learn from a consistent prediction over a long trajectory in an efficient way. Our experiments demonstrate that Online3R outperforms previous state-of-the-art methods on various benchmarks.

emporally consistent surface reconstruction of dynamic 3D objects from unstructured point cloud data remains challenging, especially for very long sequences. Existing methods either optimize deformations incrementally, risking drift and requiring long runtimes, or rely on complex learned models that demand category-specific training. We present Neu-PiG, a fast optimization method based on a novel preconditioned surface encoding that estimates coherent non-rigid deformations without sacrificing temporal stability or accuracy. Our method encodes entire deformations across all time steps at various spatial scales into a multi-resolution latent grid, parameterized by the position and normal direction of a reference surface from a single keyframe. This latent representation is then augmented for time modulation and decoded into per-frame 6-DoF deformations via a lightweight multi-layer perceptron (MLP). To achieve high-fidelity, drift-free surface reconstructions in seconds, we employ Sobolev preconditioning during gradient-based training of the latent space, completely avoiding the need for any explicit correspondences or further priors. Experiments across diverse human and animal datasets demonstrate that Neu-PiG outperforms state-the-art approaches, offering both superior accuracy and scalability to long sequences while running at least 60× faster than existing training-free methods and even matching the inference-time performance of heavy pretrained models.

We introduce a test-time framework for multiview Transformers (MVTs) that incorporates priors (e.g., camera poses, intrinsics, and depth) to improve 3D tasks, without retraining or modifying the pre-trained image-only networks. Rather than feeding priors into the architecture, we cast them as constraints on the predictions and optimize the network at inference. The optimization loss is composed of a self-supervised objective and prior penalty terms. The self-supervised objective is defined as the compatibility among multi-view predictions, implemented by the photometric or geometric loss between the renderings from other views and each view itself. Any available priors are turned into the penalty terms on the corresponding output modalities. Across a series of 3D vision benchmarks, including point map estimation and camera pose estimation, our method consistently improves performance over base MVTs by a large margin. On ETH3D, 7-Scenes, and NRGBD datasets, our method cuts the point map distance error by more than half compared to the base image-only models. Our method also outperforms those re-trained prior-aware feed-forward methods, demonstrating the effectiveness of our test-time constrained optimization (TCO) framework, in incorporating priors for 3D vision tasks.

Symmetry detection is a fundamental problem in computer vision, and symmetries serve as powerful priors for downstream tasks. However, existing learning-based methods for detecting 3D symmetries from single images have been almost exclusively trained and evaluated on object-centric or synthetic datasets, and thus fail to generalize to real-world scenes. Furthermore, due to the inherent scale ambiguity of monocular inputs, which makes localizing the 3D plane an ill-posed problem, many existing works only predict the plane's orientation. In this paper, we address these limitations by presenting the first framework for detecting 3D-grounded reflectional symmetries from single, in-the-wild RGB images, focusing on architectural landmarks. We introduce two key innovations: (1) a scalable data annotation pipeline to automatically curate a large-scale dataset of architectural symmetries, ArchSym, from SfM reconstructions by leveraging cross-view image matching; and building on the dataset, (2) a single-view symmetry detector that accurately localizes symmetries in 3D by parameterizing them as signed distance maps defined relative to predicted scene geometry. We validate our symmetry annotation pipeline against geometry-based alternatives and demonstrate that our symmetry detector significantly outperforms state-of-the-art baselines on our new benchmark.

Scene-level point cloud understanding remains challenging due to diverse geometries, imbalanced categories, and highly varied spatial layouts. Existing methods improve object-level performance but rely on static parameters during inference, limiting their adaptability to dynamic scene data. We propose Test-time Parameter Adaptation for Point Cloud Scene Perception (PointTPA), a test-time dynamic adaptation framework that constructs input-aware parameters for scene-level point clouds. PointTPA uses a Serialization-based Neighborhood Grouping (SNG) to form locally coherent patches and a Dynamic Parameter Projector (DPP) to produce patch-wise adaptive weights, enabling the backbone to adjust its behavior according to scene-specific variations while keeping parameter cost low. Integrated into PTv3, PointTPA reduces trainable parameters by over 95% and achieves competitive or superior performance to full fine-tuning. It achieves 74.9% mIoU on S3DIS and consistently surpasses existing PEFT baselines across multiple benchmarks, highlighting the efficacy of test-time dynamic parameter generation in enhancing robust 3D scene understanding. The code will be available soon.

We propose tttLRM, a novel large 3D reconstruction model that leverages a Test-Time Training (TTT) layer to enable long-context, autoregressive 3D reconstruction with linear computational complexity, further scaling the model’s capability. Our framework efficiently compresses multiple image observations into the fast weights of the TTT layer, forming an implicit 3D representation in the latent space that can be decoded into various explicit formats, such as Gaussian Splats (GS) for downstream applications. The online learning variant of our model supports progressive 3D reconstruction and refinement from streaming observations. We demonstrate that pretraining on novel view synthesis tasks effectively transfers to explicit 3D modeling, resulting in improved reconstruction quality and faster convergence. Extensive experiments show that our method achieves superior performance in feedforward 3D Gaussian reconstruction compared to state-of-the-art approaches on both objects and scenes.

Recent unified models have made remarkable strides in generating high-quality images, yet they consistently fail on reasoning-intensive tasks, i.e., solving mazes, assembling tangrams. Intriguingly, we find that vision-language models (VLMs) and large language models (LLMs) can accurately solve these tasks, but cannot generate the corresponding images because they lack a structured visual output interface. This reveals that the core bottleneck is not reasoning capacity, but the lack of a structured interface to translate high-level reasoning into precise visual output. To bridge this gap, we propose using code-structured visual hints (i.e., SVG/HTML) overlays that explicitly encode reasoning steps directly on the image plane. Accordingly, we develop an automatic data construction pipeline that can generate high-quality code-structured hints for existing datasets and train a unified model called Hint2Gen based on FLUX.1 Kontext to condition its generation on such hints. Furthermore, to comprehensively evaluate the effectiveness of our approach, we introduce Reason2Gen, a benchmark comprising 4,000 samples spanning 20 categories across 7 core dimensions, including path connectivity, spatial assembly, etc. Extensive experiments demonstrate that even simply providing such hints as extra inputs—without any retraining—boosts their performance. And our model significantly outperforms all leading open-source/closed-source methods on reasoning-aware generation and editing across all the dimensions.

Despite the rapid progress of text-to-image (T2I) models, generating images that accurately reflect complex compositional prompts (covering attribute bindings, object relationships, counting) still remains challenging. To address this, we propose BIDPO, a framework to enhance T2I model's capability of compositional text-to-image generation. We begin by introducing an carefully designed pipeline to construct a large-scale preference dataset, BICOMP, with strictly quality control. Then, we extend Diffusion DPO to jointly optimize image and text preferences, which is shown to greatly effective in improving the models to follow complex text prompt in generation. To further enhance the models for fine-grained alignment, we employ a region-level guidance method to focus on regions relevant to compositional concepts. Experimental results demonstrate that our BIDPO substantially improves compositional fidelity, consistently outperforming prior methods across multiple benchmarks. Our approach highlights the potential of preference-based fine-tuning for complex text-to-image tasks, offering a flexible and scalable alternative to existing techniques.

Compositional generalization -- the ability to understand and generate novel combinations of learned concepts -- enables models to extend their capabilities beyond limited experiences. While effective, the data structures and principles that enable this crucial capability remain poorly understood. We propose that compositional generalization fundamentally requires decomposing high-level concepts into basic, low-level concepts that can be recombined across similar contexts, similar to how humans draw analogies between concepts. For example, someone who has never seen a peacock eating rice can envision this scene by relating it to their previous observations of a chicken eating rice.In this work, we formalize these intuitive processes using principles of causal modularity and minimal changes. We introduce a hierarchical data-generating process that naturally encodes different levels of concepts and their interaction mechanisms. Theoretically, we demonstrate that this approach enables compositional generalization supporting complex relations between composed concepts, advancing beyond prior work that assumes simpler interactions like additive effects. Critically, we also prove that this latent hierarchical structure is provably recoverable (identifiable) from observable data like text-image pairs, a necessary step for learning such a generative process. To validate our theory, we apply insights from our theoretical framework and achieve significant improvements on benchmark datasets.

Significant progress has been achieved in high-fidelity video synthesis, yet current paradigms often fall short in effectively integrating identity information from multiple subjects. This leads to semantic conflicts and suboptimal performance in preserving identities and interactions, limiting controllability and applicability. To tackle this issue, we introduce ID-Crafter, a framework for multi-subject video generation that achieves superior identity preservation and semantic coherence. ID-Crafter integrates three key components: (i) a hierarchical identity-preserving attention mechanism that progressively aggregates features at intra-subject, inter-subject, and cross-modal levels; (ii) a semantic understanding module powered by a pretrained Vision-Language Model (VLM) to provide fine-grained guidance and capture complex inter-subject relationships; and (iii) an online reinforcement learning phase to further refine the model for critical concepts. Furthermore, we construct a new dataset to facilitate robust training and evaluation. Extensive experiments demonstrate that ID-Crafter establishes new state-of-the-art performance on multi-subject video generation benchmarks, excelling in identity preservation, temporal consistency, and overall video quality.

Recent years have seen impressive advances in text-to-image generation, with image generative or unified models, generating high-quality images from text. Yet these models still struggle with fine-grained color controllability, often failing to accurately match colors specified in text prompts. While existing benchmarks evaluate compositional reasoning and prompt adherence, none systematically assess the color precision. Color is fundamental to human visual perception and communication, critical for applications from art to design workflows requiring brand consistency. However, current benchmarks either neglect color or rely on coarse assessments, missing key capabilities like interpreting RGB values or aligning with human expectations. To this end, we propose GenColorBench, the first comprehensive benchmark for T2I color generation, grounded in color systems like ISCC-NBS and CSS3/X11, including numerical colors which are absent elsewhere. With 44K color-focused prompts covering 400+ colors, it reveals models’ true capabilities via perceptual and automated assessments. Evaluations of popular T2I models using GenColorBench show performance variations, highlighting which color conventions models understand best and identifying failure modes. Our GenColorBench assessments will allow to guide improvements in precise color generation. The benchmark will be made public upon acceptance.

Few-step generation has been a long-standing goal, with recent one-step generation methods exemplified by MeanFlow achieving remarkable results.Existing research on MeanFlow primarily focuses on class-to-image generation. However, an intuitive yet unexplored direction is to extend the condition from fixed class labels to flexible text inputs, enabling richer content creation.Compared to the limited class labels, text conditions pose greater challenges to the model’s understanding capability, necessitating the effective integration of powerful text encoders into the MeanFlow framework. Surprisingly, although incorporating text conditions appears straightforward, we find that integrating powerful LLM-based text encoders using conventional training strategies results in unsatisfactory performance.To uncover the underlying cause, we conduct detailed analyses and reveal that, due to the extremely limited number of refinement steps in the MeanFlow generation, such as only one step, the text feature representations are required to possess sufficiently high discriminability.This also explains why discrete and easily distinguishable class features perform well within the MeanFlow framework. Based on this insight, we propose a novel auxiliary loss design to learn the discriminative text representation space, achieving an effective adaptation of MeanFlow to text-to-image generation for the first time. Furthermore, we validate our approach on the widely used diffusion model, demonstrating significant generation performance improvements. We hope this work provides a general and practical reference for future research on text-conditioned MeanFlow generation.

Recent text-to-image (T2I) diffusion models produce visually stunning images and demonstrate excellent prompt following. But do they perform well as synthetic vision data generators? In this work, we revisit the promise of synthetic data as a scalable substitute for real training sets and uncover a surprising performance regression.We generate large-scale synthetic datasets using state-of-the-art T2I models released between 2022 and 2025, train standard classifiers solely on this synthetic data, and evaluate them on real test data.Despite observable advances in visual fidelity and prompt adherence, classification accuracy on real test data consistently declines with newer T2I models as training data generators.Our analysis reveals a hidden trend: These models collapse to a narrow, aesthetic-centric distribution that undermines diversity and label-image alignment. Overall, our findings challenge a growing assumption in vision research, namely that progress in generative realism implies progress in data realism. We thus highlight an urgent need to rethink the capabilities of modern T2I models as reliable training data generators.

Recent advances in generative video models have enabled the creation of high-quality videos based on natural language prompts. However, these models frequently lack fine-grained temporal control, meaning they do not allow users to specify when particular visual elements should appear within a generated sequence. In this work, we introduce TempoControl, a method that allows for temporal alignment of visual concepts during inference, without requiring retraining or additional supervision. TempoControl utilizes cross-attention maps, a key component of text-to-video diffusion models, to guide the timing of concepts through a novel optimization approach. Our method steers attention using three complementary principles: aligning its temporal pattern with a control signal (correlation), adjusting its strength where visibility is required (magnitude), and preserving semantic consistency (entropy). TempoControl provides precise temporal control while maintaining high video quality and diversity. We demonstrate its effectiveness across various applications, including temporal reordering of single and multiple objects, action timing, and audio-aligned video generation.

This work introduces a new task, text-conditioned selective video-to-audio (V2A) generation, which produces only the user-intended sound from a multi-object video. This capability is especially crucial in multimedia production, where audio tracks are handled individually for each sound source for precise editing, mixing, and creative control. However, current approaches generate single source-mixed sounds at once, largely because visual features are entangled, and region cues or prompts often fail to specify the source. We propose SelVA, a novel text-conditioned V2A model that treats the text prompt as an explicit selector of target source and modulates video encoder to distinctly extract prompt-relevant video features. The proposed supplementary tokens promote cross-attention by suppressing text-irrelevant activations with efficient parameter tuning, yielding robust semantic and temporal grounding. SelVA further employs a self-augmentation scheme to overcome the lack of mono audio track supervision. We evaluate SelVA on VGG-MonoAudio, a curated benchmark of clean single-source videos for such a task. Extensive experiments and ablations consistently verify its effectiveness across audio quality, semantic alignment, and temporal synchronization. Code and demo are available in the supplementary material.

Multi-subject image generation aims to synthesize user-provided subjects in a single image while preserving subject fidelity, ensuring prompt consistency, and aligning with human aesthetic preferences. Existing In-Context-Learning based methods are limited by their highly coupled training paradigm. These methods attempt to achieve both high subject fidelity and multi-dimensional human preference alignment within a single training stage, relying on a single, indirect reconstruction loss, which is difficult to simultaneously satisfy both these goals. To address this, we propose MultiCrafter, a framework that decouples this task into two distinct training stages. First, in a pre-training stage, we introduce an explicit positional supervision mechanism that effectively resolves attention bleeding and drastically enhances subject fidelity. Second, in a post-training stage, we propose Identity-Preserving Preference Optimization, a novel online reinforcement learning framework. We feature a scoring mechanism to accurately assess multi-subject fidelity based on the Hungarian matching algorithm, which allows the model to optimize for aesthetics and prompt alignment while ensuring subject fidelity achieved in the first stage. Experiments validate that our decoupling framework significantly improves subject fidelity while aligning with human preferences better.

State-of-the-art text-to-image models demonstrate impressive realism but suffer from a persistent identity crisis when generating scenes with multiple humans: producing duplicate faces, merging identities, and miscounting individuals. We present DisCo, Reinforcement with DiverSity Constraints, a novel reinforcement learning framework that directly optimizes identity diversity both within images and across groups of generated samples. DisCo fine-tunes flow-matching models using Group-Relative Policy Optimization (GRPO), guided by a compositional reward that: (i) penalizes facial similarity within images, (ii) discourages identity repetition across samples, (iii) enforces accurate person counts, and (iv) preserves visual fidelity via human preference scores. A single-stage curriculum stabilizes training as prompt complexity increases, requiring no additional annotations. On the DiverseHumans Testset, DisCo achieves 98.6\% Unique Face Accuracy and near-perfect Global Identity Spread, outperforming both open-source and proprietary models (e.g., Gemini, GPT-Image) while maintaining competitive perceptual quality. Our results establish cross-sample diversity as a critical axis for resolving identity collapse in generative models, and position DisCo as a scalable, annotation-free solution for multi-human image synthesis.

The rapid growth of the text-to-image (T2I) community has fostered a thriving online ecosystem of expert models, which are variants of pretrained diffusion models specialized for diverse generative capabilities. Yet, existing model merging methods remain limited in fully leveraging abundant online expert resources and still struggle to meet diverse in-the-wild user needs. We present DiffGraph, a novel automated agent-driven graph-based model merging framework, which automatically harnesses online experts and flexibly merges them for diverse user needs. Our DiffGraph constructs a scalable graph and organizes ever-expanding online experts within it through node registration and calibration. Then, DiffGraph dynamically activates specific subgraphs based on user needs, enabling flexible combinations of different experts to achieve user-desired generation. Extensive experiments show the efficacy of our method.

Digital characters are central to modern media, yet generating character videos with long-duration, consistent multi-view appearance and expressive identity remains challenging. Existing approaches either provide insufficient context to preserve identity or leverage non-character-centric information as the "memory", leading to suboptimal consistency.Recognizing that character video generation inherently resembles an ``outside-looking-in" scenario. In this work, we propose represent the character’s visual attributes through a compact set of anchor frames.This design provides stable references for consistency, while reference-based video generation inherently faces challenges of copy-pasting and multi-reference conflicts. To address these, we introduce two mechanisms: Superset Content Anchoring, providing intra- and extra-training clip cues to prevent duplication, and RoPE as Weak Condition, encoding positional offsets to distinguish multiple anchors.Furthermore, we construct a scalable pipeline to extract these anchors from massive videos. Experiments show our method generates high-quality character videos exceeding $10$ minutes, and achieves expressive identity and appearance consistency across views, surpassing existing methods.

Storyboard synthesis plays a crucial role in visual storytelling, aiming to generate coherent shot sequences that visually narrate cinematic events with consistent characters, scenes, and transitions. However, existing approaches are mostly adapted from text-to-image diffusion models, which struggle to maintain long-range temporal coherence, consistent character identities, and narrative flow across multiple shots. In this paper, we introduce DreamShot, a video generative model based storyboard framework that fully exploits powerful video diffusion priors for controllable multi-shot synthesis. DreamShot supports both Text-to-Shot and Reference-to-Shot generation, as well as story continuation conditioned on previous frames, enabling flexible and context-aware storyboard generation. By leveraging the spatial-temporal consistency inherent in video generative models, DreamShot produces visually and semantically coherent sequences with improved narrative fidelity and character continuity. Furthermore, DreamShot incorporates a multi-reference role conditioning module that accepts multiple character reference images and enforces identity alignment via a Role-Attention Consistency Loss, explicitly constraining attention between reference and generated roles. Extensive experiments demonstrate that DreamShot achieves superior scene coherence, role consistency, and generation efficiency compared to state-of-the-art text-to-image storyboard models, establishing a new direction toward controllable video model-driven visual storytelling.

Text-to-video generation has significantly enriched content creation and holds the potential to evolve into powerful world simulators. However, modeling the vast spatiotemporal space remains computationally demanding, particularly when employing Transformers, which incur quadratic complexity in sequence processing and thus limit practical applications. Recent advancements in linear-time sequence modeling, particularly the Mamba architecture, offer a more efficient alternative. Nevertheless, its plain design limits its direct applicability to multimodal and spatiotemporal video generation tasks. To address these challenges, we introduce M4V, a multimodal Mamba framework for efficient text-to-video generation. Specifically, a MultiModal diffusion Mamba (MM-DiM) block is designed within the framework to enable seamless integration of multimodal information and spatiotemporal modeling. In detail, we introduce a novel multimodal token re-composition design, which employs a bidirectional scheme for multimodal information integration through simple token arrangement, along with visual registers to enhance spatial–temporal consistency. As a result, the MM-DiM blocks in M4V reduce FLOPs by 45% compared with the attention-based alternative when generating videos at 768$\times$1280 resolution. Additionally, several training strategies are explored in this work to provide a better understanding of training text-to-video models using only publicly available datasets. Extensive experiments on text-to-video benchmarks demonstrate M4V's ability to produce high-quality videos while significantly lowering computational costs. Code will be made publicly available.

Designing 3D metamaterial microstructures that meet the intended functions remains a major challenge, as it typically requires domain expertise, iterative simulations, and extensive manual tuning. Existing work on inverse design that automatically generates microstructures based on desired target properties often suffers from limited design diversity and faces challenges in ensuring the physical feasibility of the generated structures. To address this issue, a property-informed diffusion-based network is proposed that enables the generation of 3D microstructures directly from textual descriptions. Unlike traditional property conditioning methods, our approach leverages rich guidance in terms of semantics and physical properties in the text input to support diverse structure synthesis. To enforce consistency between the generated structures and the target textual prompts, a dual alignment strategy is adopted, including contrastive text-structure alignment and test-time reward-guided alignment. Experimental results show that the model is capable of generating semantically meaningful and physically plausible structures across a wide range of material categories. The proposed framework has good potential for interactive microstructure design and opens up new directions for combining language-based interfaces with inverse material discovery.

Current story visualization methods tend to position subjects solely by text and face challenges in maintaining artistic consistency. To address these limitations, we introduce DreamingComics, a layout-aware story visualization framework. We build upon a pretrained video diffusion-transformer (DiT) model, leveraging its spatiotemporal priors to enhance identity and style consistency. For layout-based position control, we propose RegionalRoPE, a region-aware positional encoding scheme that re-indexes embeddings based on the target layout. Additionally, we introduce a masked condition loss to further constrain each subject's visual features to their designated region. To infer layouts from natural language scripts, we integrate an LLM-based layout generator trained to produce comic-style layouts, enabling flexible and controllable layout conditioning. We present a comprehensive evaluation of our approach, showing a 29.2\% increase in character consistency and 36.2\% increase in style similarity compared to previous methods, while displaying high spatial accuracy.

We introduce MoS (Mixture of States), a novel fusion paradigm for multimodal diffusion models that merges modalities using flexible, state-based interactions. The core of MoS is a learnable, token-wise router that creates denoising timestep- and input-dependent interactions between modalities' hidden states, precisely aligning token-level features with the diffusion trajectory. This router sparsely selects the top-$k$ hidden states and is trained with an $\epsilon$-greedy strategy, efficiently selecting contextual features with minimal learnable parameters and negligible computational overhead. We validate our design with text-to-image generation (MoS-Image) and editing (MoS-Editing), which achieve state-of-the-art results. With only 3B to 5B parameters, our models match or surpass counterparts up to $4\times$ larger. These findings establish MoS as a flexible and compute-efficient paradigm for scaling multimodal diffusion models.

Recent advances in diffusion models have demonstrated impressive capability in generating high-quality images for simple prompts. However, when confronted with complex prompts involving multiple objects and hierarchical structures, existing models struggle to accurately follow instructions, leading to issues such as concept omission, confusion, and poor compositionality. To address these limitations, we propose a Hierarchical Compositional Generative framework (HiCoGen) built upon a novel Chain of Synthesis (CoS) paradigm. Instead of monolithic generation, HiCoGen first leverages a Large Language Model (LLM) to decompose complex prompts into minimal semantic units. It then synthesizes these units iteratively, where the image generated in each step provides crucial visual context for the next, ensuring all textual concepts are faithfully constructed into the final scene. To further optimize this process, we introduce a reinforcement learning (RL) framework. Crucially, we identify that the limited exploration of standard diffusion samplers hinders effective RL. We theoretically prove that sample diversity is maximized by concentrating stochasticity in the early generation stages and, based on this insight, propose a novel Decaying Stochasticity Schedule to enhance exploration. Our RL algorithm is then guided by a hierarchical reward mechanism that jointly evaluates the image at the global, subject, and relationship levels. We also construct HiCoPrompt, a new text-to-image benchmark with hierarchical prompts for rigorous evaluation. Experiments show our approach significantly outperforms existing methods in both concept coverage and compositional accuracy.

Despite the inherent advantages of thermal infrared(TIR) imaging, large-scale data collection and annotation remain a major bottleneck for TIR-based perception. A practical alternative is to synthesize pseudo TIR data via image translation; however, most RGB-to-TIR approaches heavily rely on RGB-centric priors that overlook thermal physics, yielding implausible heat distributions. In this paper, we introduce TherA, a controllable RGB-to-TIR translation framework that produces diverse and thermally plausible images at both scene and object level. TherA couples TherA-VLM with a latent-diffusion-based translator. Given a single RGB image and a user-prompted condition pair, TherA-VLM yields a thermal-aware embedding that encodes scene, object, material, and heat-emission context reflecting the input scene-condition pair. Conditioning the diffusion model on this embedding enables realistic TIR synthesis and fine-grained control across time of day, weather, and object state. Compared to other baselines, TherA achieves state-of-the-art translation performance, demonstrating improved zero-shot translation performance up to 33% increase averaged across all metrics.

We present SWIM (See What I Mean), a novel training strategy that aligns vision and language representations to enable fine-grained object understanding solely from textual prompts. Unlike existing approaches that require explicit visual prompts, such as masks or points, SWIM leverages mask supervision only during training to guide cross-modal attention, allowing the model to automatically attend to the user-specified object at inference.Our cross-attention analysis of pretrained multimodal large language models (MLLMs) reveals a systematic discrepancy: Attribute words produce sharp, localized activations in the visual modality, whereas object nouns yield diffuse and scattered patterns due to semantic reference bias and distributed high-level representations. To address this misalignment, we construct NL-Refer, an enriched dataset, in which each object mask is paired with a precise natural language referring expression. SWIM extracts multi-layer cross-attention maps from object nouns and enforces spatial consistency with ground-truth masks. Experimental results demonstrate that SWIM substantially improves text–visual alignment and achieves superior performance over visual-prompt-based methods on fine-grained object understanding benchmarks.

Cross-modal alignment aims to learn semantically consistent latent representations across diverse modalities. Prevailing methods rely on a text-guided aggregation paradigm to achieve fine-grained alignment, while they suffer from redundant patch-word correlations and high computational costs. To address these issues, we propose CoV-Align, an effective and efficient fine-grained cross-modal alignment framework with cohesive visual semantics priority. Through a semantically convergent attention mechanism, it progressively aggregates meaningful visual patches in a text-free manner. We design a coarse visual semantic feature extractor that integrates deformable attention and consist assign attention to group patches with semantic consistency. A cohesive and discriminative feature optimization is presented to enhance intra-semantic cohesion and inter-semantic discriminability of visual region features, resulting in explicit improvements in cross-modal alignment. Extensive experiments demonstrate that CoV-Align achieves state-of-the-art performance on the Flickr30K and MS-COCO benchmarks. Notably, it delivers a 3–5$\times$ computational speedup compared to pioneer approaches, offering compelling advantages for large-scale multi-modal tasks.

Tables are pervasive in diverse documents, making table recognition (TR) a fundamental task in document analysis. Existing modular TR pipelines separately model table structure and content, leading to suboptimal integration and complex workflows.End-to-end approaches rely heavily on large-scale TR data and struggle in data-constrained scenarios.To address these issues, we propose TDATR (Table Detail-Aware Table Recognition) improves end-to-end TR through table detail-aware learning and cell-level visual alignment.TDATR adopts a “perceive-then-fuse” strategy. The model first performs table detail-aware learning to jointly perceive table structure and content through multiple structure understanding and content recognition tasks designed under a language modeling paradigm. These tasks can naturally leverage document data from diverse scenarios to enhance model robustness.The model then integrates implicit table details to generate structured HTML outputs, enabling more efficient TR modeling when trained with limited data.Furthermore, we design a structure-guided cell localization module integrated into the end-to-end TR framework, which efficiently locates cell and strengthens vision–language alignment. It enhances the interpretability and accuracy of TR.We achieve state-of-the-art or highly competitive performance on seven benchmarks without dataset-specific fine-tuning.

Pre-trained vision encoders like DINOv2 have demonstrated exceptional performance on unimodal tasks. However, we observe that their feature representations are poorly aligned across different modalities. For instance, the feature embedding for an RGB image and its corresponding depth map of the same scene exhibit a cosine similarity that is nearly identical to that of two random, unrelated images. To address this, we propose the Omnivorous Vision Encoder, a novel framework that learns a modality-agnostic feature space. We train the encoder with a dual objective: first, to maximize the feature alignment between different modalities of the same scene; and second, a distillation objective that anchors the learned representations to the output of a fully frozen teacher such as DINOv2. The resulting student encoder becomes ``omnivorous'' by producing a consistent, powerful embedding for a given scene, regardless of the input modality (RGB, Depth, Segmentation, etc.). This approach enables robust cross-modal understanding while retaining the discriminative semantics of the original foundation model.

While Vision-Language Models (VLMs) have achieved remarkable performance, their Euclidean embeddings remain limited in capturing hierarchical relationships such as part-to-whole or parent-child structures, and often face challenges in multi-object compositional scenarios. Hyperbolic VLMs mitigate this issue by better preserving hierarchical structures and modeling part-whole relations (i.e., whole scene and its part images) through entailment. However, existing approaches do not model that each part has a different level of semantic representativeness to the whole. We propose UNcertainty-guided Compositional Hyperbolic Alignment (UNCHA) for enhancing hyperbolic VLMs. UNCHA models part-to-whole semantic representativeness with hyperbolic uncertainty, by assigning lower uncertainty to more representative parts and higher uncertainty to less representative ones for the whole scene. This representativeness is then incorporated into the contrastive objective with uncertainty-guided weights. Finally, the uncertainty is further calibrated with an entailment loss regularized with entropy-based term. With the proposed losses, UNCHA learns hyperbolic embeddings with more accurate part-whole ordering, capturing the underlying compositional structure in an image and improving its understanding of complex multi-object scenes. UNCHA achieves state-of-the-art performance on zero-shot classification, retrieval, and multi-label classification benchmarks.

Multi-task learning (MTL) involves the simultaneous optimization of multiple task-specific losses, often leading to gradient conflicts and scale imbalances that result in negative transfer. While existing multi-task optimization methods attempt to mitigate these challenges, they either lack the stochasticity needed to escape poor local minima or fail to explicitly address conflicts at the gradient level. In this work, we propose TaskForce, a novel multi-task optimization framework incorporating cooperative multi-agent reinforcement learning (MARL), where agents learn to find an effective joint optimization strategy based on their respective task gradients and losses. To keep the optimization process compact yet informative, agents observe a summary of the training dynamics that consists of the gradient Gram matrix---capturing both gradient magnitudes and pairwise alignments---and task loss values. Each agent then predicts the balancing parameters that determine the weight of their contribution to the final gradient update. Crucially, we design a hybrid reward function that incorporates both gradient-based signals and loss improvement dynamics, enabling agents to effectively resolve gradient conflicts and avoid poor convergence by considering both direct gradient information and the resulting impact on loss reduction. TaskForce achieves consistent improvements over state-of-the-art MTL baselines on NYU-v2, Cityscapes, and QM9, demonstrating the promise of cooperative MARL in complex multi-task scenarios.

With the rapid development of large multimodal models, reliable judge and critic models have become essential for open-ended evaluation and preference alignment, providing pairwise preferences, numerical scores, and explanatory justifications for assessing model-generated responses. However, existing critics are primarily trained in general visual domains such as captioning or image question answering, leaving physical AI tasks involving perception, causal reasoning, and planning largely underexplored. We introduce PhyCritic, a multimodal critic model optimized for physical AI through a two-stage RLVR pipeline: a physical skill warmup stage that enhances physically oriented perception and reasoning, followed by self-referential critic finetuning, where the critic generates its own prediction as an internal reference before judging candidate responses, improving judgment stability and physical correctness. Across both physical and general-purpose multimodal judge benchmarks, PhyCritic achieves strong performance gains over open-source baselines and, when applied as a policy model, further improves perception and reasoning in physically grounded tasks.

Multimodal Large Language Models (MLLMs) suffer from a fundamental "modality gap", contradicting themselves on visual versus text views of the same content. This paper argues that this inconsistency is not a failure, but a powerful resource for self-reward multimodal learning. Instead of relying on flawed voting mechanisms that amplify systematic errors when the majority is wrong, we introduce cross-modal cycle consistency as rewards C$^3$R to improve multimodal reasoning. C$^3$R performs backward inference from an answer to a query, switches modalities, and performs forward inference to verify the answer's consistency. This cycle serves as a dense, label-free reward that guides the model to resolve its own internal conflicts, while avoiding majority-is-wrong failures of standard voting methods. On standard benchmarks, C$^3$R mitigates modality-specific biases and improves reasoning accuracy by up to 7.6%. Our results show that robust reasoning emerges not just from scaling data, but from achieving a bidirectional understanding of the multimodal world.

Reward models (RMs) are essential for training large language models (LLMs), but remain underexplored for omni models that handle interleaved image and text sequences. We introduce Multimodal RewardBench 2 (MMRB2), the first comprehensive benchmark for reward models on multimodal understanding and (interleaved) generation. MMRB2 spans four tasks: text-to-image, image editing, interleaved generation, and multimodal reasoning (“thinking-with-images”), providing 1,000 expert-annotated preference pairs per task from 23 models and agents across 21 source tasks. MMRB2 is designed with: (1) practical but challenging prompts; (2) responses from state-of-the-art models, and agents; and (3) preference pairs with strong human-expert consensus, curated via an ensemble filtering strategy. Using MMRB2, we study existing judges for each subtask, including multimodal LLM-as-a-judge and models trained with human preferences. Top judges like GPT-5 and Gemini-2.5-Pro reach 66–75% accuracy, compared to >90% for humans, and outperform the commonly used GPT-4o (59% accuracy). The best performing open-source model Qwen3-VL-32B achieves similar accuracies as Gemini-2.5-Flash (64%). We also show that MMRB2 performance strongly correlates with downstream task success, and conduct an in-depth analysis that shows key areas to improve the reward models going forward.

Parallel test-time scaling typically trains separate generation and verification models, incurring high training and inference costs. We propose Advantage Decoupled Preference Optimization (\textbf{ADPO}), a unified reinforcement learning framework that jointly learns answer generation and self-verification within a single policy. ADPO introduces two innovations: \textbf{a preference verification reward} improving verification capability and \textbf{a decoupled optimization mechanism} enabling synergistic optimization of generation and verification. Specifically, the preference verification reward computes mean verification scores from positive and negative samples as decision thresholds, providing positive feedback when prediction correctness aligns with answer correctness. Meanwhile, the advantage decoupled optimization computes separate advantages for generation and verification, applies token masks to isolate gradients, and combines masked GRPO objectives, preserving generation quality while calibrating verification scores. ADPO achieves up to \textbf{+34.1\%} higher verification AUC and \textbf{-53.5\%} lower inference time, with significant gains of \textbf{+2.8\%/+1.4\%} accuracy on MathVista/MMMU, \textbf{+1.9} cIoU on ReasonSeg, and \textbf{+1.7\%/+1.0\%} step success rate on AndroidControl/GUI Odyssey.

Multimodal Large Reasoning Models (MLRMs) exhibit remarkable performance on complex tasks by incorporating explicit multi-step reasoning. However, this capability also introduces new security vulnerabilities. Existing jailbreak studies largely overlook Cognitive-level weaknesses embedded in the reasoning process itself. In this work, we uncover a critical cognitive bias in MLRMs: the anchoring effect, where safety judgments are disproportionately influenced by the first piece of information received—the anchor. Building on this finding, we propose the Reasoning-chain Anchoring Attack (RA-Attack), a novel jailbreak framework that fully exploits this vulnerability. RA-Attack employs a cross-modal safe anchor, whose core component is a structured visual mind map. This structured format provides the model with a pre-established, safety-biased reasoning chain that subtly induces it to rationalize and execute subsequent harmful intent. Extensive experiments across seven leading closed- and open-source MLRMs demonstrate the effectiveness of RA-Attack, achieving state-of-the-art jailbreak success rates—92% on Gemini-2.5-Pro and 82% on GPT-4o. Our findings reveal that cognitive biases can be systematically exploited to manipulate multimodal reasoning chains, establishing cognitive security as a critical and underexplored frontier in AI safety research. Warning: This paper contains unsafe examples.

Test-time prompt tuning (TPT) has emerged as a powerful technique for adapting pre-trained vision-language models (VLMs) to diverse downstream tasks, including image classification and visual reasoning. With the rise of text-driven object detectors, we extend TPT to object detection, unlocking new capabilities for cross-domain adaptation. However, a critical challenge in TPT is the inherent miscalibration caused by entropy minimization: domain shifts often lead to incorrect predictions, and enforcing high confidence exacerbates miscalibration, ultimately degrading performance. To tackle this, we introduce InsCal, a novel framework designed to enhance cross-domain object detection through three key innovations: (1) extending TPT to a multi-source paradigm, enabling knowledge aggregation across diverse domains; (2) reducing domain gaps via a novel text-driven style transfer strategy that aligns features to the source domain without requiring reference images; and (3) refining the entropy minimization objective with instance-specific calibration, ensuring robust and well-calibrated adaptation. Our approach not only mitigates miscalibration but also significantly improves cross-domain object detection performance, setting a new benchmark for test-time adaptation in VLMs.

Long-tailed recognition poses a significant challenge for deep learning. The two-stage decoupling paradigm, which separates representation learning from classifier retraining, offers a promising solution. During the classifier retraining stage, adaptive norm rescaling is a popular technique. It adjusts the per-class weight norms via parameter regularization, which inevitably introduces hyperparameters. However, many studies report that long-tailed recognition is sensitive to these hyperparameters, as their setup significantly impacts performance. In this paper, we first provide a class-conditional distribution perspective to support norm rescaling methods. Furthermore, we propose a simple but effective approach called Self-Adaptive Monotonic Normalization (SAMN). SAMN avoids the need for parameter regularization. It directly enforces monotonicity on per-class weight norms using the Pool Adjacent Violators Algorithm, making the method hyperparameter-friendly. SAMN is a universal strategy that integrates seamlessly with other methods for enhanced performance. Experiments on benchmark datasets demonstrate that our method significantly boosts long-tailed recognition performance, often achieving state-of-the-art results.

Large Vision–Language Models (LVLMs) use their vision encoders to translate images into representations for downstream reasoning, but the encoders often underperform in domain-specific visual tasks such as medical image diagnosis or fine-grained classification, where representation errors can cascade through the language model, leading to incorrect responses.Existing adaptation methods modify the continuous feature interface between encoder and language model through projector tuning or other parameter-efficient updates, which still couples the two components and requires re-alignment whenever the encoder changes.We introduce CRAFT (Codebook RegulAted Fine-Tuning), a lightweight method that fine-tunes the encoder using a discrete codebook that anchors visual representations to a stable token space, achieving domain adaptation without modifying other parts of the model.This decoupled design allows the adapted encoder to seamlessly boost the performance of LVLMs with different language architectures, as long as they share the same codebook. Empirically, CRAFT achieves an average gain of 14.98\% across 10 domain-specific benchmarks such as VQARAD and PlantVillage, while preserving the LLM’s linguistic capabilities and outperforming peer methods that operate on continuous tokens.

Deploying pretrained visual models in real-world environments often suffers from significant performance degradation due to the diversity of testing scenarios. Continuous adaptation of learning models on edge devices via unlabeled data collected from the target domain is highly effective for boosting generalization capability. However, gradient-backpropagation-based optimization of the massive parameters in deep neural networks is vastly more time-consuming than forward inference, rendering online learning infeasible on low-power edge devices. To address this critical challenge, we propose a lightweight gradient-free forward-memorizing framework, namely MemFlow, which leverages a frozen backbone and enables efficient fine-tuning of the mapping between features and predictions. Specifically, MemFlow employs randomly connected neurons to memorize feature-label associations; within the network, spiking signals are propagated, and predictions are generated by associating neuron-stored memories according to their confidence levels. More notably, MemFlow supports reinforced memorization of feature mappings using unlabeled data, thereby enabling rapid adaptation to new domains. Extensive experiments on four real-world cross-domain datasets demonstrate that MemFlow achieves performance improvements of up to 10% while consuming less than 1% of the computational time required by traditional domain adaptation methods.

Source-Free Cross-Domain Few-Shot Learning (SF-CDFSL) focuses on fine-tuning with limited training data from target domains (e.g., medical or satellite images), where Vision-Language Models (VLMs) such as CLIP and SigLIP have shown promising results. Current works in traditional visual models suggest that improving visual discriminability enhances performance. However, in VLM-based SF-CDFSL tasks, we find that \textbf{strengthening visual-modal discriminability actually suppresses VLMs’ performance}. In this paper, we aim to delve into this phenomenon for an interpretation and a solution.By both theoretical and experimental proofs, our study reveals that fine-tuning with the typical cross-entropy loss ($L_{vlm}$) inherently includes a visual learning part and a cross-modal learning part, where the cross-modal part is crucial for rectifying the heavily disrupted modality misalignment in SF-CDFSL.However, we find that the visual learning essentially acts as a shortcut that encourages the model to reduce $L_{vlm}$ without considering the cross-modal part, therefore hindering the cross-modal alignment and harming the performance.Based on this interpretation, we further propose an approach to address this problem: first, we perturb the visual learning to guide the model to focus on the cross-modal alignment. Then, we use the visual-text semantic relationships to gradually align the visual and textual modalities during the fine-tuning. Extensive experiments on various settings, backbones (CLIP, SigLip, PE-Core), and tasks (4 CDFSL datasets and 11 FSL datasets) show that we consistently set new state-of-the-art results. We will release the code.

Unsupervised domain adaptation transfers knowledge from a labeled source domain to an unlabeled target domain. When source data cannot be accessed, source-free domain adaptation (SFDA) becomes a practical alternative. However, existing SFDA methods mainly rely on pseudo-label based self-training, which often accumulates noise and bias under large domain gaps. We propose VSFOT, a framework that leverages a pretrained Vision-Language Model (VLM) to guide optimal transport (OT) alignment between target features and source prototypes. Instead of relying on unreliable pseudo-labels, VSFOT employs VLM-derived semantic priors and an OT-based matching strategy to achieve stable and reliable adaptation. To further enhance domain alignment, VSFOT incorporates a bidirectional distillation mechanism in which the model learns semantic consistency from the VLM, while the VLM is refined using task-specific cues from the model. These two stages alternate during training. By combining the generalization ability of the VLM with the discriminative power of the task model, VSFOT achieves robust, source-free adaptation and consistently outperforms existing SFDA methods on four benchmark datasets.

Unsupervised domain adaptation (UDA) for semantic segmentation seeks to transfer models from a labeled source domain to an unlabeled target domain. While auxiliary self-supervised tasks such as contrastive learning have enhanced feature discriminability, masked modeling remains underexplored due to architectural constraints and misaligned objectives. We propose Masked Representation Modeling (MRM), an auxiliary task that performs representation masking and reconstruction directly in the latent space. Unlike prior masked modeling methods that reconstruct low-level signals (e.g., pixels or visual tokens), MRM targets high-level semantic features, aligning its objective with segmentation and integrating seamlessly into standard architectures like DeepLab and DAFormer. To support efficient reconstruction, we design a lightweight auxiliary module, Rebuilder, which is jointly trained with the segmentation network but removed during inference, introducing zero test-time overhead. Extensive experiments demonstrate that MRM consistently improves segmentation performance across diverse architectures and UDA benchmarks. When integrated with four representative baselines, MRM achieves an average gain of +2.3 mIoU on GTA $\rightarrow$ Cityscapes and +2.8 mIoU on Cityscapes $\rightarrow$ Synthia, establishing it as a simple, effective, and generalizable strategy for unsupervised domain-adaptive semantic segmentation.

On-device AI systems increasingly adopt a single foundation model equipped with task-specific Low-Rank Adaptation (LoRA) modules, forming a multi-LoRA LLM that supports multiple tasks.We study how to adapt such a model to a new task on memory-constrainted devices.Although LoRA reduces trainable parameters, fine-tuning a full set of modules remains memory-intensive.To improve efficiency, we apply sparse updating, training a subset of LoRA modules within the memory budget.However, existing sparse updating methods assume all candidate parameters are instantiated and cannot estimate the importance of modules that do not yet exist, while prior memory models designed for sequential networks fail to capture the blockwise parallel structure of Transformers.We propose TaskIT, a framework for memory-efficient fine-tuning via cross-task importance transfer. TaskIT predicts pre-insertion module importance by transferring from previously tuned tasks and employs a block-based memory predictor that captures parallel and sequential dependencies of Transformer blocks. A dynamic programming scheduler then selects module locations, numbers, and ranks to maximize accuracy within the memory budget.Experiments on uni-modal and cross-modal benchmarks show that TaskIT achieves superior accuracy-memory tradeoffs compared with Zero-FT, non-LoRA, and LoRA-based fine-tuning methods.

Estimating dense three dimensional motion in dynamic high speed scenes remains challenging due to motion blur, illumination variation, and the limited temporal resolution of conventional cameras. We introduce ARES, a unified framework for Asymmetric RGB-Event Stereo that addresses these issues through a hybrid setup where an event camera captures fine grained temporal dynamics and an RGB camera provides rich spatial structure. To integrate these heterogeneous modalities, we propose Multimodal Contextual Attention, a transformer based fusion mechanism that attends to spatial and temporal contexts under cross view constraints and forms a unified correspondence space for disparity and optical flow estimation. Building on this shared representation, we introduce Temporal Disparity Posterior Fusion, a probabilistic framework that models the evolution of disparity posteriors to infer disparity change and recover metrically coherent scene flow. Trained with sparse supervision and dense self consistency cues, our ARES achieves geometrically consistent and temporally stable three dimensional motion estimation across diverse driving scenarios. Experiments show that ARES attains state of the art performance in scene flow estimation, establishing a principled path toward unified asymmetric multimodal stereo sensing. Our code will be released upon paper acceptance.

This paper proposes the first task for high-speed 3D point tracking using multi-view Event-RGB hybrid cameras. We design a cuboid observation device comprising 4 RGB cameras (30fps) and 2 Event cameras to synchronously capture high-speed motions, and propose MER-Tracker, a high–frame-rate 3D point-tracking network that fuses the complementary strengths of dual modalities. We first respectively extract 2D motion-change features from the RGB and Event modalities, then apply linear interpolation and anchor sampling to fuse the discrete RGB 3D features and continuous Event 3D features after 3D lifting, and finally employ a LoRA-tuned Transformer based on temporal correlationship to predict the high-frame-rate 3D point trajectories over fast motions, accomplishing high-speed 3D point tracking. To verify the effectiveness of our method, we construct both real-world and simulated high-speed motion datasets. Experiments on these datasets show that our method achieves accurate high-speed 3D point tracking at high-frame-rate (150fps), outperforming state-of-the-art methods.

Event cameras provide accurate information at motion boundaries—exactly where disentangling ego-motion, object motion, and border ownership determines segmentation quality. We argue that the missing ingredient in dynamic scene interpretation is moving border ownership: detecting motion boundaries and assigning which side is foreground so occlusions are resolved by design.Traditional geometric motion segmentation pipelines (e.g., flow clustering, simple motion models) remain assumption-heavy and slow, while deep models often fail to generalize across sensors or datasets. We introduce a lightweight, ownership-aware predictor trained solely on synthetic events with perfect supervision for boundaries, ownership, and motion, generated via a Blender pipeline. Its key targets—a signed-distance ownership field and a motion mask—focus learning where events occur and yield stable gradients. The model runs in real time and generalizes without tuning: trained on synthetic events, it achieves zero-shot transfer on EED, EVIMO1, EVIMO2, and EMSMC, delivering state-of-the-art performance. By casting motion segmentation as ownership-aware edge understanding, we combine the robustness of model-based reasoning with the scalability of learning.

Tracking any point (TAP) is a fundamental yet challenging task in computer vision, requiring high precision and long-term motion reasoning. Recent attempts to combine RGB frames and event streams have shown promise, yet they typically rely on synchronous or non-adaptive fusion, leading to temporal misalignment and severe degradation when one modality fails. We introduce TAPFormer, a transformer-based framework that performs asynchronous temporal-consistent fusion of frames and events for robust and high-frequency arbitrary point tracking.Our key innovation is a Transient Asynchronous Fusion (TAF) mechanism, which explicitly models the temporal evolution between discrete frames through continuous event updates—bridging the gap between low-rate frames and high-rate events. In addition, a Cross-modal Locally Weighted Fusion (CLWF) module adaptively adjusts spatial attention according to modality reliability, yielding stable and discriminative features even under blur or low light.To evaluate our approach under realistic conditions, we construct a novel real-world frame-event TAP dataset under diverse illumination and motion conditions.Our method outperforms existing point trackers, achieving a $28.2$% improvement in average pixel error within threshold. Moreover, on standard point tracking benchmarks, our tracker consistently achieves the best performance. We will release the code and dataset upon acceptance to support future research.

We propose EventHub, a novel framework for training deep-event stereo networks without ground truth annotations from costly active sensors, relying instead on standard color images. From these images, we derive either proxy annotations and proxy events through state-of-the-art novel view synthesis techniques, or simply proxy annotations when images are already paired with event data. Using the training set generated by our data factory, we repurpose state-of-the-art stereo models from RGB literature to process event data, obtaining new event stereo models with unprecedented generalization capabilities.Experiments on widely used event stereo datasets support the effectiveness of EventHub and show how the same data distillation mechanism can improve the accuracy of RGB stereo foundation models in challenging conditions such as nighttime scenes.

Event-based Action Recognition (EAR) provides a promising pathway for understanding dynamic behaviors under challenging conditions. Recent progress in vision-language models has introduced a cross-modal learning paradigm into EAR, enabling models to associate event streams with textual semantics for enhancing conceptual understanding. However, existing methods typically overlook the intrinsic polarity-driven motion cues that are fundamental to event data, leading to suboptimal spatiotemporal representations. To address this limitation, we propose a POlarity Knowledge Enhanced framework (POKER), which explicitly incorporates event polarity-aware motion knowledge across visual and textual modalities.POKER consists of two synergistic components: Polarity Motion Capturer (PMC) and Polarity Motion Reasoner (PMR). Specifically, PMC decouples positive and negative polarities to capture polarity-sensitive motion cues, while PMR semantically analyzes polarity-induced motion dynamics via large language models. Through the polarity alignment, POKER couples semantic reasoning with visual dynamics, achieving more discriminative representations. Extensive experiments on multiple benchmarks demonstrate that POKER enhances performance across diverse event representations.

Understanding the geometric and semantic structure of environments is essential for embodied agents. Existing semantic mapping methods trade off between explicit geometry and multi-scale semantics,and lack a native interface for large models, thus requiring additional training of feature projection for semantic alignment. To this end, we propose the multi-scale Gaussian-Language Map (GLMap), which introduces three key designs: (1) explicit geometry, (2) multi-scale semantics covering both instance and region level concepts, and (3) a dual-modality interface where each semantic unit jointly stores a natural language description and a 3D Gaussian representation. The 3D Gaussians enable compact storage and fast rendering of task-relevant images via Gaussian splatting. To enable efficient incremental construction, we further propose a Gaussian Estimator that analytically derives Gaussian parameters from dense point clouds without gradient-based optimization. Experiments on ObjectNav, InstNav, and SQA tasks show that GLMap effectively enhances target localization and contextual reasoning, while remaining compatible with large-model-based methods in a zero-shot manner.

An ideal embodied agent should possess lifelong learning capabilities to handle long-horizon and complex tasks, enabling continuous operation in general environments. This not only requires the agent to accurately accomplish given tasks but also to leverage long-term episodic memory to optimize decision-making. However, existing mainstream one-shot embodied tasks primarily focus on task completion results, neglecting the crucial process of exploration and memory utilization. To address this, we propose Long-term Memory Embodied Exploration (LMEE), which aims to unify the agent’s exploratory cognition and decision-making behaviors to promote lifelong learning. We further construct a corresponding dataset and benchmark, LMEE-Bench, incorporating multi-goal navigation and memory-based question answering to comprehensively evaluate both the process and outcome of embodied exploration. To enhance the agent’s memory recall and proactive exploration capabilities, we propose MemoryExplorer, a novel method that fine-tunes a multimodal large language model through reinforcement learning to encourage active memory querying. By incorporating a multi-task reward function that includes action prediction, frontier selection, and question answering, our model achieves proactive exploration. Extensive experiments against state-of-the-art embodied exploration models demonstrate that our approach achieves significant advantages in long-horizon embodied tasks.

Vision Language Models (VLMs) demonstrate strong qualitative visual understanding, but struggle with metrically precise spatial reasoning required for embodied applications. The agentic paradigm promises that VLMs can use a wide variety of tools that could augment these capabilities, such as depth estimators, segmentation models, and pose estimators. Yet it remains an open challenge how to realize this vision without solely relying on handcrafted prompting strategies or enforcing fixed, predefined tool pipelines that limit VLMs' ability to discover optimal tool-use patterns. Reinforcement Learning could overcome this gap, but has so far been limited to reasoning with a single visual tool due to the large search space in multi-tool reasoning. We introduce Double Interactive Reinforcement Learning (DIRL), a two-phase training framework where VLMs learn to coordinate multiple tools through interactive exploration and feedback. In the teaching phase, we combine demonstrations from a single tool specialist trained via interactive RL with traces from a frontier model using all tools. In the exploration phase, the model further refines multi-tool coordination through continued RL. Our model, SpaceTools, with tool-augmented spatial reasoning ability, achieves state-of-the-art performance on spatial understanding benchmarks (RoboSpatial-Home, BLINK, Internal Benchmark) and demonstrates reliable real-world manipulation using a 7-DOF robot as a tool. DIRL provides substantial improvements over the vanilla SFT (+12% on RoboSpatial) and RL (+16% on RoboSpatial) baselines.

Physics-based humanoid control has achieved remarkable progress in enabling realistic and high-performing single-agent behaviors, yet extending these capabilities to cooperative human-object interaction (HOI) remains challenging. We present TeamHOI, a framework that enables a single decentralized policy to handle cooperative HOIs across any number of cooperating agents. Each agent operates using local observations while attending to other teammates through a Transformer-based policy network with teammate tokens, allowing scalable coordination across variable team sizes. To enforce motion realism while addressing the scarcity of cooperative HOI data, we further introduce a masked Adversarial Motion Prior (AMP) strategy that uses single-human reference motions while masking object-interacting body parts during training. The masked regions are then guided through task rewards to produce diverse and physically plausible cooperative behaviors. We evaluate TeamHOI on a challenging cooperative carrying task involving two to eight humanoid agents and varied object geometries. Finally, to promote stable carrying, we design a team-size- and shape-agnostic formation reward. TeamHOI achieves high success rates and demonstrates coherent cooperation across diverse configurations with a single policy.

Active 3D reconstruction enables an agent to autonomously select viewpoints to build accurate and complete scene geometry efficiently, rather than passively reconstructing scenes from pre-collected images. Existing active reconstruction methods often rely on geometric heuristics, which may result in redundant observations without improving reconstruction quality. To address this, we propose \textbf{AREA3D}, an active reconstruction agent for 3D reconstruction by leveraging feed-forward 3D models and vision-language guidance. The framework decouples view uncertainty modeling from feed-forward reconstruction, enabling precise uncertainty estimation without online optimization. Moreover, the integrated Vision-Language Model provides high-level semantic guidance that guides exploration beyond purely geometric cues. Extensive experiments on both scene-level and object-level benchmarks (Replica and OmniObject3D) demonstrate that AREA3D achieves state-of-the-art reconstruction accuracy, especially in sparse views.

Multimodal agents, such as JARVIS-1, are rapidly advancing in open-world environments. Their core workflow typically follows a perception–reasoning–action–memory cycle. Existing studies primarily emphasize improving memory representations and storage formats, treating memory mainly as an information repository. However, distilling transferable knowledge from stored experiences remains an important yet underexplored challenge.In real-world settings, structures and patterns tend to recur. If an agent can capture and reuse these latent patterns, it can infer new actionable knowledge from prior experience, enabling more efficient and flexible task execution. To explore this capability, we propose Echo. Echo decomposes knowledge into five explicit dimensions of transferability: structure, attribute, process, function, and interaction. Based on this formulation, Echo leverages In-Context Analogy Learning (ICAL) to effectively retrieve past experiences and generalize them to new tasks.Experiments show that, under a from-scratch learning setting, Echo achieves a 1.3×–1.7× speed-up in object-unlocking tasks. Moreover, Echo exhibits a burst-like chain-unlocking phenomenon, rapidly unlocking multiple similar items within a short time interval. These results demonstrate that robust knowledge transfer, driven by effective utilization of contextual examples, is a highly promising direction for advancing open-world multimodal agents.

Embodied navigation is a fundamental capability for robotic agents operating. Real-world deployment requires open vocabulary generalization and low training overhead, motivating zero-shot methods rather than task-specific RL training. However, existing zero-shot methods that build explicit 3D scene graphs often compress rich visual observations into text-only relations, leading to high construction cost, irreversible loss of visual evidence, and constrained vocabularies.To address these limitations, we introduce the Multi-modal 3D Scene Graph (M3DSG), which preserves visual cues by replacing textual relational edges with dynamically assigned images. Built on M3DSG, we propose MSGNav, a zero-shot navigation system that includes a Key Subgraph Selection module for efficient reasoning, an Adaptive Vocabulary Update module for open vocabulary support, and a Closed-Loop Reasoning module for accurate exploration reasoning. Additionally, we further identify the “last mile” problem in zero-shot navigation — determining the feasible target location with a suitable final viewpoint, and propose a Visibility-based Viewpoint Decision module to explicitly resolve it. Comprehensive experimental results demonstrate that MSGNav achieves state-of-the-art performance on GOAT-Bench and HM3D-OVON datasets. The open-source code will be publicly available.

Active perception and manipulation are crucial for embodied robots to interact with complex scenes. Existing methods struggle to unify semantic-driven perception actively with robust, viewpoint-invariant execution accordingly. To this end, we propose SaPaVe, an end-to-end framework that jointly learns these capabilities in a data-efficient manner. Central to our approach is a decoupling of camera and manipulation actions, contrary to shared-action-space, and learning in a bottom-up strategy: we first train semantic camera control on our proposed large-scale dataset, then jointly optimize both action types via hybrid data. To support this learning, we introduce ActiveViewPose-200K, comprising 200k image-language-camera movement pairs for semantic camera movement learning, and a 3D geometry-aware module that improves execution robustness under dynamic viewpoints. We further present ActiveManip-Bench, the first benchmark filling the gap to evaluate active manipulation. Extensive experiments in both simulation and real-world settings show that SaPaVe outperforms recent VLA models such as GR00T and $\pi_0$, achieving up to 31.25\% higher success rates in real-world tasks. Our results show that tightly coupled perception and execution, when trained with decoupled yet coordinated strategies, enable efficient and generalizable active manipulation.

Real-world robotic tasks are long-horizon and often span multiple floors, requiring complex spatial reasoning. Existing embodied benchmarks, however, are largely confined to single-floor homes, failing to evaluate agents on realistic, building-scale tasks. We introduce MANSION, a language-driven framework for generating building-scale, multi-floor 3D environments for long-horizon tasks. Using this framework, we release MansionWorld, a large-scale dataset featuring over 1,000 diverse, non-residential buildings. These environments support cross-floor skills and long-horizon task generation on reusable building layouts. Experiments show that current methods degrade sharply on our multi-floor tasks, highlighting both the challenge and the value of this setting for advancing embodied AI.

Existing appliance assets suffer from poor rendering, incomplete mechanisms, and misalignment with manuals, leading to simulation-reality gaps that hinder appliance manipulation development. In this work, we introduce the RealAppliance dataset, comprising 100 high-fidelity appliances with complete physical, electronic mechanisms, and program logic aligned with their manuals. Based on these assets, we propose the RealAppliance-Bench benchmark, which evaluates multimodal large language models and embodied manipulation planning models across key tasks in appliance manipulation planning: manual page retrieval, appliance part grounding, open-loop manipulation planning, and closed-loop planning adjustment. Our analysis of model performances on RealAppliance-Bench provides insights for advancing appliance manipulation research

Vision-Language-Action (VLA) models convert abstract language instructions into concrete, executable actions, a task that is especially challenging in open-world environments. We present \textit{Visually Grounded Planning}, a general and efficient high-level planner that guides a VLA step-by-step using imagined future observations and subtask descriptions. With an imagined future observation, the VLA can focus on visuomotor inference rather than high-level semantic reasoning, leading to improved accuracy and generalization. Our planner comprises a highly efficient foresight image-generation module that predicts a high-quality 640×480 future observation from the current visual input and language instruction within only 0.33 s on an H100 GPU, together with a vision–language component that reasons over the task and produces subtask descriptions for both the generator and the VLA. Importantly, state-of-the-art VLAs can integrate our planner seamlessly by simply augmenting their visual inputs, without any architectural modification. The foresight generator is pretrained on approximately 10 million multi-task, cross-embodiment samples, enabling it to learn robust embodied dynamics and achieve strong real-world generalization. We evaluate our framework on a benchmark consists of 11 diverse, multi-step real-world tasks. It achieves an average success rate of 87.4\%, demonstrating a +40.9\% absolute improvement over the $\pi_0$ baseline (46.5\%) and a +30.3\% absolute improvement over $\pi_0$ augmented with textual subtask guidance (57.1\%).

Vision-Language-Action (VLA) models have shown great performance in robotic manipulation by mapping visual observations and language instructions directly to actions. However, they remain brittle under distribution shifts: when test scenarios change, VLAs often reproduce memorized trajectories instead of adapting to the updated scene, which is a failure mode we refer to as the ``Memory Trap''. This limitation stems from the end-to-end design, which lacks explicit 3D spatial reasoning and prevents reliable identification of actionable regions in unfamiliar environments. To compensate for this missing spatial understanding, 3D Spatial Affordance Fields (SAFs) can provide a geometric representation that highlights where interactions are physically feasible, offering explicit cues about regions the robot should approach or avoid. We therefore introduce Affordance Field Intervention (AFI), a lightweight hybrid framework that uses SAFs as an on-demand plug-in to guide VLA behavior. Our system detects memory traps through proprioception, repositions the robot to recent high-affordance regions, and proposes affordance-driven waypoints that anchor VLA-generated actions. A SAF-based scorer then selects trajectories with the highest cumulative affordance. Extensive experiments demonstrate that our method achieves an average improvement of 23.5\% across different VLA backbones ($\pi_{0}$ and $\pi_{0.5}$) under out-of-distribution scenarios on real-world robotic platforms, and 20.2\% on the LIBERO-Pro benchmark, validating its effectiveness in enhancing VLA robustness to distribution shifts.

RAW images captured by different camera sensors exhibit substantial domain shifts due to varying spectral responses, noise characteristics, and tone behaviors, complicating their direct use in downstream vision tasks. Prior methods address this by training one-to-one RAW-to-RAW translators for each source-target domain pair, but such approaches do not scale to real-world scenarios with multiple cameras. We introduce MERIT, the first unified framework for multi-domain RAW image translation, which leverages a single model to perform translations across arbitrary camera domains. To address domain-specific noise discrepancies, we propose a sensor-aware noise modeling loss that explicitly aligns the signal-dependent noise statistics of the generated images with those of the target domain. Additionally, we enhance the generator’s context modeling with a conditional multi-scale large kernel attention module, enabling efficient capture of both global illumination and fine-grained sensor cues. To support standardized evaluation, we construct MDRAW, a new dataset of paired and unpaired RAW images from five diverse camera sensors. Extensive experiments on existing and newly proposed benchmarks demonstrate that MERIT significantly outperforms prior models in both accuracy and scalability, offering a practical and generalizable solution to cross-domain RAW image harmonization.

Recent advances in multimodal models have demonstrated remarkable text-guided image editing capabilities, with systems like GPT-4o and Nano-Banana setting new benchmarks. However, the research community's progress remains constrained by the absence of large-scale, high-quality, and openly accessible datasets built from real images. We introduce Pico-Banana-400K, a comprehensive 400K-image dataset for instruction-based image editing. Our dataset is constructed by leveraging Nano-Banana to generate diverse edit pairs from real photographs in the OpenImages collection. What distinguishes Pico-Banana-400K from previous synthetic datasets is our systematic approach to quality and diversity. We employ a fine-grained image editing taxonomy to ensure comprehensive coverage of edit types while maintaining precise content preservation and instruction faithfulness through MLLM-based quality scoring and careful curation. Beyond single turn editing, Pico-Banana-400K enables research into complex editing scenarios. The dataset includes three specialized subsets: (1) a 72K-example multi-turn collection for studying sequential editing, reasoning, and planning across consecutive modifications; (2) a 56K-example preference subset for alignment research and reward model training; and (3) paired long-short editing instructions for developing instruction rewriting and summarization capabilities. By providing this large-scale, high-quality, and task-rich resource, Pico-Banana-400K establishes a robust foundation for training and benchmarking the next generation of text-guided image editing models.

Recent advances in vision–language foundation models have enabled text-driven evaluation of image aesthetics and visual quality. However, existing models are typically optimized for fixed prompts or specific datasets, limiting their adaptability to diverse evaluation criteria. This paper presents \textit{Probabilistic Prompt Adaptation (PPA)}, a unified probabilistic framework that flexibly predicts aesthetic and quality scores conditioned on arbitrary text prompts. PPA formulates score prediction as a mixture over prompts, dynamically estimating prompt suitability based on both image content and task context. By marginalizing over prompts pre-sampled from a large language model (LLM), it enables annotation-free training using only triplets of task, image, and score. Experiments across multiple IAA and IQA benchmarks demonstrate that PPA achieves consistent and perceptually aligned prompt-based scoring, allowing fine-grained control over evaluation semantics.

The widespread adoption of text-to-image (T2I) generation has raised concerns about privacy, bias, and copyright violations. Concept erasure techniques offer a promising solution by selectively removing undesired concepts from pre-trained models without requiring full retraining. However, these methods are often evaluated on a limited set of concepts, relying on overly simplistic and direct prompts. To test the boundaries of concept erasure techniques, and assess whether they truly remove targeted concepts from model representations, we introduce EMMA, a benchmark that evaluates five key dimensions of concept erasure over 12 metrics. EMMA goes beyond standard metrics like image quality and time efficiency, testing robustness under challenging conditions, including indirect descriptions, visually similar non-target concepts, and potential gender and ethnicity bias, providing a socially aware analysis of method behavior. Using EMMA, we analyze five concept erasure methods across five domains (objects, celebrities, art styles, NSFW, and copyright).Our results show that existing methods struggle with implicit prompts (i.e. generating the erased concept when it is indirectly referenced) and visually similar non-target concepts (i.e. failing to generated non-targeted concepts resembling the erased one), while some amplify gender and ethnicity bias compared to the original model.

Recent advances in text-to-image (T2I) generation have greatly improved visual quality, yet producing images that appear visually authentic to real-world photography remains challenging. This is partly due to biases in existing evaluation paradigms: human ratings and preference-trained metrics often favor visually vivid images with exaggerated saturation and contrast, which make generations often $\textit{too vivid to be real}$ even when prompted for realistic-style images.To address this issue, we present $\textbf{Color Fidelity Dataset (CFD)}$ and $\textbf{Color Fidelity Metric (CFM)}$ for objective evaluation of color fidelity in realistic-style generations. CFD contains over 1.3M real and synthetic images with ordered levels of color realism, while CFM employs a multimodal encoder to learn perceptual color fidelity. In addition, we propose a training-free $\textbf{Color Fidelity Refinement (CFR)}$ that adaptively modulates spatial–temporal guidance scale in generation, thereby enhancing color authenticity.Together, CFD supports CFM for assessment, whose learned attention further guides CFR to refine T2I fidelity, forming a progressive framework for assessing and improving color fidelity in realistic-style T2I generation. All datasets and code will be publicly released.

Recent image editing models boast next-level intelligent capabilities, facilitating cognition- and creativity-informed image editing. Yet, existing benchmarks provide too narrow a scope for evaluation, failing to holistically assess these advanced abilities. To address this, we introduce WiseEdit, a knowledge-intensive benchmark for comprehensive evaluation of cognition- and creativity-informed image editing, featuring deep task depth and broad knowledge breadth. Drawing an analogy to human cognitive creation, WiseEdit decomposes image editing into three cascaded steps—Awareness, Interpretation, and Imagination—each corresponding to a task that poses a challenge for models to complete at the specific step. It also encompasses complex tasks, where none of the three steps can be finished easily. Furthermore, WiseEdit incorporates three fundamental types of knowledge: Declarative, Procedural, and Metacognitive knowledge. Ultimately, WiseEdit comprises 1,220 test cases, objectively revealing the limitations of SoTA image editing models in knowledge-based cognitive reasoning and creative composition capabilities.

With the rapid advances of powerful multimodal models such as GPT-4o, Nano Banana, and Seedream 4.0 in Image Editing, the performance gap between closed-source and open-source models is widening, primarily due to the scarcity of large-scale, high-quality training data and comprehensive benchmarks capable of diagnosing model weaknesses across diverse editing behaviors. Existing data construction methods face a scale-quality trade-off: human annotations are high-quality but not scalable, while automated pipelines suffer from error propagation and noise. To address this, we introduce a lightweight data pipeline that replaces multi-toolchains with an end-to-end model and a unified post-verification stage. For scalable quality control, we train a 7B dual-task expert model, \textbf{Qwen-Verify}, for efficient failure detection and instruction recaptioning. This pipeline yields \textbf{UnicEdit-10M}, a 10M-scale dataset spanning diverse basic and complex editing tasks. We also propose \textbf{UnicBench}, a general benchmark that extends beyond basic edits to explicitly assess spatial and knowledge-driven reasoning. To enable fine-grained diagnosis, we introduce novel metrics, including \textit{Non-edit Consistency} and \textit{Reasoning Accuracy}. Our analysis of mainstream models on UnicBench reveals their limitations and provides clear directions for future research. The dataset, benchmark, and code will be released.

Precise and controllable image editing remains a significant challenge. Current methods often rely on text prompts, but achieving accurate spatial localization solely through descriptions is inherently difficult. Mask-based approaches, though offering better control, typically require overly precise user annotations, thus increasing user burden and leading to unnatural results. To bridge this gap, we introduce the **I**nteractive **I**nstruction-based **I**mage **E**diting ($I_3E$) task, which generates high-quality edits from a more intuitive combination: concise text instructions and imprecise spatial guidance. To address the critical lack of suitable data, we propose an efficient pipeline to generate Inter-Edit, a new million-scale training dataset that simulates realistic user masks---not strictly segment-aligned. We also present a comprehensive benchmark, featuring a meticulously human-annotated test set that captures diverse, localization-dependent editing scenarios and realistic user interaction patterns. To evaluate this task, we introduce a new suite of position-aware metrics that strongly correlate with human perceptual judgments. Finally, we develop three baseline models trained on Inter-Edit. Extensive experiments demonstrate that our methods significantly enhance $I_3E$ performance, achieving substantial improvements in localization and edit quality, and outperforming existing state-of-the-art models. The Inter-Edit dataset and all related code will be made publicly available.

Diffusion models are promising for sparse-view novel view synthesis (NVS), as they can generate pseudo-ground-truth views to aid 3D reconstruction pipelines like 3D Gaussian Splatting (3DGS). However, these synthesized images often contain photometric and geometric inconsistencies, and their direct use for supervision can impair reconstruction. To address this, we propose Partial-Reference Image Quality Assessment (PR-IQA), a framework that evaluates diffusion-generated views using reference images from different poses, eliminating the need for ground truth. PR-IQA first computes a geometrically consistent partial quality map in overlapping regions. It then performs quality completion to inpaint this partial map into a dense, full-image map. This completion is achieved via a cross-attention mechanism that incorporates reference-view context, ensuring cross-view consistency and enabling thorough quality assessment. When integrated into a diffusion-augmented 3DGS pipeline, PR-IQA restricts supervision to high-confidence regions identified by its quality maps. Experiments demonstrate that PR-IQA outperforms existing IQA methods, achieving full-reference-level accuracy without ground-truth supervision. Thus, our quality-aware 3DGS approach more effectively filters inconsistencies, producing superior 3D reconstructions and NVS results.

In 3D reconstruction, the problem of inverse rendering, namely recovering the illumination of the scene and the material properties, is fundamental. Existing Gaussian Splatting-based methods primarily target static scenes and often assume simplified or moderate lighting to avoid entangling shadows with surface appearance. This limits their ability to accurately separate lighting effects from material properties, particularly in real-world conditions. We address this limitation by leveraging dynamic elements - regions of the scene that undergo motion - as a supervisory signal for inverse rendering. Motion reveals the same surfaces under varying lighting conditions, providing stronger cues for disentangling material and illumination. This thesis is supported by our experimental results which show we improve LPIPS by 23\% for albedo estimation and by 15% for scene relighting relative to next-best baseline. To this end, we introduce LumiMotion, the first Gaussian-based approach that leverages dynamics for inverse rendering and operates in arbitrary dynamic scenes. Our method learns a dynamic 2D Gaussian Splatting representation that employs a set of novel constraints which encourage the dynamic regions of the scene to deform, while keeping static regions stable. As we demonstrate, this separation is crucial for correct optimization of the albedo. Finally, we release a new synthetic benchmark comprising five scenes under four lighting conditions, each in both static and dynamic variants, for the first time enabling systematic evaluation of inverse rendering methods in dynamic environments and challenging lighting.

3D Gaussian Splatting has recently enabled fast and photorealistic reconstruction of static 3D scenes. However, dynamic editing of such scenes remains a significant challenge. We introduce a novel framework, Physics-Guided Score Distillation, to address a fundamental conflict: physics simulation provides a strong motion prior that is insufficient for photorealism , while video-based Score Distillation Sampling (SDS) alone cannot generate coherent motion for complex, multi-particle scenarios. We resolve this through a unified optimization framework where physics simulation guides Score Distillation to jointly refine the motion prior for photorealism while simultaneously optimizing appearance. Specifically, we learn a neural dynamics model that predicts particle motion and appearance, optimized end-to-end via a combined loss integrating Video-SDS for photorealism with our physics-guidance prior. This allows for photorealistic refinements while ensuring the dynamics remain plausible. Our framework enables scene-wide dynamic weather effects, including snowfall, rainfall, fog, and sandstorms, with physically plausible motion. Experiments demonstrate our physics-guided approach significantly outperforms baselines, with ablations confirming this joint refinement is essential for generating coherent, high-fidelity dynamics.

Feed-forward 3D modeling has emerged as a promising approach for rapid and high-quality 3D reconstruction. In particular, directly generating explicit 3D representations, such as 3D Gaussian splatting, has attracted significant attention due to its fast and high-quality rendering. However, many state-of-the-art methods, primarily based on transformer architectures, suffer from severe scalability issues because they rely on full attention across image tokens from multiple input views, resulting in prohibitive computational costs as the number of views or image resolution increases. Toward a scalable and efficient feed-forward 3D reconstruction, we introduce an iterative Large 3D Reconstruction Model (iLRM) that generates 3D Gaussian representations through an iterative refinement mechanism, guided by three core principles: (1) decoupling the scene representation from input images to enable compact 3D representations; (2) decomposing global multi-view interactions into a two-stage attention scheme to reduce computational costs; and (3) injecting high-resolution information at every layer to achieve high-fidelity reconstruction. Experimental results on widely used datasets, such as RE10K and DL3DV, demonstrate that iLRM outperforms existing methods in both reconstruction quality and speed.

Multi-view inverse rendering aims to recover geometry, materials, and illumination consistently across multiple viewpoints. Existing single-view approaches often ignore cross-view relationships, leading to inconsistent results, while multi-view optimization methods rely on slow differentiable rendering and per-scene refinement, making them computationally expensive and hard to scale. To address these limitations, we introduce a feed-forward multi-view inverse rendering framework that directly predicts spatially varying albedo, metallicity, roughness, diffuse shading, and surface normals from sequences of RGB images. By alternating attention across views, our model captures both intra-view long-range lighting interactions and inter-view material consistency, enabling coherent scene-level reasoning within a single forward pass. Due to the scarcity of real-world training data, models trained on existing synthetic datasets often struggle to generalize to real-world scenes. To overcome this limitation, we propose a consistency-based finetuning strategy that leverages unlabeled real-world videos to enhance both multi-view coherence and robustness under in-the-wild conditions. Extensive experiments on benchmark datasets demonstrate that our method achieves state-of-the-art performance in terms of multi-view consistency, material and normal estimation quality, and generalization to real-world imagery.

In this paper, we introduce NAS3R, a self-supervised feed-forward framework that jointly learns explicit 3D geometry and camera parameters with no ground-truth annotations and no pretrained priors.Given uncalibrated and unposed multi-view images, NAS3R reconstructs 3D Gaussian primitives from context views and renders target views using its self-predicted camera parameters, enabling self-supervised training from 2D photometric supervision.To ensure stable convergence, NAS3R integrates scene reconstruction and camera estimation within a shared transformer backbone regulated by masked attention, and adopts a depth-based Gaussian formulation that facilitates well-conditioned optimization.The framework is compatible with state-of-the-art architectures and can incorporate pretrained priors or intrinsic information when available.Extensive experiments show that NAS3R achieves superior results to other self-supervised methods, establishing a scalable and geometry-aware paradigm for 3D learning from unconstrained data.

Recent advances in 4D Gaussian Splatting (4DGS) have extended the high-speed rendering capability of 3D Gaussian Splatting (3DGS) into the temporal domain, enabling real-time rendering of dynamic scenes.However, one of the major remaining challenges lies in modeling long-range motion-contained dynamic videos, where a naïve extension of existing methods leads to severe memory explosion, temporal flickering, and failure to handle appearing or disappearing occlusions over time. To address these challenges, we propose a novel 4DGS framework characterized by an Anchor Relay-based Bidirectional Blending (ARBB) mechanism, named MoRel, which enables temporally consistent and memory-efficient modeling of long-range dynamic scenes.Our method progressively constructs locally canonical anchor spaces at key-frame time index and models inter-frame deformations at the anchor level, enhancing temporal coherence. By learning bidirectional deformations between KfA and adaptively blending them through learnable opacity control, our approach mitigates temporal discontinuities and flickering artifacts.We further introduce a Feature-variance-guided Hierarchical Densification (FHD) scheme that effectively densifies KfA's while keeping rendering quality, based on an assigned level of feature-variance.To effectively evaluate our model's capability to handle real-world long-range 4D motion, we newly compose long-range 4D motion-contained dataset, called SelfCap$_{\text{LR}}$. It has larger average dynamic motion magnitude, captured at spatially wider spaces, compared to previous dynamic video datasets.Overall, our MoRel achieves temporally coherent and flicker-free long-range 4D reconstruction while maintaining bounded memory usage, demonstrating both scalability and efficiency in dynamic Gaussian-based representations. The code and project page will be publicly released.

We propose Multi-view Pyramid Transformer (MVP), a scalable multi-view transformer architecture that directly reconstructs large 3D scenes from tens to hundreds of images in a single forward pass. Drawing on the idea of ``looking broader to see the whole, looking finer to see the details," MVP is built on two core design principles: 1) a local-to-global inter-view hierarchy that gradually broadens the model's perspective from local views to groups and ultimately the full scene, and 2) a fine-to-coarse intra-view hierarchy that starts from detailed spatial representations and progressively aggregates them into compact, information-dense tokens. This dual hierarchy achieves both computational efficiency and representational richness, enabling fast reconstruction of large and complex scenes. We validate MVP on diverse datasets and show that, when coupled with 3D Gaussian Splatting as the underlying 3D representation, it achieves state-of-the-art generalizable reconstruction quality while maintaining high efficiency and scalability across a wide range of view configurations.

Recent breakthroughs in 3D Gaussian Splatting (3DGS) have advanced neural rendering with high fidelity and speed. However, its performance degrades significantly in large-scale scenes due to the computational burden of tile-based rasterization. Existing optimization efforts either require costly scene re-training or focus on narrow aspects of the pipeline, overlooking critical inefficiencies in real-world deployments. Through a comprehensive analysis, we identify three primary sources of redundancy and low GPU utilization: redundant inter-frame pre-processing, viewpoint-based occlusion redundancy, and severe tile-level load imbalance. To address these issues, we propose CaT-GS, a novel and efficient 3DGS rendering pipeline. CaT-GS introduces a speculative multi-frame preprocessing method to eliminate redundant computations across consecutive frames, and an inter-frame caching mechanism to eliminate viewpoint redundant rendering stages. Furthermore, it refactors rasterization tasks with a dedicated kernel to mitigate tile load imbalance, significantly boosting GPU utilization. Extensive experiments demonstrate that CaT-GS achieves a speedup of up to 10\times over the original 3DGS and up to 70\% over previous state-of-the-art methods, establishing a new benchmark for high-fidelity, real-time rendering of large-scale scenes.

Blind 360° image quality assessment (IQA) aims to predict perceptual quality for panoramic images without a pristine reference. Unlike conventional planar images, 360° content in immersive environments restricts viewers to a limited viewport at any moment, making viewing behaviors critical to quality perception. Although existing scanpath-based approaches have attempted to model viewing behaviors by approximating the human view‑then‑rate paradigm, they treat scanpath generation and quality assessment as separate steps, preventing end-to-end optimization and task-aligned exploration. To address this limitation, we propose RL‑ScanIQA, a reinforcement‑learned framework for blind 360° IQA. RL-ScanIQA optimize a PPO-trained scanpath policy and a quality assessor, where the policy receives quality-driven feedback to learn task-relevant viewing strategies. To improve training stability and prevent mode collapse, we design multi-level rewards, including scanpath diversity and equator-biased priors. We further boost cross‑dataset robustness using distortion‑space augmentation together with rank‑consistent losses that preserve intra‑image and inter‑image quality orderings. Extensive experiments on three benchmarks show that RL‑ScanIQA achieves superior in‑dataset performance and cross‑dataset generalization. Code will be released upon publication.

In endoscopic surgery, a clear and high-quality visual field is critical for surgeons to make accurate intraoperative decisions. However, persistent visual degradation, including smoke generated by energy devices, lens fogging from thermal gradients, and lens contamination due to blood or tissue fluid splashes during surgical procedures, severely impairs visual clarity. These degenerations can seriously hinder surgical workflow and pose risks to patient safety. To systematically investigate and address various forms of surgical scene degradation, we introduce a real-world open-source surgical image restoration dataset covering laparoscopic environments, called SurgClean, which involves multi-type image restoration tasks from two medical sites, i.e., desmoking, defogging, and desplashing. SurgClean comprises 3,113 images with diverse degradation types and corresponding paired reference labels. Based on SurgClean, we establish a standardized evaluation benchmark and provide performance for 22 representative generic task-specific image restoration approaches, including 12 generic and 10 task-specific image restoration approaches. Experimental results reveal substantial performance gaps relative to clinical requirements, highlighting a critical opportunity for algorithm advancements in intelligent surgical restoration. Furthermore, we explore the degradation discrepancies between surgical and natural scenes from structural perception and semantic understanding perspectives, providing fundamental insights for domain-specific image restoration research. Our work aims to empower restoration algorithms and improve the efficiency of clinical procedures. Data and code are available.

Underwater images often exhibit dominant blue-green hues due to wavelength-dependent light attenuation. While existing enhancement methods have achieved promising performance, they typically overlook the subjective nature of visual preferences. To address this gap, we propose SDUIE, a level-aware Semi-supervised Diffusion framework for Underwater Image Enhancement that enables dual control through both quantitative and textual inputs. SDUIE-Quant allows continuous, numerical adjustment of enhancement levels via low-rank adaptation weight merging within a dual-branch diffusion model. This model comprises a supervised branch trained on synthetic underwater-terrestrial pairs and a self-supervised branch designed to preserve the natural hues of real-world underwater scenes. Building on this, SDUIE-Text introduces intuitive, language-guided control by aligning semantic prompts with visual enhancement effects, leveraging the learned fusion weights. This dual-modality design offers both precise control and flexible, user-preferred enhancement. Experimental results demonstrate that SDUIE achieves state-of-the-art results while better preserving the aesthetic qualities often missed by conventional methods. The source code will be made publicly available.

Thermal infrared (TIR) imaging enables robust perception in adverse conditions. However, it often suffers from complex degradations (\textit{e.g.,} fixed-pattern noise and low-resolution) due to sensor limitations and environmental dynamics. Existing methods, whether traditional or learning-based, easily fail under composite and varying degradation. Pre-trained generative models showcase powerful capabilities for alleviating degradations but lack effective tools to adapt visible generative priors to TIR-specific characteristics. To overcome these challenges, we propose a Hierarchical Degradation Representation and Adaptation (HiDRA) framework to decompose the enhancement procedure into degradation representation estimation and generative model fine-tuning. The degradation representation estimation aims to disentangle TIR degradation patterns, which then guide the parameter adaptation for thermal image enhancement. Additionally, we introduce a hierarchical adaptation solution that aggregates learning across varying degradation levels, further improving robustness under various scenarios. Experiments across diverse types and degrees demonstrate the robustness of our approach and further validate its effectiveness on downstream tasks.

This paper introduces a novel unsupervised approach for image deblurring that utilizes a simple process for training data collection, thereby enhancing the applicability and effectiveness of deblurring methods. Our technique does not require meticulously paired data of blurred and corresponding sharp images; instead, it uses unpaired blurred and sharp images of similar scenes to generate pseudo-ground truth data by leveraging a dense matching model to identify correspondences between a blurry image and reference sharp images. Thanks to the simplicity of the training data collection process, our approach does not rely on existing paired training data or pre-trained networks, making it more adaptable to various scenarios and suitable for networks of different sizes, including those designed for low-resource devices. We demonstrate that this novel approach achieves state-of-the-art performance, marking a significant advancement in the field of image deblurring.

Low-Light Image Enhancement (LLIE) is a challenging task, as severe information loss means a single input can correspond to multiple plausible restorations. This inherent ambiguity causes conventional regression-based models to produce overly-smooth results that lack detail. While recent generative models can create richer details, their common unidirectional design often compromises content fidelity by distorting original structures. We introduce Bi-Bridge, a unified framework that models both enhancement and its inverse degradation within a single symmetric diffusion bridge. By compelling the network to preserve essential content structures across both transformations, this bidirectional learning acts as a powerful constraint, leading to significantly more faithful and realistic restorations. Extensive experiments show that Bi-Bridge outperforms state-of-the-art (SOTA) methods across multiple benchmarks, establishing a new standard for fidelity and perceptual quality.

All-in-One Image Restoration (AiOIR) has emerged as a promising yet challenging research direction. To address the core challenges of diverse degradation modeling and detail preservation, we propose UniLDiff, a unified framework enhanced with degradation- and detail-aware mechanisms, unlocking the power of diffusion priors for robust image restoration. Specifically, we introduce a Degradation-Aware Feature Fusion (DAFF) to dynamically inject low-quality features into each denoising step via decoupled fusion and adaptive modulation, enabling implicit modeling of diverse and compound degradations. Furthermore, we design a Detail-Aware Expert Module (DAEM) in the decoder to enhance texture and fine-structure recovery through expert routing. Extensive experiments across multi-task and mixed degradation settings demonstrate that our method consistently achieves state-of-the-art performance, highlighting the practical potential of diffusion priors for unified image restoration. Our code will be released.

Video matting remains limited by the scale and realism of existing datasets. While leveraging segmentation data can enhance semantic stability, the lack of effective boundary supervision often leads to segmentation-like mattes lacking fine details. To this end, we introduce a learned Quality Evaluator (QE) that assesses semantic and boundary quality of alpha mattes without ground truth. It produces a pixel-wise evaluation map that identifies reliable and erroneous regions, enabling fine-grained quality assessment. The QE scales up video matting in two ways: (1) as an online matting-quality feedback during training to suppress erroneous regions, providing comprehensive supervision, and (2) as an offline selection module for data curation, improving annotation quality by combining the strengths of leading video and image matting models. This process allows us to build a large-scale real-world video matting dataset, VMReal, containing 28K clips and 2.4M frames. To handle large appearance variations in long videos, we introduce a reference-frame training strategy that incorporates long-range frames beyond the local window for effective training. Our MatAnyone 2 achieves state-of-the-art performance on both synthetic and real-world benchmarks, surpassing prior methods across all metrics.

Shooting video with handheld shooting devices often results in blurry frames due to shaking hands and other instability factors. Although previous video deblurring methods have achieved impressive progress, they still struggle to perform satisfactorily on real-world handheld video due to the blur domain gap between training and testing data. To address the issue, we propose a self-supervised method for handheld video deblurring, which is driven by sharp clues in the video. First, to train the deblurring model, we extract the sharp clues from the video and take them as misalignment labels of neighboring blurry frames. Second, to improve the deblurring ability of the model, we propose a novel Self-Enhanced Video Deblurring (SEVD) method to create higher-quality paired video data. Third, we propose a Self-Constrained Spatial Consistency Maintenance (SCSCM) method to regularize the model, preventing position shifts between the output and input frames. Moreover, we construct synthetic and real-world handheld video datasets for handheld video deblurring. Extensive experiments on these and other common real-world datasets demonstrate that our method significantly outperforms existing self-supervised ones. The code and datasets will be publicly available.

Motion blur arises when rapid scene changes occur during the exposure period, collapsing rich intra-exposure motion into a single RGB frame. Without explicit structural or temporal cues, RGB-only deblurring is highly ill-posed and often fails under extreme motion.Inspired by the human visual system, neuromorphic sensors introduce temporally dense information to alleviate this problem; however, event cameras still suffer from event rate saturation under rapid motion, while the event modality entangles edge features and motion cues, which limits their effectiveness.As a recent breakthrough, the complementary vision sensor (CVS) captures synchronized RGB frames together with high-frame-rate, multi-bit spatial difference ($\mathcal{SD}$, encoding structural edges) and temporal difference ($\mathcal{TD}$, encoding motion cues) data within a single RGB exposure, offering a promising solution for RGB deblurring under extreme dynamic scenes. To fully leverage these complementary modalities, we propose Spatio-Temporal Difference Guided Deblur Net (STGDNet), which adopts a recurrent multi-branch architecture that iteratively encodes and fuses $\mathcal{SD}$ and $\mathcal{TD}$ sequences to restore structure and color details lost in blurry RGB inputs. Our method outperforms current RGB or event-based approaches in both synthetic CVS dataset and real-world evaluations. Moreover, STGDNet exhibits strong generalization capability across over 100 extreme real-world scenarios. Our code, dataset and pre-trained weights will be fully publicly available.

No-reference image quality assessment (NR IQA) has recently benefited from deep and multimodal models, yet many SOTA systems still violate at least one basic requirement: they either discard critical quality cues via aggressive resizing, fail to generalize across resolutions, cannot be jointly trained on heterogeneous IQA datasets with mismatched MOS scales, or require prohibitive computation. We present $\textbf{ReLIQS}$, a model for $\textbf{Re}$solution-agnostic $\textbf{L}$earning for $\textbf{I}$mage $\textbf{Q}$uality with $\textbf{S}$aliency, which is resolution-agnostic, preserves original-resolution quality cues, learns from multiple subjective studies, and remains computationally efficient and budget-adaptive. ReLIQS is a CLIP-based multiscale patch-driven architecture that learns both \emph{where to look} and \emph{how to judge} quality. Fixed-size patches are sampled across multiple resolutions, including the original resolution, and encoded with a CLIP vision backbone. A lightweight Perceptual Importance Estimator then predicts IQA-specific importance maps to select a small set of informative patches, and a Quality Aspect Module aggregates their embeddings into a single image-level score. Across authentic, synthetic, and AIGC benchmarks spanning diverse resolutions and distortions, ReLIQS generalizes better than strong CNN-, CLIP-, and MLLM-based baselines with matching or reduced computational cost.

Nuclei instance segmentation in histopathology images is essential for diagnostic accuracy and downstream computational tasks, yet this task relies heavily on expensive pixel level annotations. Although point level annotations substantially reduce the annotation burden for pathologists, many existing methods utilize only a single type of image and overlook the complementary information contained in alternative representations. To address this limitation, we propose DFGNet, a weakly supervised framework that utilizes dual-representation complementary fusion and interleaved guidance learning by jointly modeling RGB images and their corresponding Hematoxylin components. From the complementary fusion perspective, we propose a Reciprocal Cross-scale Dynamic Fusion Module. (RCDF) and an Entropy Confidence Aggregation Unit (ECAU) to integrate multi-scale complementary cues and adaptively combine the outputs of the dual branches. In terms of interleaved guidance, we also propose an Interleaved point-Guided Attention (IGA) that enables bidirectional refinement between the segmentation task and the kernel prediction task. Extensive experiments on three benchmark datasets show that DFGNet achieves state-of-the-art performance across multiple metrics and significantly outperforms existing approaches. DFGNet also demonstrates strong generalization ability across different tissue types and exhibits remarkable robustness to annotation shifts, providing a low-cost and scalable solution for practical clinical applications.

Multimodal magnetic resonance imaging (MRI) is crucial for brain tumor segmentation, with many methods leveraging its four key modalities to capture complementary information for effective sub-region analysis. However, the absence of several modalities is very common in practice, leading to severe performance degradation in existing full-modality segmentation methods. Limited by the structured data model, recent works often adopt a multi-stage training strategy for full-modality and missing-modality scenarios, which increases training costs and inadequately addresses the interference of miss. In this work, we propose a graph-based one-stage framework for robust brain tumor segmentation with missing modalities. Specifically, we introduce modality-specific virtual nodes that serve as supplementary information sources to compensate for missing modalities. To enhance model robustness against arbitrary modality combinations, we leverage the inherent flexibility of graph networks to devise a dynamic connection strategy. This mechanism dynamically adjusts the adjacency matrix based on modality availability, preserving beneficial information flow while mitigating interference effects caused by missing modalities. Furthermore, we enhance the graph network through heterogeneous weight matrices, enhancing its adaptability to multimodal scenarios. Extensive experiments on the BRATS-2018 and BRATS-2020 datasets demonstrate that our method outperforms the state-of-the-art methods on almost all subsets of incomplete modalities.

We introduce VoxTell, a vision–language model for text-prompted volumetric medical image segmentation. It maps free-form descriptions, from single words to full clinical sentences, to 3D masks. Trained on 62K+ CT, MRI, and PET volumes spanning 1K+ anatomical and pathological classes, VoxTell uses multi-stage vision–language fusion across decoder layers to align textual and visual features at multiple scales. It achieves state-of-the-art zero-shot performance across modalities on unseen datasets, excelling on familiar concepts while generalizing to related unseen classes. Extensive experiments further demonstrate strong cross-modality transfer, robustness to linguistic variations and clinical language, as well as accurate instance-specific segmentation from real-world text. Code and model will be published at: www.github.com/anonymous

Accurate 3D tooth segmentation is fundamental for digital dentistry, orthodontic analysis, and clinical simulation. Intraoral scan (IOS) models often suffer from incomplete or unreliable texture information, making it difficult to delineate fine boundaries between teeth and gingiva, while 2D intraoral images provide rich semantic and chromatic information that can complement 3D geometry. Thus, we propose a novel Photo-guided 3D Model Tooth Segmentation framework, PMTSeg, that enhances 3D tooth segmentation by integrating texture cues from intraoral photos. Our framework introduces three key components: a Camera Alignment Module (CAM) for accurate image-model registration, a Feature Filtering Gate (FFG) for adaptive multi-view feature selection, and a Consistent Feature Learning (CFL) mechanism for learning texture-geometry correspondence. Our method supports arbitrary numbers and views of intraoral photos. Experiments show significant improvements in distinguishing adjacent teeth and tooth–gingiva boundaries, demonstrating that intraoral photographs serve as an efficient, semantically rich supplement to 3D scans for precise dental segmentation.

Anatomical structures in MRI exhibit strong spatial priors, including well-defined boundaries, low inter-subject variability, and consistent topology. These properties naturally induce clustered patterns in the latent space, which are difficult to capture using conventional continuous generative priors that assume smooth manifold distributions. To address this limitation, we propose DiCoS (Discrete–Continuous Synthesis), a generative reconstruction framework that integrates discrete structural reasoning with continuous refinement. DiCoS models an anatomy-aware discrete distribution and generates diverse reconstructions in one coarse-to-fine pass through a Discrete Prior Network (DPN). A Dual-domain Balanced Scoring (DBS) mechanism adaptively evaluates candidates using both image-domain fidelity and k-space consistency. To further enhance realism, Micro Diffusion Cycles (MDC) perform efficient score-guided refinement to enhance texture realism without disturbing global topology. Experiments on the fastMRI knee and brain datasets demonstrate that DiCoS achieves state-of-the-art reconstruction quality with sharper boundaries and improved anatomical consistency. Beyond pixel metrics, segmentation-based evaluations further confirm superior structural overlap and semantic alignment, highlighting DiCoS's advantages in anatomy-aware reconstruction. Code and models will be released upon publication.

Digital hematopathology requires cell-level analysis across diverse disease categories, including malignant disorders (e.g., leukemia), infectious conditions (e.g., malaria), and non-malignant red blood cell disorders (e.g., sickle cell disease). Whether single-task, vision-language, WSI-optimized, or single-cell hematology models, these approaches share a key limitation: they cannot provide unified, multi-task, multi-modal reasoning across the complexities of digital hematopathology. To overcome these limitations, we propose \textbf{Uni-Hema}, a multi-task, unified model for digital hematopathology integrating detection, classification, segmentation, morphology prediction, and reasoning across multiple diseases. Uni-Hema leverages 46 publicly available datasets, encompassing over 700K images and 21K question–answer pairs, and is built upon \textbf{Hema-Former}, a multimodal module that bridges visual and linguistic representations at the hierarchy level for the different tasks (detection, classification, segmentation, morphology, mask language modeling and visual question answer) at different granularity. Extensive experiments demonstrate that Uni-Hema achieves comparable or superior performance to train on a single-task and single dataset models, across diverse hematological tasks, while providing interpretable, morphologically relevant insights at the single-cell level. Our framework establishes a new standard for multi-task and multi-modal digital hematopathology. The code will be made publicly available.

Deep neural networks have achieved strong performance in fundus image classification, yet their flattened feature representations are often highly redundant. Such redundancy can lead to poor generalization across imaging devices, reduced interpretability, and inefficient use of model capacity. To address this issue, this study proposes a post-training feature pruning framework, termed {greedy feature pruning (GFP)}, which removes weak or redundant dimensions from the flattened features of trained backbones. GFP employs a greedy build-up process guided by performance metrics on the training set, constrained by a minimum feature keeping ratio, to identify compact yet discriminative subsets of features. Experiments are conducted on five public fundus datasets covering multiple tasks, including diabetic retinopathy detection (DDR, Messidor-2), glaucoma detection (PAPILA), multi-label classification (ODIR) and multi-class retinal disease classification (RETINA), using EfficientNetV2, ViT, and CoAtNet as backbones. Results show that GFP consistently improves AUROC and AUPRC across datasets while reducing the number of flattened features by up to 96\%. Feature visualizations and quantitative analyses confirm that GFP enhances the compactness and separability of latent features. Moreover, cross-dataset evaluation demonstrates that GFP improves transferability between datasets, indicating better domain robustness. Overall, the proposed GFP framework provides a simple yet effective approach for compressing feature representations and improving both discriminability and generalization in fundus image classification.

Diffusion probabilistic models have demonstrated significant potential in generating high-quality, realistic medical images, providing a promising solution to the persistent challenge of data scarcity in the medical field. Nevertheless, producing 3D medical volumes with anatomically consistent structures under multimodal conditions remains a complex and unresolved problem. We introduce Sketch2CT, a multimodal diffusion framework for structure-aware 3D medical volume generation, jointly guided by a user-provided 2D sketch and a textual description that captures 3D geometric semantics. The framework initially generates 3D segmentation masks of the target organ from random noise, conditioned on both modalities. To effectively align and fuse these inputs, we propose two key modules that refine sketch features with localized textual cues and integrate global sketch-text representations. Built upon a capsule-attention backbone, these modules leverage the complementary strengths of sketches and text to produce anatomically accurate organ shapes. The synthesized segmentation masks subsequently guide a latent diffusion model for 3D CT volume synthesis, enabling realistic reconstruction of organ appearances that are consistent with user-defined sketches and descriptions. Extensive experiments on public CT datasets demonstrate that Sketch2CT achieves superior performance in generating multimodal medical volumes. Its controllable, low-cost generation pipeline enables principled, efficient augmentation of medical datasets.

Recent Vision-Language Models (VLMs) have become increasingly susceptible to jailbreak attacks, where adversarial prompts exploit subtle manipulation to circumvent safety alignment.The diversity and adaptability of such jailbreakers necessitate a defense mechanism with strong generalization capability.However, fine-tuning large-scale VLMs is computationally expensive, and introducing excessive visual or textual defense prompts is impractical for preserving image realism and model usability.We propose SafeLogo, which tunes a logo-sized visual prompt into a universal shield against diverse jailbreak attacks through micro-regional adversarial training.We are the first to integrate min–max adversarial optimization into visual defense prompt generation.Specifically, in the outer loop, SafeLogo injects compact, bounded perturbations into extremely small image regions ($\leq 2\%$ pixel coverage), effectively preserving both visual fidelity and semantic consistency.Meanwhile, overcoming the limitations of prior defenses constrained to a single attack direction or fixed benign supervision, the inner loop dynamically generates and selects the strongest one from a variety of jailbreakers.Extensive experiments on LLaVA-1.5-13B,MiniGPT-4, and Qwen3-VL show that SafeLogo markedly lowers jailbreak success rates on MM-SafetyBench, VLGuard, and FigStep, while preserving benign performance on MM-Vet and MME.

Advances in diffusion-based video generation models, while significantly improving human animation, poses threats of misuse through the creation of fake videos from a specific person's photo and text prompts. Recent efforts have focused on adversarial attacks that introduce crafted perturbations to protect images from diffusion-based models. However, most existing approaches target image generation, while relatively few explicitly address image-to-video diffusion models (VDMs), and most primarily focus on UNet-based architectures. Hence, their effectiveness against Diffusion Transformer (DiT) models remains largely under-explored, as these models demonstrate improved feature retention, and stronger temporal consistency due to larger capacity and advanced attention mechanisms. In this work, we introduce Anti-I2V, a novel defense against malicious human image-to-video generation, applicable across diverse diffusion backbones. Instead of restricting noise updates to the RGB space, Anti-I2V operates in both the $L$*$a$*$b$* and frequency domains, improving robustness and concentrating on salient pixels. We then identify the network layers that capture the most distinct semantic features during the denoising process to design appropriate training objectives that maximize degradation of temporal coherence and generation fidelity. Through extensive validation, Anti-I2V demonstrates state-of-the-art defense performance against diverse video diffusion models, offering an effective solution to the problem.

Unified Multimodal Models (UMMs) have shown impressive performance in both understanding and generation with a single architecture. However, UMMs still exhibit a fundamental inconsistency: understanding favors compact embeddings, whereas generation favors reconstruction-rich representations. This structural trade-off produces misaligned decision boundaries, degraded cross-modal coherence, and heightened vulnerability under distributional and adversarial shifts. In this paper, we present UniGame, a self-adversarial post-training framework that directly targets the inconsistencies. By applying a lightweight perturber at the shared token interface, UniGame enables the generation branch to actively seek and challenge fragile understanding, turning the model itself into its own adversary. Experiments demonstrate that UniGame significantly improves the consistency (+4.6%). Moreover, it also achieves substantial improvements in understanding (+3.6%), generation (+0.02), out-of-distribution and adversarial robustness (+4.8% and +6.2% on NaturalBench and AdVQA). The framework is architecture-agnostic, introduces <1% additional parameters, and is complementary to existing post-training methods. These results position adversarial self-play as a general and effective principle for enhancing the coherence, stability, and unified competence of future multimodal foundation models.

Vision-Language Models (VLMs) can perform zero-shot classification but are susceptible to adversarial attacks. While robust fine-tuning improves their robustness, existing approaches align fixed text embeddings with an image embedding, sacrificing natural performance and robustness. A robustness degradation also occurs when a model faces adversarial attacks targeting superclasses (parent classes, e.g., \texttt{mammal}) in addition to their base (leaf) classes (e.g., \texttt{cat}). Thus, to enhance adversarial robustness and leverage the inherent hierarchical properties of class space, we propose a novel adversarial fine-tuning framework based on hierarchical embeddings and several levels of adversarially robust alignment of image-text modalities. Additional mechanisms place visual embeddings at the desired depth of hierarchy, and we provide theoretical connection between the depth of embedding in the hierarchy and the maximum viable margin size. Our model naturally realizes several margin sizes boosting generalization of adversaries for robustification. As various trees with different parent labels can share the same leaf labels, we also consider aligning over multiple trees to boost semantic variety. Experiments across several datasets are performed.

Text-to-image generative models have achieved impressive fidelity and diversity, but can inadvertently produce unsafe or undesirable content due to implicit biases embedded in large-scale training datasets.Existing concept erasure methods, whether text-only or image-assisted, face trade-offs: textual approaches often fail to fully suppress concepts, while naive image-guided methods risk over-erasing unrelated content. We propose $\textbf{TICoE}$, a Text–Image Collaborative Erasing framework that achieves precise and faithful concept removal through a continuous convex concept manifold and hierarchical visual representation learning. TICoE precisely removes target concepts while preserving unrelated semantic and visual content. To objectively assess the quality of erasure, we further introduce a fidelity-oriented evaluation strategy that measures post-erasure usability. Experiments on multiple benchmarks show that TICoE surpasses prior methods in concept removal precision and content fidelity, enabling safer, more controllable text-to-image generation.

The integration of vision-language models (VLMs) is driving a new generation of embodied agents capable of operating in human-centered environments. However, as deployment expands, these systems face growing safety risks, particularly when executing hazardous instructions. Current safety evaluation benchmarks remain limited: they cover only narrow scopes of hazards and focus primarily on final outcomes, neglecting the agent's full perception-planning-execution process and thereby obscuring critical failure modes. Therefore, we present SAFE, a benchmark for systematically assessing the safety of embodied VLM agents on hazardous instructions. SAFE comprises three components: SAFE-THOR, an extensible adversarial simulation sandbox with a universal adapter that maps high-level VLM outputs to low-level embodied controls, supporting diverse agent workflow integration; SAFE-VERSE, a risk-aware task suite inspired by Asimov's Three Laws of Robotics, comprising 45 adversarial scenarios, 1,350 hazardous tasks, and 9,900 instructions that span risks to humans, environments, and agents; and SAFE-DIAGNOSE, a multi-level and fine-grained evaluation protocol measuring agent performance across perception, planning, and execution. Applying SAFE to nine state-of-the-art VLMs and two embodied agent workflows, we uncover systematic failures in translating hazard recognition into safe planning and execution. Our findings reveal fundamental limitations in current safety alignment and demonstrate the necessity of a comprehensive, multi-stage evaluation for developing safer embodied intelligence.

Vision Transformers (ViTs) achieve state-of-the-art performance on a wide range of vision tasks, yet they remain highly vulnerable to adversarial perturbations due to the lack of explicit region-level semantic modeling. Adversarial perturbations are typically local and spatially structured, whereas the globally coupled self-attention and spatially uniform feed-forward networks in ViTs propagate local corruptions across the whole image without enforcing consistency within semantically coherent regions. To mitigate this mismatch, we propose Region-aware Mixture-of-Experts, namely "ReMoE", a plug-and-play module that replaces the standard feed-forward network (FFN) with a region-aware expert layer. Specifically, our ReMoE strategically introduces multi-granularity experts (i.e., global, center, and regional) and couples them with an attention-guided routing mechanism that operates on patch-to-region (P2R) and region-to-patch (R2P) transformations. This mechanism adaptively activates the most relevant experts for each spatial location according to its attention profile, enabling the model to capture region-level semantics and local context while preserving global consistency, thereby providing a stronger inductive bias for adversarially robust ViT representations. Extensive experiments demonstrate that our ReMoE substantially improves the adversarial robustness of ViTs with only marginal additional computational cost.

The rapid advancement of Vision-Language Models (VLMs) has brought their safety vulnerabilities into sharp focus. However, existing red teaming methods are fundamentally constrained by an inherent linear exploration paradigm, confining them to optimizing within a predefined strategy set and preventing the discovery of novel, diverse exploits. To transcend this limitation, we introduce TreeTeaming, an automated red teaming framework that reframes strategy exploration from static testing to a dynamic, evolutionary discovery process. At its core lies a strategic Orchestrator, powered by a Large Language Model (LLM), which autonomously decides whether to evolve promising attack paths or explore diverse strategic branches, thereby dynamically constructing and expanding a strategy tree. A multimodal actuator is then tasked with executing these complex strategies. In the experiments across 12 prominent VLMs, TreeTeaming achieves state-of-the-art attack success rates on 11 models, outperforming existing methods and reaching up to 87.60\% on GPT-4o. The framework also demonstrates superior strategic diversity over the union of previously public jailbreak strategies. Furthermore, the generated attacks exhibit an average toxicity reduction of 23.09\%, showcasing their stealth and subtlety. Our work introduces a new paradigm for automated vulnerability discovery, underscoring the necessity of proactive exploration beyond static heuristics to secure frontier AI models.Warning: This paper contains examples of harmful texts and images, and reader discretion is recommended.

Multimodal large language models (MLLMs) are increasingly deployed in real-world, agentic settings where outputs must not only be correct, but also conform to pre-defined data schemas. Despite recent progress in structured generation in textual domain, there is still no benchmark that systematically evaluates schema-grounded information extraction and reasoning over visual inputs. In this work, we conduct a comprehensive study of visual structural output capabilities for MLLMs with our carefully designed SO-BENCH benchmark. Covering four visual domains, including UI screens, natural images, documents, and charts, SO-BENCH is built from over 6.5K diverse JSON schemas and 1.8K curated image–schema pairs with human-verified quality. Benchmarking experiments on open-sourced andfrontier proprietary models reveal persistent gaps in predicting accurate, schema compliant outputs, highlighting the need for better multimodal structured reasoning. Beyond benchmarking, we further conduct training experiments to largely improve the model’s structured output capability. We plan to make the benchmark available to the community.

With the rise of visual-language models, multi-modal ReID retrieves specific targets by integrating different spectra and textual descriptions. Existing methods merely adopt descriptive representation learning for image-text, ignoring the relationships among the intrinsic logical hierarchies of semantic features. Since Chain-of-Thought (CoT) can provide textual logical context and enhance semantic perception in large-model reasoning, we propose CoT-ReID, a CoT-guided framework that injects the Multi-modal Large Language Models (MLLMs) reasoning into multi-modal ReID. Specifically, we simulate the joint visual-textual logical decision-making of human reasoning, leveraging CoT textual logical reasoning to guide visual feature learning at the early, late, and decision-making level: At the early level, we embed the semantic reversion of CoT hierarchical reasoning into visual features to calibrate bottom-level features and emphasize visual hierarchical reasoning. Next, we take CoT hierarchical reasoning text as an anchor condition to constrain the consistency of visual cross-modal semantics. Finally, through the hierarchical reasoning process of CoT, we embed logically reasoned text attribute features into multi-modal decision-making, providing logical support for selecting discriminative identity features. By constructing CoT textual benchmarks and our proposed modules, our framework generates more robust multi-modal features in complex scenarios. Comprehensive experiments on four datasets (RGBNT100, MSVR310, WMVeID863, RGBNT201) demonstrate that our method outperforms existing approaches. Code will be released upon acceptance.

Establishing accurate correspondence between sparse line representations and rich textured imagery remains a formidable challenge. While diffusion features excel in semantic correspondence, they struggle to bridge the fundamental gap between abstract sketches and texture-rich photographs. We identify two critical disparities: spatial domain misalignment from structural abstraction differences, and frequency domain inconsistencies from texture density variations. Based on this analysis, we propose SFA-DIFT, a novel approach that learns spatial-frequency aligned diffusion features for robust cross-modal correspondence. Unlike previous methods focusing solely on spatial alignment, our key innovation performs dual-domain alignment by learning unified clean diffusion features while strategically aggregating low-frequency components in the frequency domain. This comprehensive spatial-frequency alignment enables equitable understanding between sparse abstractions and rich textures. To validate our approach, we extend the existing sketch-photo correspondence dataset (PSC6K) by generating multi-style textured imagery, creating MS-PSC6K, a comprehensive correspondence benchmark. Extensive experiments demonstrate that SFA-DIFT achieves state-of-the-art performance, delivering substantial improvements with an average of 0.87\% on PCK@1, 2.20\% on PCK@5, and 0.95\% on PCK@10 over previous best methods, validating the effectiveness and robustness of our dual-domain alignment approach.

The flexibility and accuracy of methods for automatically counting objects in images and videos are limited by the way the object can be specified. While existing methods allow users to describe the target object with text and visual examples, the visual examples must be manually annotated inside the image, and there is no way to specify what not to count. To address these gaps, we introduce novel capabilities that expand how the target object can be specified. Specifically, we extend the prompt to enable what not to count to be described with text and/or visual examples, introduce the concept of `pseudo-exemplars' that automate the annotation of visual examples at inference, and extend counting models to accept visual examples from both natural and synthetic external images. We also use our new counting model, CountGD++, as a vision expert agent for an LLM. Together, these contributions expand the prompt flexibility of multi-modal open-world counting and lead to significant improvements in accuracy, efficiency, and generalization across multiple datasets.

Recent advances in text-to-audio (TTA) generation excel at synthesizing short audio clips but struggle with long-form narrative audio, which requires temporal coherence and compositional reasoning. To fill this gap, we propose AudioStory, a unified framework that integrates large language models (LLMs) with TTA systems to generate structured, long-form audio narratives. AudioStory possesses strong instruction-following reasoning generation capabilities. It employs LLMs to decompose complex narrative queries into temporally ordered sub-tasks with contextual cues, enabling coherent scene transitions and emotional tone consistency. AudioStory has two appealing features: (1) Decoupled bridging mechanism: AudioStory disentangles LLM-diffuser collaboration into two specialized components, i.e., a bridging query for intra-event semantic alignment and a residual query for inter-event coherence preservation. (2) End-to-end training: By unifying instruction comprehension and audio generation within a single end-to-end framework, AudioStory eliminates the need for modular training pipelines while enhancing synergy between components. Furthermore, we establish a benchmark AudioStory-10K, encompassing diverse domains such as animated soundscapes and natural sound narratives. Extensive experiments show the superiority of AudioStory on both single-audio generation and narrative audio generation, surpassing prior TTA baselines in both instruction-following ability and audio fidelity. Our code and dataset will be publicly available.

Parameter-efficient fine-tuning (PEFT) has become a compelling approach for adapting large language models (LLMs) into multimodal large language models (MLLMs), enabling them to handle diverse modalities with substantially lower memory and computational costs. However, most existing PEFT methods neglect the issue of modality-imbalanced learning, which is characterized by the excessive dominance of text modality in updating parameters, thus incurring insufficient learning of non-text modalities and leading to performance degradation. To address this issue, we propose a novel parameter-efficient adaptation method for MLLMs, namely Implicit Modality Decomposition (IMoD), based on LoRA. It firstly decomposes the learnable parameters into the non-overlapped text-specific, non-text-specific and modality-sharing components, thereby alleviating modality imbalance. To further guide the optimization of these components toward specific modalities, we propose Modality-Specific Decoupling Constraint that suppresses cross-modal interference among modality-specific parameters, and Modality-Agnostic Alignment Constraint that encourages modality-sharing component to capture well-aligned, modality-invariant semantics. Extensive experiments across diverse multimodal settings and LLM architectures demonstrate that our method consistently delivers significant performance gains, particularly achieving an averaged 3.3\% improvements on the audio-visual-text tasks without sacrificing the parameter and inference efficiency. We will release the source code upon acceptance.

Multimodal contrastive learning typically relies on pairwise similarities for alignment, but recent work has shown that Gramian volumes can capture higher-order correlations across modalities.However, Euclidean Gramian volumes suffer from volume collapse under L2 normalization, concentrating near unity with minimal discriminative variance.Hyperbolic geometry's exponential volume growth naturally addresses this via variance preservation, motivating us to extend Gramian alignment to hyperbolic space.Yet preliminary experiments reveal that pure hyperbolic geometry alone is insufficient: while it preserves variance, it underperforms Euclidean baselines on cross-category discrimination.We introduce HyperGRAM, a hybrid geometry framework that combines Euclidean discriminative stability with hyperbolic semantic variance through learnable mixing.Using the numerically stable Lorentz model, HyperGRAM enables volumes to serve dual roles: discriminating matched from mismatched triplets while preservingsemantic sensitivity within matched pairs that reflects interpretation spaces (the set of valid multimodal realizations).Evaluation across four video-text benchmarks demonstrates that hybrid geometry consistently outperforms both pure Euclidean and pure hyperbolic variants, achievingsignificant zero-shot improvements with cross-dataset semantic sensitivity exhibiting contrasting correlation patterns.

The widespread use of smartphones has made photography ubiquitous, yet a clear gap remains between ordinary users and professional photographers, who can identify aesthetic issues and provide actionable shooting guidance during capture. We define this capability as aesthetic guidance (AG) — an essential but largely underexplored domain in computational aesthetics. Existing multimodal large language models (MLLMs) primarily offer overly positive feedback, failing to identify issues or provide actionable guidance. Without AG capability, they cannot effectively identify distracting regions or optimize compositional balance, thus also struggling in aesthetic cropping, which aims to refine photo composition through reframing after capture. To address this, we introduce AesGuide, the first large-scale AG dataset and benchmark with 10,748 photos annotated with aesthetic scores, analyses, and guidance. Building upon it, we propose Venus, a two-stage framework that first empowers MLLMs with AG capability through progressively complex aesthetic questions and then activates their aesthetic cropping power via CoT-based rationales. Extensive experiments show that Venus substantially improves AG capability and achieves state-of-the-art (SOTA) performance in aesthetic cropping, enabling interpretable and interactive aesthetic refinement across both stages of photo creation.

Short-form videos have become a primary medium for digital advertising, requiring scalable and efficient content creation.However, current workflows and AI tools remain disjoint and modality-specific, leading to high production costs and low overall efficiency. To address this issue, we propose AutoCut, an end-to-end video ad editing framework based on multimodal discretization and controllable generation. AutoCut employs dedicated encoders to extract video and audio features, then applies residual vector quantization to discretize them into unified tokens aligned with textual representations, constructing a shared video–audio–text token space. Built upon a foundation model, we further develop a multimodal large language model for intelligent video editing through combined multimodal alignment and supervised fine-tuning, supporting tasks covering video selection and ordering, script generation, and background music selection within a unified editing framework. Finally, a complete production pipeline converts generated tokens into deployable long video outputs. Experiments on real-world advertisement datasets show that AutoCut reduces production cost and iteration time while substantially improving consistency and controllability, paving the way for scalable generative video creation.

Multi-task learning for dense prediction is limited by the need for extensive annotation for every task, although recent works have explored training with partial task labels.Leveraging the generalization power of diffusion models, we extend the partial learning setup to a zero-shot setting, training a multi-task model on multiple synthetic datasets, each labeled for only a subset of tasks. Our method, StableMTL, repurposes an image generator for latent regression. Adapting a denoising framework with task encoding, task conditioning and a tailored training scheme. Instead of per-task losses requiring careful balancing, a unified latent loss is adopted, enabling seamless scaling to more tasks. To encourage inter-task synergy, we introduce a multi-stream model with a task-attention mechanism that converts N-to-N task interactions into efficient N-to-one attention, promoting effective cross-task sharing. StableMTL outperforms baselines on 7 tasks across 8 benchmarks.

Modality gap significantly restricts the effectiveness of multimodal fusion. Previous methods often use techniques such as diffusion models and adversarial learning to reduce the modality gap, but they typically focus on one-to-one alignment without exposing the data points of the source modality to the global distribution information of the target modality. To this end, leveraging the characteristic of rectified flow that can map one distribution to another via a straight trajectory, we extend rectified flow for modality distribution mapping. Specifically, we leverage the 'one-to-many mapping' strategy in rectified flow that allows each data point of source modality to observe the overall target distribution. This also alleviates the issue of insufficient paired data within each sample, enabling a more robust distribution transformation. Moreover, to achieve more accurate distribution mapping and address the ambiguous flow directions in one-to-many mapping, we design 'adaptive relaxed alignment', enforcing stricter alignment for modality pairs belonging to the same sample, while applying relaxed mapping for pairs not belonging to the same sample or category. Additionally, to prevent information loss during distribution mapping, we introduce 'cyclic rectified flow' to ensure the transferred features can be translated back to the original features, allowing multimodal representations to learn sufficient modality-specific information. After distribution alignment, our approach achieves very competitive results on multiple tasks of multimodal affective computing even with a simple fusion method, and visualizations verify that it can effectively reduce modality gap.

Most vision-language models assume images have a single literal meaning, even though images are polysemous. We propose a retrieval paradigm that models many-to-many relationships between images and text using interpretive lenses and introduce Lenses, a multi-prompt embedding model and dataset for polysemous image-text retrieval. The Lenses dataset contains (105,669) images and (732,405) captions, with each image paired with multiple captions and image-side prompts annotated across five categories: Literal, Figurative, Emotional, Abstract, and Background. Building on a multimodal large language model, the Lenses model uses learned lens tokens to extract lens-specific embeddings for every image and caption and compares these using a lens-masking similarity function with a global fallback that prioritizes same-lens matches while retaining a global pathway. Training uses a category-aware multi-positive contrastive loss and intra-set diversity regularization to align corresponding perspectives while preventing semantic collapse across lenses. We further propose lens-aware evaluation protocols, including category-aware ranking, that better reflect how humans match images and text. Experiments on the Lenses dataset and public benchmarks show that our model outperforms baselines on literal and non-literal retrieval and reduces over-reliance on literal cues.

Dynamic multimodal fusion methods lack robust theoretical guidance for handling modal conflicts and inconsistent data quality. While recent theory-based works correlate weights with indirect scalar proxies (e.g., loss or confidence), this paradigm struggles to comprehensively capture the risk driven by direct distribution inconsistencies. In this paper, we propose a **Co**nflict-driven **Ri**sk **M**inimization (**CoRiM**) dynamic fusion paradigm. Specifically, we redefine dynamic fusion as a principled, per-sample, direct risk minimization task. To this end, we first design a novel, differentiable Modality Conflict Risk (MCR) function, $\mathcal{R}(w)$, which quantifies risk by directly modeling fused uncertainty and inter-modal consistency. Second, we identify that minimizing $\mathcal{R}(w)$ is fundamentally a non-convex constrained optimization problem over the probabilistic simplex. To efficiently solve this specific challenge, we innovatively introduce the projection free Frank-Wolfe (FW) algorithm, as it is perfectly suited for optimization on the simplex.We prove that our designed $\mathcal{R}(w)$ possesses L-smoothness, which provides theoretical guarantees for the convergence of the FW algorithm on our non-convex objective. Extensive experiments on multiple benchmark datasets demonstrate that CoRiM outperforms current state-of-the-art methods in high-conflict and noisy environments, validating the robustness of our method.

Direct Preference Optimization (DPO) has proven to be an effective solution for mitigating hallucination in Multimodal Large Language Models (MLLMs) by learning from preference pairs. One of its key challenge lies in how to transfer the sequence-level preference into fine-grained supervision on visual fidelity. To safeguard vision-related tokens that are prone to hallucination, existing methods typically allocate training emphasis according to the model's self-assessed visual sensitivity signals. However, such sensitivity, estimated by a model still under training, introduces self-referential bias: reinforcing already well-learned visual cues while neglecting hard-to-perceive but critical details, thereby limiting deeper alignment. In this work, we propose an Uncertainty-aware Exploratory Direct Preference Optimization (UE-DPO) method for MLLMs, which enables the model to uncover its cognitive deficiencies and actively explore for self-correction, guided by token-level epistemic uncertainty. Specifically, we first quantify the uncertainty from the model's failure to ground token predictions in the given image. Then, based on an uncertainty-aware exploration intensity, we encourage more learning pressure on visually deficient tokens in preferred samples, and alleviate the over-penalization of beneficial knowledge in dispreferred samples. Further, we provide a theoretical justification for our method, and extensive experiments on hallucination benchmarks demonstrate its effectiveness and robustness. All code will be released.

Multimodal sentiment analysis requires integrating language, visual, and acoustic cues, yet these modalities are often noisy, incomplete, or contradictory, making fusion unreliable. Most existing methods assume uniformly trustworthy modalities and thus degrade when signals conflict.To address this, we propose CICA, a framework that couples Confidence-Aware Pretraining with Confidence-Informed Attention. In pretraining, each modality encoder learns to estimate the reliability of its own representation, producing both embeddings and confidence scores. These scores then guide a confidence-informed attention mechanism, which strengthens contributions from reliable modalities while suppressing noisy or conflicting ones, enabling adaptive fusion under varying signal conditions.CICA achieves state-of-the-art performance across four major benchmarks on MOSI, MOSEI, CH-SIMS, and CH-SIMSv2. It achieves MAE 0.630 and Corr 0.855 on MOSI, and MAE 0.489 and Corr 0.856 on MOSEI, significantly surpassing prior methods. Consistent improvements are also observed across Acc-7, Acc-2, and F1 metrics. Under noisy and missing-modality conditions, CICA maintains significantly more stable performance, indicating improved robustness and interpretability.

Pixel-wise capabilities are essential for building interactive intelligent systems. However pixel-wise multi-modal LLMs (MLLMs) remain difficult to scale due to complex region-level encoders, specialized segmentation decoders, and incompatible training objectives. To solve these challenges, we present SAMTok, a discrete mask tokenizer that converts any region mask into two textual special tokens and reconstructs masks from these tokens with high fidelity. By treating masks as a new language, SAMTok enables base MLLMs (such as the QwenVL series) to learn pixel-wise capabilities through standard next-token prediction and simple reinforcement learning, without architectural modifications and specialized loss design. SAMTok builds on SAM2 and is trained on 209M diverse masks using a mask encoder and residual vector quantizer to produce discrete, compact, and information-rich tokens. With 5M SAMTok formatted mask understanding and generation data samples, QwenVL-SAMTok attains state-of-the-art or comparable results on region captioning, region VQA, grounded conversation, referring segmentation, scene graph parsing, and multi-round interactive segmentation. We further introduce a textual answer-matching reward that enables efficient reinforcement learning for mask generation, delivering substantial improvements on GRES and GCG benchmarks. Our results demonstrate a simple and scalable paradigm for equipping MLLMs with strong pixel-wise capabilities. Code and models will be available.

Multimodal sentiment analysis (MSA) aims to identify human emotions through multimodal data. Despite considerable advances in MSA, we find that emotional class centers often overlap when integrating data from different modalities into the same representation space. In this paper, we propose a novel $\textbf{M}$ulti-$\textbf{M}$etric $\textbf{R}$epresentation l$\textbf{e}$arning $\textbf{s}$trategy based on clus$\textbf{t}$ering (MMRest) to alleviate this issue through flexible multi-metric representation learning, enabling the model to learn fine-grained sentiments. Specifically, we first design a module termed Multi-metric Multimodal learning on Clusters (MMC), which minimizes distances within similar sentiment pairs while maximizing dissimilar ones, aiming to learn a global metric and local metrics in each cluster from multimodal data. Afterwards, we develop a Projection and Decision-Level Fusion (PDLF) module, including two parts. One part utilizes the optimal global and local metrics to obtain a projection value. The other part combines the projection value with an intermediate score which is obtained through the fusion of unimodal and multimodal representations to obtain the final sentiment prediction score. Extensive experiments on the CMU-MOSI and CMU-MOSEI datasets demonstrate that our method is significantly superior to state-of-the-art methods on various evaluation indicators and parameter count, by effectively learning fine-grained emotional boundaries. The code will be made open-source if the paper is accepted.

Cinematic Audio Source Separation (CASS) aims to decompose mixed film audio into speech, music, and sound effects, enabling applications like dubbing and remastering. Existing CASS approaches are audio-only, overlooking the inherent audio-visual nature of films, where sounds often align with visual cues. We present the first framework for audio-visual CASS (AV-CASS), leveraging visual context to enhance separation quality. Our method formulates CASS as a conditional generative modeling problem using conditional flow matching, enabling multimodal audio source separation. To address the lack of cinematic datasets with isolated sound tracks, we introduce a training data synthesis pipeline that pairs in-the-wild audio and video streams (e.g., facial videos for speech, scene videos for effects) and design a dedicated visual encoder for this dual-stream setup. Trained entirely on synthetic data, our model generalizes effectively to real-world cinematic content and achieves strong performance on synthetic, real-world, and audio-only CASS benchmarks.

Despite progress in multimodal sarcasm detection, existing datasets and methods predominantly focus on single-image scenarios, overlooking potential semantic and affective relations across multiple images. This leaves a gap in modeling cases where sarcasm is triggered by multi-image cues in real-world settings. To bridge this gap, we introduce MMSD3.0, a new benchmark composed entirely of multi-image samples curated from tweets and Amazon reviews. We further propose a Cross-Image Reasoning Model (CIRM), integrating a Dual-Stage Bridge Module and Relevance-Guided Fusion Module to model inter-image dependencies and cross-modal correspondences. Complementarily, we establish a comprehensive suite of strong and representative baselines and conduct extensive experiments, showing that MMSD3.0 is an effective and reliable benchmark that better reflects real-world conditions. Moreover, CIRM demonstrates state-of-the-art performance across MMSD, MMSD2.0, and MMSD3.0, validating its effectiveness in both single-image and multi-image scenarios.

Vision-language pre-training (VLP) has achieved remarkable performance across diverse multimodal learning (MML) tasks. Recently, many efforts have focused on reconstructing missing modalities to improve the adaptability of VLP models in incomplete MML scenarios. However, these approaches overlook the learning imbalance under severe missing-modality conditions, i.e., the optimization process is dominated by reconstructed samples, thereby weakening complete-sample representations. In this paper, we propose a novel ANchor-guided Gradient Alignment (ANGA) framework to address these issues. Specifically, we first retrieve similar instances to reconstruct the missing modalities, thereby alleviating information deficiency. We then introduce an entropy-driven curriculum that progressively integrates reliable reconstructed samples with complete ones to form an optimization anchor, which guides gradient alignment to mitigate learning imbalance. Furthermore, we design a semantic-enhanced adapter that leverages the retrieved instances to generate dynamic prompts, further enhancing the robustness of the VLP model. Extensive experiments on widely used datasets demonstrate the superiority of ANGA over state-of-the-art (SOTA) baselines across various missing-modality scenarios.

Discrete video VAEs underpin modern text-to-video generation and video understanding systems, yet existing tokenizers typically learn visual codebooks at a single scale with limited vocabularies and shallow language supervision, leading to poor cross-modal alignment and zero-shot transfer. We introduce PyraTok, a language-aligned pyramidal tokenizer that learns semantically structured discrete latents across multiple spatial-temporal resolutions. PyraTok builds on a pretrained video VAE and a novel Language aligned Pyramidal Quantization (LaPQ) module that discretizes encoder features at several depths using a shared large binary codebook, yielding compact yet expressive video token sequences. To tightly couple visual tokens with language, PyraTok jointly optimizes multi-scale text-guided quantization and a global autoregressive objective over the token hierarchy. Across ten benchmarks, PyraTok delivers state-of-the-art (SOTA) video reconstruction, consistently improves text-to-video quality, and sets new SOTA zero-shot performance on video instance segmentation, temporal action localization, and video understanding, scaling robustly to up to 4K/8K resolutions.

Though rectified flow models have achieved remarkable performance in image, video, and 3D generation, their practical deployments are challenged by slow inference speeds. Previous acceleration methods rely on caching and reusing, neglecting the growing mismatch between static cached values and evolving input, leading to reduced generated content fidelity.This work proposes Velocity Decomposition and Estimation (VDE), a training-free acceleration method that shifts the paradigm from caching-and-reusing to decomposing-and-estimating.VDE periodically anchors the model’s state with a full forward pass and estimates subsequent outputs analytically. VDE first decomposes the model’s velocity output into components parallel and orthogonal to the input, then exploiting the temporal predictability of the components' coefficients and the consistency of the orthogonal direction for precise, input-adaptive estimation at each timestep.Extensive experiments on image and video generation tasks demonstrate that VDE achieves up to 2.04-3.22× acceleration with minimal loss in visual quality. For example, in image generation, VDE achieves a 2.21× speedup while preserving nearly identical visual quality, outperforming the best baseline by 19.5% in SSIM, 30.3% in PSNR, and reducing LPIPS by 55.4%.

Instruction-based image editing offers a powerful and intuitive way to manipulate images through natural language. Yet, relying solely on text instructions limits fine-grained control over the extent of edits. We introduce Kontinuous Kontext, an instruction-driven editing model that provides a new dimension of control over edit strength, enabling users to adjust edits gradually from no change to a fully realized result in a smooth and continuous manner. Kontinuous Kontext extends a state-of-the-art image editing model to accept an additional input, a scalar edit strength which is then paired with the edit instruction, enabling explicit control over the extent of the edit. To inject this scalar information, we train a lightweight projector network that maps the input scalar and the edit instruction to coefficients in the model's modulation space. For training our model, we synthesize a diverse dataset of image-edit-instruction-strength quadruplets using existing generative models, followed by a filtering stage to ensure quality and consistency. Kontinuous Kontext provides a unified approach for fine-grained control over edit strength for instruction driven editing from subtle to strong across diverse operations such as stylization, attribute, material, background, and shape changes, without requiring attribute-specific training.

Existing video editing methods face a critical trade-off: expert models offer precision but rely on task-specific priors like masks, hindering unification; conversely, unified temporal in-context learning models are mask-free but lack explicit spatial cues, leading to weak instruction-to-region mapping and imprecise localization. To resolve this conflict, we propose VideoCoF, a novel Chain-of-Frames approach inspired by Chain-of-Thought reasoning. VideoCoF enforces a "seeing, reasoning, then editing" procedure by compelling the video diffusion model to first predict reasoning tokens (edit-region latents) before generating the target video tokens. This explicit reasoning step removes the need for user-provided masks while achieving precise instruction-to-region alignment and fine-grained video editing. Furthermore, we introduce a RoPE alignment strategy that leverages these reasoning tokens to ensure motion alignment and enable length extrapolation beyond the training duration. We demonstrate that with a minimal data cost of only 50k video pairs, VideoCoF achieves state-of-the-art performance on VideoCoF-Bench, validating the efficiency and effectiveness of our approach.

Visual Autoregressive (VAR) models have demonstrated competitive performance with diffusion models in image generation by adopting a "next-scale" prediction paradigm that significantly reduces inference steps. However, VAR's progressive multi-scale generation leads to severe memory overhead due to KV cache accumulation across all scales, limiting practical deployment. Existing solutions either require training and deploying multiple specialized models or sacrifice generation quality.We observe a critical scale-depth asymmetric dependency in VAR: small scales (low-resolution tokens) are highly sensitive to network depth and require deep layers to capture global semantic information, while large scales (high-resolution tokens) exhibit remarkable robustness to depth reduction.Motivated by this insight, we propose VARiant, a unified supernet framework that enables dynamic depth adjustment within a single model through parameter sharing. Our approach employs an even-spacing layer selection strategy to construct quality-preserving subnetworks, and introduces a dynamic-ratio progressive training strategy that gradually transitions from joint optimization (full network to subnetwork ratio 2:8) to subnetwork optimization (ratio 10:0), effectively resolving the inherent optimization conflicts between full network and subnetworks in supernet training.Extensive experiments on ImageNet demonstrate that our method achieves Pareto-optimal trade-offs between generation quality and inference efficiency: by using full depth (30 layers) for the first 7 scales and a 16-layer subnetwork (50% depth) for subsequent scales, we obtain 50% cache reduction and 1.8× inference speedup with minimal quality loss (FID increases by only 0.3).Unlike approaches requiring deployment of multiple models, our single-model solution eliminates deployment complexity, supports zero-cost runtime depth switching, and seamlessly integrates into standard transformer inference frameworks, making it highly practical for resource-constrained scenarios.

Text-driven instruction-based video editing in complex scenes remains challenging: purely textual prompts often fail to capture precise spatial relationships and physical constraints, resulting in target ambiguity and physically implausible outcomes. To address this, we propose a plan--guide--edit framework that explicitly bridges semantic intent and spatial execution. In our framework, a Chain-of-Thought (CoT)-enhanced multimodal large language model (MLLM) serves as a planner, performing structured reasoning over the video and instructions to derive a precise sequence of bounding boxes and attribute-enriched editing directives. These spatial priors then guide a box-conditioned mask generator, transforming ambiguous global retrieval into localized, context-aware refinement and producing masks that more accurately capture object scale, contact relationships, and placement. Building on these spatial and semantic signals, a diffusion-based editor integrates the masks, enriched instructions, and frame features to render high-fidelity edits that remain temporally coherent and spatially well aligned. Trained first in a modular manner and then jointly, our framework achieves superior performance with reduced data requirements, delivering precise localization in scenes with multiple similar objects and physically consistent object additions, and extensive experiments demonstrate state-of-the-art performance over multiple strong baseline methods.

Instruction-based video editing promises to democratize content creation, yet its progress is severely hampered by the scarcity of large-scale, high-quality training data. We introduce Ditto, a holistic framework designed to tackle this fundamental challenge. At its heart, Ditto features a novel data generation pipeline that fuses the creative diversity of a leading image editor with an in-context video generator, overcoming the limited scope of existing models. To make this process viable, our framework resolves the prohibitive cost-quality trade-off by employing an efficient, distilled model architecture augmented by a temporal enhancer, which simultaneously reduces computational overhead and improves temporal coherence. Finally, to achieve full scalability, this entire pipeline is driven by an intelligent agent that crafts diverse instructions and rigorously filters the output, ensuring quality control at scale. Using this framework, we invested over 12,000 GPU-days to build Ditto-1M, a new dataset of one million high-fidelity video editing examples. We trained our model, Editto, on Ditto-1M with a curriculum learning strategy. The results demonstrate superior instruction-following ability and establish a new state-of-the-art in instruction-based video editing. We will release our dataset and models for reproducibility.

Existing generative models, such as diffusion and auto-regressive networks, are inherently static, relying on a fixed set of pretrained parameters to handle all inputs. In contrast, humans flexibly adapt their internal generative representations to each perceptual or imaginative context. Inspired by this capability, we introduce Composer, a new paradigm for adaptive generative modeling based on test-time instance-specific parameter composition. Composer generates input-conditioned parameter adaptations at inference time, which are injected into the pretrained model’s weights, enabling per-input specialization without fine-tuning or retraining. Adaptation occurs once prior to multi-step generation, yielding higher-quality, context-aware outputs with minimal computational and memory overhead. Experiments show that Composer substantially improves performance across diverse generative models and use cases, including lightweight/quantized models and test-time scaling. By leveraging input-aware parameter composition, Composer establishes a new paradigm for designing generative models that dynamically adapt to each input, moving beyond static parameterization. The code will be available at \url{https://anonymous.4open.science/r/Composer-IPC}.

MeanFlow promises high-quality generative modeling in few steps, by jointly learning instantaneous and average velocity fields. Yet, the underlying training dynamics remain unclear. We analyze the interaction between the two velocities and find: (i) well-established instantaneous velocity is a prerequisite for learning average velocity; (ii) learning of instantaneous velocity benefits from average velocity when the temporal gap is small, but degrades as the gap increases; and (iii) task-affinity analysis indicates that smooth learning of large-gap average velocities, essential for one-step generation, depends on the prior formation of accurate instantaneous and small-gap average velocities. Guided by these observations, we design an effective training scheme that accelerates the formation of instantaneous velocity, then shifts emphasis from short- to long-interval average velocity. Our enhanced MeanFlow training yields faster convergence and significantly better few-step generation: With the same DiT-XL backbone, our method reaches an impressive FID of 2.87 on 1-NFE ImageNet 256x256, compared to 3.43 for the conventional MeanFlow baseline. Alternatively, our method matches the performance of the MeanFlow baseline with 2.5x shorter training time, or with a smaller DiT-L backbone.

Unified multi-modal understanding/generative models have shown improved image editing performance by incorporating fine-grained understanding into their Chain-of-Thought (CoT) process. However, a critical question remains underexplored: what forms of CoT and training strategy can jointly enhance both the understanding granularity and generalization? To address this, we propose Meta-CoT, a paradigm that performs a two-level decomposition of any single-image editing operation with two key properties: (1) Decomposability. We observe that any editing intention can be represented as a triplet — (task, target, required understanding ability). Inspired by this, Meta-CoT decomposes both the editing task and the target, generating task-specific CoT and traversing editing operations on all targets. This decomposition enhances the model's understanding granularity of editing operations and guides it to learn each element of the triplet during training, substantially improving the editing capability.(2) Generalizability. In the second decomposition level, we further break down editing tasks into five fundamental meta-tasks. We find that training on these five meta-tasks, together with the other two elements of the triplet, is sufficient to achieve strong generalization across diverse, unseen editing tasks.To further align the model's editing behavior with its CoT reasoning, we introduce the CoT–Editing Consistency Reward, which encourages more accurate and effective utilization of CoT information during editing. Experiments demonstrate that our method achieves an overall 15.8\% improvement across 21 editing tasks, and generalizes effectively to unseen editing tasks when trained on only a small set of meta-tasks. Our code, benchmark, and model will be released publicly.

Video generation using recurrent architectures offers compelling efficiency advantages over attention-based transformers, particularly for long-sequence generation. However, chunked processing in recurrent models creates temporal discontinuities that harm long-range consistency. We introduce two complementary memory mechanisms to address this challenge at different granularities: \textbf{(1) Context Memory} maintains persistent global context within attention chunks through learnable \textit{sink columns} and \textit{boundary buffers}, adding only 150K parameters (\textless 0.1\% overhead); \textbf{(2) Latent Context-as-Memory (LCaM)} extends memory across video segments by storing and retrieving historical latent embeddings, enabling cross-segment consistency without requiring camera annotations or frame reconstruction. Applied to Generalized Spatial-temporal Propagation Networks (GSTPN), our dual-memory approach achieves \textbf{1.54$\times$ faster} inference than attention-based transformers, while excelling in visual quality metrics. Our approach is particularly effective for knowledge distillation scenarios where only pre-extracted latent embeddings are available. This work demonstrates compelling efficiency-quality trade-offs for practical long video generation.

Transformers have emerged as a universal backbone across 3D perception, video generation, and world models for autonomous driving and embodied AI, where understanding camera geometry is essential for grounding visual observations in three-dimensional space. However, existing camera encoding methods often rely on simplified pinhole assumptions, restricting generalization across the diverse intrinsics and lens distortions in real-world cameras. We introduce Relative Ray Encoding, a geometry-consistent representation that unifies complete camera information, including 6-DoF poses, intrinsics, and lens distortions. To evaluate its capability under diverse controllability demands, we adopt camera-controlled text-to-video generation as a testbed task. Within this setting, we further identify pitch and roll as two components effective for Absolute Orientation Encoding, enabling full control over the initial camera orientation. Together, these designs form UCPE (Unified Camera Positional Encoding), which integrates into a pretrained video Diffusion Transformer through a lightweight spatial attention adapter, adding less than 1% trainable parameters while achieving state-of-the-art camera controllability and visual fidelity. To facilitate systematic training and evaluation, we construct a large video dataset covering a wide range of camera motions and lens types. Extensive experiments validate the effectiveness of UCPE in camera-controllable video generation and highlight its potential as a general camera representation for Transformers across future multi-view, video, and 3D tasks.

Recent advances in diffusion models (DMs) have achieved exceptional visual quality in image editing tasks. However, the global denoising dynamics of DMs inherently conflate local editing targets with the full-image context, leading to unintended modifications in non-target regions. In this paper, we shift our attention beyond DMs and turn to Masked Generative Transformers (MGTs) as an alternative approach to tackle this challenge. By predicting multiple masked tokens rather than holistic refinement, MGTs exhibit a localized decoding paradigm that endows them with the inherent capacity to explicitly preserve non-relevant regions during the editing process. Building upon this insight, we introduce the first MGT-based image editing framework, termed EditMGT. We first demonstrate that MGT's cross-attention maps provide informative localization signals for localizing edit-relevant regions and devise a multi-layer attention consolidation scheme that refines these maps to achieve fine-grained and precise localization. On top of these adaptive localization results, we introduce region-hold sampling, which restricts token flipping within low-attention areas to suppress spurious edits, thereby confining modifications to the intended target regions and preserving the integrity of surrounding non-target areas. To train EditMGT, we construct Crisp-2M, a high-resolution (>1024) dataset spanning seven diverse editing categories. Without introducing additional parameters, we adapt a pre-trained text-to-image MGT into an image editing model through attention injection. Extensive experiments across four standard benchmarks demonstrate that, with fewer than 1B parameters, our model achieves state-of-the-art image similarity performance while enabling faster editing. Moreover, it delivers comparable or superior editing quality, with improvements of 3.6% and 17.6% on style change and style transfer tasks, respectively.

Generalizable Neural Radiance Fields (GeNeRF) enable high-quality scene reconstruction from a limited number of views and can generalize to unseen scenes. However, in real-world environments, transient distractors disrupt structural consistency across views, leading to deviated supervision signals and degraded reconstruction quality. Existing distractor-free NeRF methods rely on per-scene optimization and they estimate uncertainty from per-view reconstruction errors to remove distractors, but this is unreliable to GeNeRF, because it may misjudge inconsistent static structures from source views as distractors. To address this issue, we propose MUGeNeRF: a multi-view uncertainty-guided distractor-aware GeNeRF method, aim to effectively alleviate GeNeRF's robust modeling challenges in dynamic scenes with transient distractions. We explicitly decompose distractor awareness into two complementary uncertainty modeling tasks: Source-view uncertainty, serving as a transferable prior during the feed-forward process, captures structural inconsistencies across source views caused by viewpoint changes or dynamic factors; Target-view uncertainty focuses on observation anomalies caused by transient changes to infer distractor spatial distribution. These two uncertainties are integrated into a heteroscedastic reconstruction loss that guides adaptive supervision weighting, boosting the model's capability to detect and suppress distractors, and enabling more robust geometric modeling. To our knowledge, this is the first attempt to explore GeNeRF modeling with in scenes with transient distractors. Extensive experiments demonstrate that our method not only outperforms existing GeNeRF approaches but also rivals the performance of scene-specific distractor-free NeRFs.

Recent advances in generative AI have dramatically improved photorealistic image synthesis, yet they fall short for studio-level multi-object compositing. This task demands simultaneous (i) near‑perfect preservation of each item’s identity, (ii) precise background and color fidelity, (iii) layout and design elements control, and (iv) complete, appealing displays showcasing all objects. However, current state-of-the-art models often alter object details, omit or duplicate objects, and produce layouts with incorrect relative sizing or inconsistent item presentations. To bridge this gap, we introduce PLACID, a framework that transforms a collection of object images into an appealing multi-object composite. Our approach makes two main contributions. First, we leverage a pretrained image-to-video (I2V) diffusion model with text control to preserve objects consistency, identities and background details by exploiting temporal priors from videos. Second, we propose a novel data curation strategy that generates synthetic sequences where randomly placed objects smoothly move to their target positions. This synthetic data aligns with the video model’s temporal priors during training. At inference, objects initialized at random positions consistently converge into coherent layouts guided by text, with the final frame serving as the composite image. Extensive quantitative evaluations and user studies demonstrate that PLACID surpasses state-of-the-art methods in multi-object compositing, achieving superior identity, background, and color preservation, with less omitted objects and visually appealing results.

In this paper, we introduce Object-WIPER, a training-free framework for removing dynamic objects and their associated visual effects from videos, and inpainting them with semantically consistent and temporally coherent content. Our approach leverages a pre-trained text-to-video diffusion transformer (DiT). Given an input video, a user-provided object mask, and query tokens describing the target object and its effects, we localize relevant visual tokens via visual-text cross-attention and visual self-attention. This produces an intermediate effect mask that we fuse with the user mask to obtain a final foreground token mask to replace. We first invert the video through the DiT to obtain structured noise, then reinitialize the masked tokens with Gaussian noise while preserving background tokens. During denoising, we copy values for the background tokens saved during inversion to maintain scene fidelity. To address the lack of suitable evaluation, we introduce a new object removal metric that rewards temporal consistency among foreground tokens across consecutive frames, coherence between foreground and background tokens within each frame, and dissimilarity between the input and output foreground tokens. Experiments on DAVIS and a newly curated real-world associated effect benchmark WIPER-Bench show that Object-WIPER surpasses both training-based and training-free baselines in terms of the metric, achieving clean removal and temporally stable reconstruction without any retraining. Our new benchmark, source code, and pre-trained models will be publicly available.

Virtual try-on (VTON) has recently achieved impressive visual fidelity, but most existing systems require uploading personal photos to cloud-based GPUs, raising privacy concerns and limiting on-device deployment. To address this, we present \textsc{Mobile-VTON}, a high-quality, privacy-preserving framework that enables fully offline virtual try-on on commodity mobile devices using only a single user image and a garment image. \textsc{Mobile-VTON} introduces a modular TeacherNet--GarmentNet--TryonNet (TGT) architecture that integrates knowledge distillation, garment-conditioned generation, and garment alignment into a unified pipeline optimized for on-device efficiency. Within this framework, we propose a Feature-Guided Adversarial (FGA) Distillation strategy that combines teacher supervision with adversarial learning to better match real-world image distributions. GarmentNet is trained with a trajectory-consistency loss to preserve garment semantics across diffusion steps, while TryonNet uses latent concatenation and lightweight cross-modal conditioning to enable robust garment-to-person alignment without large-scale pretraining. By combining these components, \textsc{Mobile-VTON} achieves high-fidelity generation with low computational overhead. Experiments on VITON-HD and DressCode at $1024{\times}768$ show that it matches or outperforms strong server-based baselines while running entirely offline. These results demonstrate that high-quality VTON is not only feasible but also practical on-device, offering a secure solution for real-world applications.

Recent visual autoregressive (AR) models have shown promising capabilities in text-to-image generation, operating in a manner similar to large language models. While test-time computation scaling has brought remarkable success in enabling reasoning-enhanced outputs for challenging natural language tasks, its adaptation to visual AR models remains unexplored and poses unique challenges. Naively applying test-time scaling strategies such as Best-of-$N$ can be suboptimal: they consume full-length computation on erroneous generation trajectories, while the raster-scan decoding scheme lacks a blueprint of the entire canvas, limiting scaling benefits as only a few prompt-aligned candidates are generated. To address these, we introduce GridAR, a test-time scaling framework designed to elicit the best possible results from visual AR models. GridAR employs a grid-partitioned progressive generation scheme in which multiple partial candidates for the same position are generated within a canvas, infeasible ones are pruned early, and viable ones are fixed as anchors to guide subsequent decoding. Coupled with this, we present a layout-specified prompt reformulation strategy that inspects partial views to infer a feasible layout for satisfying the prompt. The reformulated prompt then guides subsequent image generation to mitigate the blueprint deficiency. Together, GridAR achieves higher-quality results under limited test-time scaling: with $N{=}4$, it even outperforms Best-of-$N$ ($N{=}8$) by 14.4% on T2I-CompBench++ while reducing cost by 25.6%. It also generalizes to autoregressive image editing, showing comparable edit quality and a 13.9% gain in semantic preservation on PIE-Bench over larger-$N$ baselines. The source code will be publicly released.

Recent advances in visual generative models have enabled high-fidelity image editing guided by human instructions. However, these models often struggle with complex instructions involving combinatorial editing operations or inter-step dependencies. This difficulty stems from the limitations of two canonical paradigms: (1) single-turn editing, which attempts to apply all instructed edits in one pass, often fails to parse the complex instruction accurately and causes undesired edits; and (2) sequential editing can decompose the task into simpler steps but suffers from compounding errors introduced by the sequential execution, leading to low-fidelity results. To derive a robust solution for complex image editing, we examine editing behaviors of different paradigms under a unified in-context editing framework, and study how the benefits of sequential decomposition can be balanced against its error-accumulation drawbacks. We further develop a synthetic data pipeline that constructs editing tasks of varying instruction complexity, allowing us to curate a large-scale editing dataset with high-quality decomposed sequences. By finetuning on synthetic data, we discovered that with properly designed editing paradigms, sequential decomposition yields robust improvements even as task complexity increases. Furthermore, the decomposition skills learned from synthetic tasks can transfer to real images by co-training with real-world editing data, demonstrating the promise of sim-to-real generalization for tackling complex image editing across broader domains.

Recent rapid advancement of generative models has significantly improved the fidelity and accessibility of AI-generated synthetic images. While enabling various innovative applications, the unprecedented realism of these synthetics makes them increasingly indistinguishable from authentic photographs, posing serious security risks, such as media credibility and content manipulation. Although extensive efforts have been dedicated to detecting synthetic images, most existing approaches suffer from poor generalization to unseen data due to their reliance on model-specific artifacts or low-level statistical cues. In this work, we identify a previously unexplored distinction that real images maintain consistent semantic attention and structural coherence in their latent representations, exhibiting more stable feature transitions across network layers, whereas synthetic ones present discernible distinct patterns. Therefore, we propose a novel approach termed latent transition discrepancy (LTD), which captures the inter-layer consistency differences of real and synthetic images. LTD adaptively identifies the most discriminative layers and assesses the transition discrepancies across layers. Benefiting from the proposed inter-layer discriminative modeling, our approach exceeds the base model by 14.35\% in mean Acc across three datasets containing diverse GANs and DMs. Extensive experiments demonstrate that LTD outperforms recent state-of-the-art methods, achieving superior detection accuracy, generalizability, and robustness.

Physically Plausible Video Generation (PPVG) has emerged as a promising avenue for modeling real-world physical phenomena. PPVG requires an understanding of commonsense knowledge, which remains a challenge for video diffusion models. Current approaches leverage commonsense reasoning capability of large language models to embed physical concepts into prompts. However, generation models often render physical phenomena as a single moment defined by prompts, due to the lack of conditioning mechanisms for modeling causal progression. In this paper, we view PPVG as generating a sequence of causally connected and dynamically evolving events. To realize this paradigm, we design two key modules: (1) Physics-driven Event Chain Reasoning. This module decomposes the physical phenomena described in prompts into multiple elementary event units, leveraging chain-of-thought reasoning. To mitigate causal ambiguity, we embed physical formulas as constraints to impose deterministic causal dependencies during reasoning. (2) Transition-aware Cross-modal Prompting (TCP). To maintain continuity between events, this module transforms causal event units into temporally aligned vision-language prompts. It summarizes discrete event descriptions to obtain causally consistent narratives, while progressively synthesizing visual keyframes of individual events by interactive editing. Comprehensive experiments on PhyGenBench and VideoPhy benchmarks demonstrate that our framework achieves superior performance in generating physically plausible videos across diverse physical domains. \textcolor[RGB]{237,0, 140}{Our code will be released soon}.

Recent research in long-form video generation has shifted from bidirectional to autoregressive models, yet these methods commonly suffer from error accumulation and a loss of long-term coherence. While attention sink frames have been introduced to mitigate this performance decay, they often induce a critical failure mode we term sink-collapse: the generated content repeatedly reverts to the sink frame, resulting in abrupt scene resets and cyclic motion patterns. Our analysis reveals that sink-collapse originates from an inherent conflict between the periodic structure of Rotary Position Embedding (RoPE) and the multi-head attention mechanisms prevalent in current generative models. To address it, we propose a lightweight, training-free approach that effectively suppresses this behavior by introducing multi-head RoPE jitter that breaks inter-head attention homogenization and mitigates long-horizon collapse. Extensive experiments show that our method successfully alleviates sink-collapse while preserving generation quality. To the best of our knowledge, this work achieves the first demonstration of real-time, streaming, and infinite-length video generation with little quality decay. As an illustration of this robustness, we generate continuous videos up to 12 hours in length, which, to our knowledge, is among the longest publicly demonstrated results in streaming video generation.

Video motion transfer aims to generate a target video that inherits motion patterns from a source video while rendering new scenes. Existing training-free approaches focus on constructing motion guidance based on the intermediate outputs of pre-trained T2V models, which results in heavy computational overhead and limited flexibility. In this paper, we present FlowMotion, a novel training-free framework that enables efficient and flexible motion transfer by directly leveraging the predicted outputs of flow-based T2V models. Our key insight is that early latent predictions inherently encode rich temporal information. Motivated by this, we propose flow guidance, which extracts motion representations based on latent predictions to align motion patterns between source and generated videos. We further introduce a velocity regularization strategy to stabilize optimization and ensure smooth motion evolution. By operating purely on model predictions, FlowMotion achieves superior time and resource efficiency as well as competitive performance compared with state-of-the-art methods.

Flow Matching has limited ability in achieving one-step generation due to its reliance on learned curved trajectories. Previous studies have attempted to address this limitation by either modifying the coupling distribution to prevent interpolant intersections or introducing consistency and mean-velocity modeling to promote straight trajectory learning. However, these approaches often suffer from discrete approximation errors, training instability, and convergence difficulties. To tackle these issues, in the present work, we propose \textbf{S}traight \textbf{V}ariational \textbf{F}low \textbf{M}atching (\textbf{S-VFM}), which integrates a variational latent code representing the ``generation overview'' into the Flow Matching framework. \textbf{S-VFM} explicitly enforces trajectory straightness, ideally producing linear generation paths. The proposed method achieves competitive performance across three challenge benchmarks and demonstrates advantages in both training and inference efficiency compared with existing methods.

Recent unified models such as Bagel demonstrate that paired image–edit data can effectively align multiple visual tasks within a single diffusion transformer. However, these models remain limited to single-condition inputs and lack the flexibility needed to synthesize results from multiple heterogeneous sources. We present SIGMA (Selective-Interleaved Generation with Multi-Attribute Tokens), a unified post-training framework that enables interleaved multi-condition generation within diffusion transformers. SIGMA introduces selective multi-attribute tokens, including style, content, subject and identity tokens, which allow the model to interpret and compose multiple visual conditions in an interleaved text–image sequence. Through post-training on the Bagel unified backbone with 700K interleaved examples, SIGMA supports compositional editing, selective attribute transfer and fine-grained multimodal alignment. Extensive experiments show that SIGMA improves controllability, cross-condition consistency and visual quality across diverse editing and generation tasks, with substantial gains over Bagel on compositional tasks.

Benefiting from the powerful generative priors of diffusion models, diffusion-based real-world image super-resolution (Real-ISR) methods have demonstrated impressive performance.To achieve efficient Real-ISR, several recent works have designed one-step diffusion-based models.Howerver, unmediatedly feeding LR into a diffusion model creates a distributional gap with the model's original input.A straightforward approach to reduce the distribution gap is to introduce noise to the LR latents. However, directly adding noise inevitably corrupts the content of the LR images.In this study, we propose \textbf{DNF‑SR}, a \textbf{D}ual‑input and \textbf{N}egative‑aware \textbf{F}eature fine‑tuning method for Real-ISR.Specifically, we use a \textbf{dual-input} strategy that concatenates the original LR image with the noisy LR input and feeds them into a diffusion-based image editing model, ensuring both high-fidelity one-step super-resolution and improved perceptual and content consistency.Additionally, the noise present in the noisy LR input introduces randomness and diversity into the outputs. We exploit this property and propose a post-training optimization method, Negative-aware Feature Fine-Tuning (NF²T), which guides the model toward producing higher-quality results.NF$^2$T classifies multiple outputs into positive and negative subsets and then defines implicit policy improvement directions in both the image and feature spaces, thereby further enhancing the stability of the optimization.Extensive experiments show that DNF-SR outperforms other methods.Code will be released.

Diffusion models have recently achieved remarkable success in real-world image super-resolution (ISR), typically balancing a trade-off between fidelity (i.e., similarity to HR images) and realism (i.e., perceptual naturalness). To better account for subjective preferences in image quality, controllable diffusion-based methods have been explored, allowing personalized adjustment of this trade-off via tunable parameters. While existing controllable methods have shown effective control, they operate in the latent space and require repeated network inference during adjustment, eventually limiting their practicality. In this paper, we propose IFCSR, a simple yet practical approach for one-step diffusion-based real-world ISR that enables inference-free control between fidelity and realism. The key idea behind IFCSR is to design a controllable model that adjusts the fidelity-realism trade-off in the image space, rather than in the latent space. Such an image-space control allows users to seamlessly adjust the trade-off without extra inference after an initial inference of fidelity- and realism-specific images. We further introduce a two-stage training scheme and specialized losses that encourage the controllable space to span a broad spectrum of fidelity and realism. Our method achieves quality competitive with state-of-the-art models while providing a practical advantage through inference-free control.

Omnidirectional image super-resolution (ODISR) remains challenging due to extreme magnification factors (e.g., 8×, 16×) and projection-specific distortions, which degrade edge integrity and limit model performance. This paper proposes an edge-focused framework combined with spherical geometric augmentation to address these issues. Our approach includes an Edge Focused Block (EFB) that integrates spatial-channel attention via Edge Enhanced and Refined Blocks, strengthening edge feature capture and optimization. We also design an Edge-Aware Multi-Scale (EAM) pipeline, leveraging shallow convolutions for initial feature extraction, local modules for deep mining, and a Global Integration Block for multi-scale aggregation, ensuring coherent edge reconstruction in distorted regions. To mitigate data scarcity, we introduce a rotation-translation augmentation strategy based on spherical projections, expanding datasets while preserving scene continuity. Extensive experiments show our method outperforms state-of-the-art approaches on public datasets.

Diffusion-based generative models have achieved remarkable success in real-world image super-resolution (SR). With tiled diffusion techniques, these models can produce high-resolution images that exceed their native-supported resolution. However, the quality of such high-resolution (e.g $2048^2$) outputs often remains extremely poor, primarily due to two factors we consider: the image upsampling ratio (e.g $\times8$) exceeding the model's native-supported upsampling ratio (e.g $\times4$), and the model's native-supported resolution. In practice, training a native high-resolution model requires larger architectures, which incur significant computational overhead and GPU memory costs, making it hard on limited-resource equipment. Thus, we present \textbf{TUDSR}, a \textbf{T}wice \textbf{U}psampling–\textbf{D}iffusion framework for higher SR. The TUDSR framework mainly consists of two stages: the first involves training at $R$-resolution, and the second introduces a looped chunk-based training strategy at $NR$-resolution. Each stage adapts a one-step GAN architecture comprising a generator and a discriminator. Based on SD2.1-base, we develop TUDSR-S, which achieves state-of-the-art performance across multiple benchmarks. Extensive experiments further demonstrate that TUDSR-S generates high-quality images at the resolutions of $1024^2$ and even $2048^2$, significantly outperforming existing approaches.

Video Super-Resolution (VSR) fundamentally struggles with a critical trade-off: single-step models offer unmatched efficiency but often lack the high-frequency detail, creativity, and visual quality of their multi-step diffusion counterparts, which are computationally prohibitive for practical use. In this paper, we propose PS-SR, a novel "pseudo" single-step VSR framework that transcends this trade-off through a computationally asymmetric sampling pipeline. The key to PS-SR lies in its speculative diffusion mechanism: a powerful base model performs only a single, comprehensive sampling step, establishing the global structure and content fidelity, after which a lightweight draft model, directly augmented by the base model's features, speculatively performs subsequent refinements. Crucially, we further enforce a frequency-domain update rule that constrains these refinements to exclusively inject high-frequency details, preserving the foundational low-frequency content and preventing semantic drift across sampling steps. By doing so, PS-SR creates the "illusion" of a single-step model—delivering the similar inference speeds and input-output content consistency—while achieving the visual richness and creativity typically reserved for costly multi-step generative models. We demonstrate that our "pseudo-single-step" paradigm achieves state-of-the-art quality with a comparable speed to single-step models, paving the way for real-time, high-fidelity video enhancement.

Image Super-Resolution (SR) aims to reconstruct high-resolution images from degraded low-resolution inputs. While diffusion-based SR methods offer powerful generative capabilities, their performance heavily depends on how semantic priors are structured and integrated into the generation process. Existing approaches often rely on entangled or coarse-grained priors that mix global layout with local details, or conflate structural and textural cues, thereby limiting semantic controllability and interpretability. In this work, we propose DTPSR, a novel diffusion-based SR framework that introduces disentangled textual priors along two complementary dimensions: spatial hierarchy (global vs. local) and frequency semantics (low- vs. high-frequency). By explicitly separating these priors, DTPSR enables the model to simultaneously capture scene-level structure and object-specific details with frequency-aware semantic guidance. The corresponding embeddings are injected via specialized cross-attention modules, forming a progressive generation pipeline that reflects the semantic granularity of visual content—from global layout to fine-grained textures. To support this paradigm, we construct DisText-SR, a large-scale dataset containing approximately 95,000 image-text pairs with carefully disentangled global, low-frequency, and high-frequency descriptions. To further enhance controllability and consistency, we adopt a multi-branch classifier-free guidance strategy with frequency-aware negative prompts to suppress hallucinations and semantic drift. Extensive experiments on synthetic and real-world benchmarks show that DTPSR achieves high perceptual quality, competitive fidelity, and strong generalization across diverse degradation scenarios.

Generative diffusion priors have recently achieved state-of-the-art performance in Natural Image Super-Resolution, demonstrating a powerful capability to synthesize photorealistic details. However, their direct application to Remote Sensing Image Super-Resolution (RSISR) reveals significant shortcomings. Remote sensing images present a unique challenge: ground objects often exhibit globally stochastic yet locally clustered patterns. This characteristic leads to highly imbalanced texture distributions, posing a significant hurdle to the model's spatial perception. To address this, we propose TexADiff, a novel framework that begins by estimating a Relative Texture Density Map (RTDM) that reflects the underlying texture distribution. TexADiff then leverages this RTDM in three synergistic ways: as an explicit spatial conditioning to guide the diffusion process, as a loss modulation term to prioritize texture-rich regions, and as a dynamic adapter for the sampling schedule. These modifications are designed to endow the model with explicit texture-aware capabilities. Experiments demonstrate that TexADiff achieves superior quantitative metrics. Furthermore, qualitative results show that our model generates faithful high-frequency details while effectively suppressing texture hallucinations. This improved reconstruction quality also results in significant gains in downstream task performance.

Diffusion model (DM) based Video Super-Resolution (VSR) approaches achieve impressive perceptual quality. However, they suffer from error accumulation, spatial artifacts, and a trade-off between perceptual quality and fidelity, primarily caused by inaccurate alignment and insufficient compensation between video frames. In this paper, within the DM-based VSR pipeline, we revisit the role of alignment and compensation between adjacent video frames and reveal two crucial observations: (a) the feature domain is better suited than the pixel domain for information compensation due to its stronger spatial and temporal correlations, and (b) warping at an upscaled resolution better preserves high-frequency information, but this benefit is not necessarily monotonic. Therefore, we propose a novel Densely Guided diffusion model with Aligned Features for Video Super-Resolution (DGAF-VSR), with an Optical Guided Warping Module (OGWM) to maintain high-frequency details in the aligned features and a Feature-wise Temporal Condition Module (FTCM) to deliver dense guidance in the feature domain. Extensive experiments on synthetic and real-world datasets demonstrate that DGAF-VSR surpasses state-of-the-art methods in key aspects of VSR, including perceptual quality (35.82\% DISTS reduction), fidelity (0.20 dB PSNR gain), and temporal consistency (30.37\% tLPIPS reduction).

Large-scale pre-trained diffusion models have been extensively adopted for real-world image Super-Resolution because of their powerful generative priors through textual guidance. However, when super-resolving high-resolution images with patch-wise inference strategy, most existing diffusion-based SR methods tend to suffer from over-generation, due to the misalignment between the global prompt from LR image and the incomplete semantic information of local patches during each inference step. On the other hand, most existing methods also failed to generate detailed texture in local patches due to the overemphasis on global generation capabilities in network designs and training strategies.To address this issue, we present DreamSR, a novel SR model that suppresses local over-generation and improves fine-detail synthesis, thereby achieving visually faithful results with ultra-high-quality details. Specifically, we propose a dual-branch MM-ControlNet, where the ControlNet generates local textual feature with patch-level prompts while the pre-trained DiT provides global textual feature with global prompts, thereby mitigating over-generation and ensuring semantic consistency across patches.We also design a comprehensive training strategy with stage-specific data processing pipelines and a Receptive-Field Enhancement strategy, enhancing the model’s capability to capture patch information and effectively restore local textures.Extensive experiments demonstrate that DreamSR outperforms state-of-the-art methods, providing high-quality SR results.

Diffusion-based approaches have recently driven remarkable progress in real-world image super-resolution (SR). However, existing methods still struggle to simultaneously preserve fine details and ensure high-fidelity reconstruction, often resulting in suboptimal visual quality. In this paper, we propose FiDeSR, a high-fidelity and detail-preserving one-step diffusion super-resolution framework. During training, we introduce a detail-aware weighting strategy that adaptively emphasizes regions where the model exhibits higher prediction errors. During inference, low- and high-frequency adaptive enhancers further refine the reconstruction without requiring model retraining, enabling flexible enhancement control. To further improve the reconstruction accuracy, FiDeSR incorporates a residual-in-residual noise refinement, which corrects prediction errors in the diffusion noise and enhances fine detail recovery. FiDeSR achieves superior real-world SR performance compared to existing diffusion-based methods, producing outputs with both high perceptual quality and faithful content restoration.

We present STCDiT, a video super-resolution framework built upon a pre-trained video diffusion model, aiming to restore structurally faithful and temporally stable videos from degraded inputs, even under complex camera motions. The main challenges lie in maintaining temporal stability during reconstruction and preserving structural fidelity during generation. To address these challenges, we first develop a motion-aware VAE reconstruction method that performs segment-wise reconstruction, with each segment clip exhibiting uniform motion characteristic, thereby effectively handling videos with complex camera motions. Moreover, we observe that the first-frame latent extracted by the VAE encoder in each clip, termed the anchor-frame latent, remains unaffected by temporal compression and retains richer spatial structural information than subsequent frame latents. We further develop an anchor-frame guidance approach that leverages structural information from anchor frames to constrain the generation process and improve structural fidelity of video features. Coupling these two designs enables the video diffusion model to achieve high-quality video super-resolution. Extensive experiments show that STCDiT outperforms state-of-the-art methods in terms of structural fidelity and temporal consistency.

Generating human motion that satisfies customized zero-shot goal functions, enabling applications such as controllable character animation and behavior synthesis for virtual agents, is a critical capability. While current approaches handle many unseen constraints, they fail on tasks with very challenging spatiotemporal restrictions, such as severe spatial obstacles or specified numbers of walking steps. To equip motion generators for these highly constrained tasks, we present a retrieval-guided method built on the training-free diffusion noise optimization framework. The key idea is to search within large motion datasets for guidance that can potentially satisfy difficult constraints. We introduce relational task parsing to group target constraints and identify the difficult ones to be handled by retrieved reference. A better initialization for diffusion noise is then obtained via a reward-guided mask that combines random noise with retrieved noise. By optimizing diffusion noise from this improved initialization, we successfully solve highly constrained generation tasks. By leveraging LLM for relational task parsing, the whole framework is further enabled to automatically reason for what to retrieve, improving the intelligence of moving agents under a training-free optimization scheme. Code will be released upon publication.

Video data is more cost-effective than motion capture data for learning 3D character motion controllers, yet synthesizing realistic and diverse behaviors directly from videos remains challenging. Previous approaches typically rely on off-the-shelf motion reconstruction techniques to obtain 3D trajectories for physics-based imitation. These reconstruction methods struggle with generalizability, as they either require 3D training data (potentially scarce) or fail to produce physically plausible poses, hindering their application to challenging scenarios like human-object interaction (HOI) or non-human characters. We tackle this challenge by introducing Mimic2DM, a novel motion imitation framework that learns the control policy directly and solely from widely available 2D keypoint trajectories extracted from videos. By minimizing the reprojection error, we train a general single-view 2D motion tracking policy capable of following arbitrary 2D reference motions in physics simulation, using only 2D motion data. The policy, when trained on diverse 2D motions captured from different or slightly different viewpoints, can further acquire 3D motion tracking capabilities by aggregating multiple views. Moreover, we develop a transformer-based autoregressive 2D motion generator and integrate it into a hierarchical control framework, where the generator produces high-quality 2D reference trajectories to guide the tracking policy. We show that the proposed approach is versatile and can effectively learn to synthesize physically plausible and diverse motions across a range of domains, including dancing, soccer dribbling, and animal movements, without any reliance on explicit 3D motion data.

Generating realistic human geometry animations remains a challenging task, as it requires modeling natural clothing dynamics with fine-grained geometric details under limited data. To address these challenges, we propose two novel designs. First, we propose a compact distribution-based latent representation that enables efficient and high-quality geometry generation. We improve upon previous work by establishing a more uniform mapping between SMPL and avatar geometries. Second, we introduce a generative animation model that fully exploits the diversity of limited motion data. We focus on short-term transitions while maintaining long-term consistency through an identity-conditioned design. These two designs formulate our method as a two-stage framework: the first stage learns a latent space, while the second learns to generate animations within this latent space. We conducted experiments on both our latent space and animation model. We demonstrate that our latent space produces high-fidelity human geometry surpassing previous methods (90% lower Chamfer Dist.). The animation model synthesizes diverse animations with detailed and natural dynamics (2.2 x higher user study score), achieving the best results across all evaluation metrics.

Controllable hand image generation aims to synthesize geometrically accurate images with consistent appearance. Recently, diffusion models have achieved remarkable success in image generation and have been applied for hand image synthesis. Through input-level fusion or feature-level modulation, existing methods inject control signals with fixed strength across all denoising timesteps. However, this static modulation ignores the progressive characteristic of the denoising process. In this paper, we reveal that the modulation process of control signals depends on the denoising state and the conditions complexity. Due to their distinct semantic distributions and information densities, it remains challenging to achieve effective interaction of these heterogeneous representations. To address this, we propose a Temporal and Content Co-Awareness Latent Diffusion method that designs a temporal- and content-driven modulation mechanism for controllable hand image generation. To achieve temporal and content co-awareness among the heterogeneous representations, we propose a query-based interaction mechanism designed to mitigate information redundancy and align semantic distributions. Leveraging this cross-domain interaction, the model infers the control information required at the current denoising state and dynamically adjust pose and appearance injection strengths. To obtain a stable appearance representation from multi-pose images of the same identity, we design the Pose-Invariant Appearance Encoder that captures both global appearance consistency and local texture details. Furthermore, we employ a feature orthogonal decomposition to mitigate pose leakage into appearance subspaces. Both quantitative and qualitative experimental results demonstrate the superiority of our method over the state of the arts.

Human motion analysis tasks, such as temporal 3D pose estimation, motion prediction, and motion in-betweening, play an essential role in computer vision. However, current paradigms suffer from severe fragmentation. First, the field is split between "perception" models that understand motion from video but only output text, and "generation" models that cannot perceive from raw visual input. Second, generative MLLMs are often limited to single-frame, static poses using dense, parametric SMPL models, failing to handle temporal motion. Third, existing motion vocabularies are built from skeleton data alone, severing the link to the visual domain. To address these challenges, we introduce Superman, a unified framework that bridges visual perception with temporal, skeleton-based motion generation. Our solution is twofold. First, to overcome the modality disconnect, we propose a Vision-Guided Motion Tokenizer. Leveraging the natural geometric alignment between 3D skeletons and visual data, this module pioneers robust joint learning from both modalities, creating a unified, cross-modal motion vocabulary. Second, grounded in this motion language, a single, unified MLLM architecture is trained to handle all tasks. This module flexibly processes diverse, temporal inputs, unifying 3D skeleton pose estimation from video (perception) with skeleton-based motion prediction and in-betweening (generation). Extensive experiments on standard benchmarks, including Human3.6M, demonstrate that our unified method achieves state-of-the-art or competitive performance across all motion tasks. This showcases a more efficient and scalable path for generative motion analysis using skeletons. Code is in the supplementary material.

Humanoid robotics has strong potential to transform daily service and caregiving applications. Although recent advances in general motion tracking within physics engines (GMT) have enabled virtual characters and humanoid robots to reproduce a broad range of human motions, these behaviors remain largely isolated and non-interactive. Assistive scenarios, by contrast, require continuous awareness of a human partner and rapid adaptation to their evolving posture and dynamics.In this paper, we formulate the imitation of closely interacting, force-exchanging human–human motion sequences as a multi-agent reinforcement learning problem. We jointly train partner-aware policies for both the supporter (assistant) and the recipient in a physics simulator to track assistive motion references. To make this problem tractable, we introduce a partner policies initialization scheme that transfers priors from single-human motion-tracking controllers, greatly improving exploration. We further propose dynamic reference retargeting and contact-promoting reward, which adapt the assistant’s reference motion to the recipient’s real-time pose and encourage physically meaningful support.We show that AssistMimic is the first method capable of successfully tracking assistive interaction motions on established benchmarks, demonstrating the benefits of a multi-agent RL formulation for physically grounded and socially aware humanoid control. We will make our code available upon acceptance.

Co-manipulation requires multiple humans to synchronize their motions with a shared object while ensuring reasonable interactions, maintaining natural poses, and preserving stable states. However, most existing motion generation approaches are designed for single-character scenarios or fail to account for payload-induced dynamics. In this work, we propose a flow-matching framework that ensures the generated co-manipulation motions align with the intended goals while maintaining naturalness and effectiveness. Specifically, we first introduce a generative model that derives explicit manipulation strategies from the object’s affordance and spatial configuration, which guide the motion flow toward successful manipulation. To improve motion quality, we then design an adversarial interaction prior that promotes natural individual poses and realistic inter-person interactions during co-manipulation. In addition, we also incorporate a stability-driven simulation into the flow matching process, which refines unstable interaction states through sampling-based optimization and directly adjusts the vector field regression to promote more effective manipulation. The experimental results demonstrate that our method achieves higher contact accuracy, lower penetration, and better distributional fidelity compared to state-of-the-art human-object interaction baselines. The code will be made publicly available.

Recent advances in motion diffusion models have substantially improved the realism of human motion synthesis. However, existing approaches either rely on full-sequence diffusion models with bidirectional generation, which limits temporal causality and real-time applicability, or autoregressive models that suffer from instability and cumulative errors. In this work, we present Causal Motion Diffusion Models (CMDM), a unified framework for autoregressive motion generation based on a causal diffusion transformer that operates in a semantically aligned latent space. CMDM builds upon a Motion-Language-Aligned Causal VAE (MAC-VAE), which encodes motion sequences into temporally causal latent representations. On top of this latent representation, an autoregressive diffusion transformer is trained using causal diffusion forcing to perform temporally ordered denoising across motion frames. To achieve fast inference, we introduce a frame-wise sampling schedule with causal uncertainty, where each subsequent frame is predicted from partially denoised previous frames. The resulting framework supports high-quality text-to-motion generation, streaming synthesis, and long-horizon motion generation at interactive rates. Experiments on HumanML3D and SnapMoGen demonstrate that CMDM outperforms existing diffusion and autoregressive models in both semantic fidelity and temporal smoothness, while substantially reducing inference latency.

Animation relies heavily on effective cinematography to enhance narrative clarity and emotional resonance, yet crafting optimal character interactions and camera positioning remains a resource-intensive challenge. Existing methods typically require extensive, predefined datasets, which restrict their effectiveness when encountering unfamiliar character interactions or novel animation contexts. We introduce an innovative approach to jointly generate character interactions and camera placements through unconditional diffusion-based generative models. Our method leverages a unified framework to simultaneously synthesize realistic two-person motions and corresponding cinematographic compositions without relying on predefined visual datasets. By integrating 3D motion representations and Toric features, our diffusion model effectively captures spatial orientation and relative positioning, enabling coherent and expressive scene generation. Experiments demonstrate that our approach can autonomously produce diverse and plausible dual-character interactions coupled with compelling camera movements, enhancing creative flexibility in animated storytelling.

We introduce MoLingo, a text-to-motion (T2M) model that generates realistic, lifelike human motion by denoising in a continuous latent space. Recent works perform latent space diffusion, either on the whole latent at once or auto-regressively over multiple latents. In this paper, we study how to make diffusion on continuous motion latents work best. We focus on two questions: (1) how to build a semantically aligned latent space so diffusion becomes more effective, and (2) how to best inject text conditioning so the motion follows the description closely. We propose a semantic-aligned motion encoder trained with frame-level text labels so that latents with similar text meaning stay close, which makes the latent space more diffusion-friendly. We also compare single-token conditioning with a multi-token cross-attention scheme and find that cross-attention gives better motion realism and text–motion alignment. With semantically aligned latents, auto-regressive generation, and cross-attention text conditioning, our model sets a new state of the art in human motion generation on standard metrics and in a user study. We will release our code and models for further research and downstream usage.

Existing humanoid control systems often rely on teleoperation or modular generation pipelines that separate language understanding from physical execution. However, the former is entirely human-driven, and the latter lacks tight alignment between language commands and physical behaviors. In this paper, we present SENTINEL, a fully end-to-end language–action model for humanoid whole-body control. We construct a large-scale dataset by tracking human motions in simulation using a pretrained whole body controller, combined with their text annotations. The model directly maps language commands and proprioceptive inputs to low-level actions without any intermediate representation. The model generates action chunks using flow matching, which can be subsequently refined by a residual action head for real-world deployment. Our method exhibits strong semantic understanding and stable execution on humanoid robots in both simulation and real-world deployment, and also supports multi-modal extensions by converting inputs into texts.

Recent text-to-image (T2I) diffusion models can produce highly realistic images from natural language prompts. In practice, users usually generate multiple candidates and select only a small subset for downstream use, guided by automatic metrics like CLIPScore and ImageReward. However, this post-hoc quality assessment is highly resource-intensive since quality is assessed after dozens to hundreds of denoising steps per image, leading to substantial waste on low-quality samples. To address this issue, we propose \textbf{Probe-Select}, a plug-in framework for early quality assessment in T2I generation. Our key observation is that certain intermediate features within the denoiser—often as early as 20\% of the reverse process—already encode stable structural cues (e.g., object layout, spatial composition, and color harmony) that strongly correlate with final image fidelity. Building upon this phenomenon, Probe-Select attaches lightweight probes to these stable activations at an early checkpoint and trains them to align with external evaluators. During inference, the probes forecast image quality on the fly, enabling early pruning of unpromising trajectories so that computation is concentrated on promising ones. Experiments on MS-COCO across multiple generative backbones show that this early assessment mechanism reduces sampling cost by over 60\% while improving the quality of the generated images, demonstrating that early structural signals can effectively guide efficient text-to-image generation.

Existing diffusion codecs typically build on text-to-image diffusion foundation models like Stable Diffusion.However, text conditioning is suboptimal from a compression perspective, hindering the potential of downstream diffusion codecs, particularly at ultra-low bitrates.To address it, we introduce CoD, the first Compression-oriented Diffusion foundation model, trained from scratch to enable end-to-end optimization of both compression and generation. CoD is not a fixed codec but a general foundation model designed for various diffusion-based codecs.It offers several advantages: High compression efficiency, replacing Stable Diffusion with CoD in downstream codecs like DiffC achieves SOTA results, especially at ultra-low bitrates (e.g., 0.0039 bpp); Low-cost and reproducible training, 300$\times$ faster training than Stable Diffusion ($\sim$ 20 vs. $\sim$ 6,250 A100 GPU days) on entirely open image-only datasets; Providing new insights, e.g., We find pixel-space diffusion can achieve VTM-level PSNR with high perceptual quality and can outperform GAN-based codecs using fewer parameters.We hope CoD lays the foundation for future diffusion codec research.Codes will be released.

Diffusion Magnetic Resonance Imaging (dMRI) plays a critical role in studying microstructural changes in the brain. It is, therefore, widely used in clinical practice; yet progress in learning general-purpose representations from dMRI has been limited. A key challenge is that existing deep learning approaches are not well-suited to capture the unique properties of diffusion signals. Brain dMRI is normally composed of several brain volumes, each with different attenuation characteristics dependent on the direction and strength of the diffusion-sensitized gradients. Thus, there is a need to jointly model spatial, diffusion-weighting, and directional dependencies in dMRI. Furthermore, varying acquisition protocols (e.g., differing numbers of directions) further limit traditional models. To address these gaps, we introduce a diffusion space rotatory positional embedding (D-RoPE) plugged into our dMRI transformer to capture both the spatial structure and directional characteristics of diffusion data, enabling robust and transferable representations across diverse acquisition settings and an arbitrary number of diffusion directions. After self-supervised masked autoencoding pretraining, tests on several downstream tasks show that the learned representations and the pretrained model can provide competitive or superior performance compared to several baselines in these downstream tasks (even compared to a fully trained baseline); the finetuned features from our pretrained encoder resulted in a 6\% higher accuracy in classifying mild cognitive impairment and a 0.05 increase in the correlation coefficient when predicting cognitive scores.

Nighttime photography is susceptible to flare caused by strong light sources, which degrades visual quality and disrupts structural information required by downstream vision tasks. Existing nighttime flare removal methods generally lack semantic priors for flare-occluded regions and thus tend to introduce artifacts and lose details under severe degradation. To address this problem, we propose a language-guided one-step diffusion framework that explicitly aligns flare-occluded regions with the underlying scene content at the semantic level. Specifically, we develop the first flare-specific vision–language model, Flare-VLM, which extracts fine-grained textual descriptions to guide one-step diffusion for high-quality restoration of severely damaged areas. Then, we propose semantics-aware distribution distillation to constrain the noise distribution with high-level semantics, suppressing redundant perturbations on clean backgrounds and improving the stability of distillation. In addition, we design an instruction-driven data synthesis pipeline to generate geometrically and semantically aligned nighttime flare samples, narrowing the gap to real degradations. Experimental results demonstrate that the proposed method achieves better restoration and enhances the performance of downstream vision tasks.

RAW images preserve superior fidelity and rich scene information compared to RGB, making them essential for tasks in challenging imaging conditions. To alleviate the high cost of data collection, recent RGB-to-RAW conversion methods aim to synthesize RAW images from RGB. However, they overlook two key challenges: (i) the reconstruction difficulty varies with pixel intensity, and (ii) multi-camera conversion requires camera-specific adaptation. To address these issues, we propose SpiralDiff, a diffusion-based framework tailored for RGB-to-RAW conversion with a signal-dependent noise weighting strategy that adapts reconstruction fidelity across intensity levels. In addition, we introduce CamLoRA, a camera-aware lightweight adaptation module that enables a unified model to handle diverse ISP characteristics. Extensive experiments on four benchmark datasets demonstrate the superiority of SpiralDiff in RGB-to-RAW conversion quality and its downstream benefits in RAW-based object detection.

Diffusion models have found extensive use in solving numerous inverse problems. Such diffusion inverse problem solvers aim to sample from the posterior distribution of data given the measurements, using a combination of the unconditional score function and an approximation of the posterior related to the forward process. Recently, consistency models (CMs) have been proposed to directly predict the final output from any point on the diffusion ODE trajectory, enabling high-quality sampling in just a few NFEs. CMs have also been utilized for inverse problems, but existing CM-based solvers either require additional task-specific training or utilize data fidelity operations with slow convergence, not amenable to large-scale problems. In this work, we reinterpret CMs as proximal operators of a prior, enabling their integration into plug-and-play (PnP) frameworks. We propose a solver based on PnP-ADMM, which enables us to leverage the fast convergence of conjugate gradient method. We further accelerate this with noise injection and momentum, dubbed PnP-CM, and show it maintains the convergence properties of the baseline PnP-ADMM. We evaluate our approach on a variety of inverse problems, including inpainting, super-resolution, Gaussian deblurring, and magnetic resonance imaging (MRI) reconstruction. To the best of our knowledge, this is the first CM trained for MRI datasets. Our results show that PnP-CM achieves high-quality reconstructions in as few as 4 NFEs, and can produce meaningful results in 2 steps, highlighting its effectiveness in real-world inverse problems while outperforming comparable CM-based approaches.

Accuracy camera viewpoint estimation under sparse-view conditions remains challenging, particularly in two-view scenarios. Recent approaches leverage diffusion models such as Zero123, which synthesize novel views conditioned on relative viewpoint, and have demonstrated promising performance when repurposed for viewpoint estimation via optimization with MSE loss. However, existing methods often suffer from non-convex loss landscape with numerous local minima, which makes them sensitive to initialization and reliant on na\"ive multi-start strategies to achieve reasonable results. We analyze these optimization challenges and visualize failure cases, showing that ambiguities in object geometry, such as symmetry and self-similarity, can mislead gradient-based updates toward incorrect viewpoints. To address these limitations, we propose a score-based method that reshapes the optimization landscape to guide updates toward the ground-truth viewpoint, followed by a refinement stage using a viewpoint-conditioned diffusion model. Experiments show that our method improves convergence, reduces reliance on brute-force sampling, and achieves competitive accuracy with higher sample-efficiency.

Diffusion models (DMs) have recently achieved remarkable success across diverse modalities, including high-fidelity image and video synthesis.However, their inherent step sequential denoising process introduces substantial cumulative latency, which significantly degrades user experience. While existing multi-GPU parallelization motheds can alleviate latency, they often incur prohibitive GPU-GPU communication overhead, offsetting much of the performance gain.We present Otil (Only Transmit Informative Latents), a communication-efficient parallel framework for accelerating diffusion inference.% Otil can minimizes redundant data exchange across GPUs while preserving generation quality.Our key insight is that latent activations change only marginally between consecutive denoising process. Leveraging this property, Otil identifies and synchronizes only the most informative latent sub-blocks and introduces a dynamic polling mechanism that periodically revisits all spatial regions, ensuring complete coverage without unnecessary communication. The framework is fully plug-and-play and remains compatible with fast sample and architectural acceleration algorithms, without requiring any retraining or architectural modification.Otil reduces GPU–GPU communication up to 87.5\% compared with SOTA parallelism methods, achieving $1.8\times$ speedup on two GPUs with Stable Diffusion v1.5 and $2.6\times$ on four GPUs with Stable Diffusion XL. When combined with few-step samplers ($30$ steps) and LoRA models, the acceleration further increases to $2.46\times$–$2.84\times$ on 2 GPUs. These demonstrate the strong potential of Otil for scalable and efficient multi-GPU diffusion inference while preserving generation fidelity.

Diffusion policies have emerged as a powerful paradigm for robot learning, but their inherent multi-modality can lead to a diverse set of plausible—though not always optimal—actions from a single observation. We posit that for a given task, an optimal action exists within this distribution. Inspired by negative prompting in generative models, we introduce a novel method that leverages an error detector to identify out-of-distribution (OOD) execution histories and uses them to construct negative action prompts. This allows our policy to steer away from suboptimal behaviors and converge towards higher-performance actions. We present a comprehensive ablation study demonstrating the effectiveness of positive, and negative prompts, and validate our approach on a suite of simulated benchmarks and real-world robotic tasks. Our results show that the proposed Negative-Prompt-guided Diffusion Policy achieves significant improvement in task performance by effectively filtering undesirable action modes.

Multimodal Large Language Models (MLLMs) have exhibited immense potential across numerous medical specialties, yet dentistry remains underexplored, in part due to limited domain-specific data, scarce dental expert annotations, insufficient modality-specific modeling, and challenges in reliability. In this paper, we present OralGPT-Omni, the first dental-specialized MLLM designed for comprehensive and trustworthy analysis across diverse dental imaging modalities and clinical tasks. To explicitly capture dentists’ diagnostic reasoning, we construct TRACE-CoT, a clinically grounded chain-of-thought dataset that mirrors dental radiologists’ decision-making processes. This reasoning supervision, combined with our proposed four-stage training paradigm, substantially strengthens the model’s capacity for dental image understanding and analysis. In parallel, we introduce MMOral-Uni, the first unified multimodal benchmark for dental image analysis. It comprises 2,809 open-ended question–answer pairs spanning five modalities and five tasks, offering a comprehensive evaluation suite to date for MLLMs in digital dentistry. OralGPT-Omni achieves an overall score of 51.84 on the MMOral-Uni benchmark and 45.31 on the MMOral-OPG benchmark, dramatically outperforming the scores of GPT-5. Our work promotes intelligent dentistry and paves the way for future advances in dental image analysis. All code, benchmark, and models will be made publicly available.

Human-object interaction (HOI) detection has traditionally been addressed using task-specific models, sometimes augmented by early vision-language models such as CLIP. With the emergence of large, generative VLMs, a natural question arises: can standalone VLMs perform HOI detection effectively, and how do they compare to specialized HOI methods? Existing benchmarks like HICO-DET rely on exact label matching under incomplete annotations, counting any unmatched prediction as wrong. This leads to incorrect penalization, especially for VLMs whose outputs are less constrained, making fair comparison between the two paradigms difficult. To address this limitation, we introduce a multi-choice HOI benchmark with explicitly defined positives and curated negatives, enabling unified and correct evaluation of both VLMs and HOI-specific models. We further focus on challenging scenarios, such as multi-person scenes and fine-grained interaction distinctions, which are crucial for revealing real differences between the two paradigms. Experiments show that large VLMs achieve competitive, sometimes superior, zero-shot performance, yet they struggle with multiple concurrent actions and with correctly assigning interactions to the target person. Conversely, HOI-specific methods remain weaker in general HOI reasoning but demonstrate stronger multi-action recognition and more reliable identification of which person performs which action. These findings expose complementary strengths and weaknesses of VLMs and HOI-specific methods, which existing benchmarks fail to reveal due to incorrect penalization.

This paper reveals that many open-source large language models (LLMs) lack hierarchical knowledge about our visual world, unaware of even well-established biology taxonomies. This shortcoming makes LLMs a bottleneck for vision LLMs' hierarchical visual recognition (e.g., recognizing Anemone Fish but not Vertebrate). We arrive at these findings using about one million four-choice visual question answering (VQA) tasks constructed from six taxonomies and four image datasets. Interestingly, finetuning a vision LLM using our VQA tasks reaffirms LLMs' bottleneck effect because the VQA tasks improve the LLMs' hierarchical consistency more than the vision LLMs'. We conjecture that one cannot make open-source vision LLMs understand visual concepts hierarchically until LLMs possess corresponding taxonomy knowledge.

Reading measurement instruments is effortless for humans and requires relatively little domain expertise, yet it remains surprisingly challenging for current vision-language models (VLMs) as we find in preliminary evaluation. In this work, we introduce MeasureBench, a benchmark on visual measurement reading covering both real-world and synthesized images of various types of measurements, along with an extensible pipeline for data synthesis. Our pipeline procedurally generates a specified type of gauge with controllable visual appearance, enabling scalable variation in key details such as pointers, scales, fonts, lighting, and clutter. Evaluation on popular proprietary and open-weight VLMs shows that even the strongest frontier VLMs struggle measurement reading in general. We have also conducted preliminary experiments with reinforcement finetuning (RFT) over synthetic data, and find a significant improvement on both in-domain synthetic subset and real-world images. Our analysis highlights a fundamental limitation of current VLMs in fine-grained spatial grounding. We hope this resource and our code releases can help future advances on visually grounded numeracy and precise spatial perception of VLMs, bridging the gap between recognizing numbers and measuring the world.

Progress in computer vision relies on the interplay of data, algorithms, and computation. For foundational tasks such as object detection, supervised learning with human-annotated data remains the state-of-the-art approach. However, this "gold-standard" data is notoriously error-prone, which is a fundamental bottleneck that hinders both model training and evaluation. As a result, benchmarking improvements have become negligible or non-existent in the last year. This issue does not stem from algorithms or computation, but from problem specifications and the dataset creation process. This ultimately leads to ill-defined tasks with noisy labels. Although statistical methods for Inter-Annotator Agreement (IAA) exist, they are often applied inconsistently and lack standardization, which makes dataset quality comparisons unreliable.We propose a unified meta-algorithm for dataset quality evaluation called K$\alpha$LOS (Krippendorff's $\alpha$ Localization Object Sensing) that serves as a tool for dataset creation and final assessment. Our framework conceptually incorporates existing methods and extends upon them. This provides a broader scope, as our method applies to any combined localization and classification task. It provides greater analytical depth than competing methods, enabling downstream tasks such as evaluating intra-annotator-consistency, rater vitality, and localization sensitivity. Crucially, it is modular, flexible, and extensible, allowing components to be interchanged for specific use-cases and enabling comparability across datasets and tasks.Validating such a metric is challenging, as no "real" ground truth exists. Typically, what we evaluate is considered the ground truth and starting point in the modeling process. Prior validation often relies on heuristics or machine-generated labels that fail to capture the complexity of real annotation noise. Therefore, we introduce an experimental validation approach using an empirical noise generator from real, multi-annotated datasets, which also scrutinizes heuristic assumptions about the noise distribution.

Automatic album organization has been studied extensively over the past decades due to significant progress in digital photography. Recent vision-language models (VLMs) have shown strong performance on multi-image understanding, making them natural candidates for automating album organization workflows. While VLMs' abilities in multi-image understanding have been widely studied, their performance on album organization remains underexplored. To bridge this gap, we introduce AlbumBench, the first comprehensive benchmark for automatic album organization. Specifically, we (1) define album organization tasks as photo selection for album-specific user objectives, photo rating according to how well user intents are fulfilled, and album-specific photo grouping given a user query which requires contextual understanding of the album; (2) establish AlbumBench, a benchmark dataset containing 27051 images across 641 albums with 5 annotations per image; and (3) evaluate mainstream open-source and proprietary VLMs on AlbumBench. We show that AlbumBench presents unique challenges compared to traditional multi-image understanding benchmarks due to its requirement for understanding album context and user intent. Our findings reveal a significant performance gap between open-source and proprietary VLMs on album organization tasks. Despite this gap, even the best-performing proprietary models sometimes struggle with tasks that humans find relatively easy. We hope that AlbumBench can serve as a foundation for unifying album organization research and motivate improvements in VLMs' performance on these tasks.

Visual–Language–Action (VLA) models report impressive success rates exceeding 95\% on robotic manipulation benchmarks, yet these results may mask fundamental weaknesses in robustness. Current simulation-based robustness evaluations suffer from narrow perturbation coverage, manual design constraints, and coarse-grained analysis that fails to reveal when and how models fail. To address this gap, we propose LIBERO-Plus, a comprehensive, automatic, and fine-grained evaluation framework with controlled perturbations across seven dimensions: object layouts, camera viewpoints, robot initial states, language instructions, lighting conditions, background textures, and sensor noise. Our systematic analysis of ten state-of-the-art models reveals consistent brittleness beneath apparent competence, with performance dropping from 95\% to below 30\% under modest perturbations. Our findings challenge the assumption that high benchmark scores equate to true competency and highlight the need for evaluation practices that assess reliability under realistic variation.

Recent advances in 3D vision–language models (VLMs) highlight a strong potential for 3D scene understanding and reasoning.However, effectively tokenizing 3D scenes into holistic scene tokens, and leveraging these tokens across diverse 3D understanding tasks, remain highly challenging. We present NDTokenizer3D, a generalist 3D VLM that performs a wide range of 3D scene understanding tasks while naturally supporting human interactions, thereby bridging language-level reasoning with 3D spatial understanding. The core of our approach is a novel three-stage scene tokenization pipeline built upon a Multi-Scale Normal Distributions Transform (NDT) representation, paired with a Multi-Scale NDT Decoder (MSDec).Specifically, NDTokenizer3D first constructs a multi-scale NDT representation from raw high-resolution point clouds, preserving both global context and fine-grained geometric details. Next, the MSDec progressively fuses cross-scale NDT features, producing holistic scene tokens consumable by LLM endpoints. Beyond tokenization, MSDec is repurposed as a general interface for human-interactive prompting (points, boxes, masks) and segmentation-mask decoding, unifying diverse 3D scene understanding tasks within a single architecture. With this compact and unified design, NDTokenizer3D offers a fine-grained, general-purpose 3D VLM, achieving remarkable improvements in 3D Referring Segmentation, 3D Visual Question Answering, and 3D Dense Captioning.

Language-Guided 3D segmentation is crucial for linking 3D perception with semantic understanding, yet it remains vulnerable to the incomplete and occluded views common in real-world RGB-D data. To overcome this, we present a real-time framework that leverages 3D Gaussian Splatting (3DGS) to build a semantically continuous and differentiable embedding field from partial observations. Our approach integrates two key components: a Dirichlet Process (DP) for the adaptive discovery of novel object categories, and a gradient low-rank mechanism that enhances class separability by reducing feature redundancy. This combination enables robust open-vocabulary segmentation guided directly by text prompts. Extensive experiments on challenging benchmarks demonstrate that our method achieves strong performance, exhibiting superior accuracy, robustness to incomplete inputs, and a powerful capacity for novel class discovery.

Many audio-visual learning methods have focused on aligning audio and visual information, either through semantic or temporal correspondence. However, most of these works have utilized monaural audio, which does not contain information about the spatial location of the sound source. In contrast, humans and other animals utilize binaural hearing to perceive this spatial information. Combining spatial sound and visual perception enables powerful high-level reasoning: for example, a person looking for their phone may hear the ringing sound coming from a backpack sitting on a table, and quickly infer that the missing phone is inside the backpack. In this paper, we investigate the problem of Audio-Visual Spatial Reasoning. We design a spatial audio-visual question answering dataset to cover scenarios where semantic correspondence between audio and visual signals is absent but spatial alignment exists, as well as cases with multiple audio-visual semantic correspondences that require spatial reasoning to disambiguate. We propose a model that learns spatial comprehension across the audio and vision modalities by connecting them with a large language model and experimentally demonstrate that spatial sound perception is an essential part of our task.

Multimodal large language models (MLLMs) are increasingly applied to spatial cognition tasks, where they are expected to understand and interact with complex environments. Most concurrent works enhance spatial reasoning by introducing 3D priors or geometric supervision, which improves performance but incurs substantial data preparation and alignment costs. Purely 2D approaches, however, struggle with multi-frame spatial reasoning due to missing viewpoint transitions and overlooked implicit objects that act as spatial bridges. To address these limitations, we propose EgoMind, a Chain-of-Thought framework that enables geometry-free spatial reasoning through Role-Play Captioning and Progressive Spatial Analysis, jointly constructing a coherent linguistic scene graph across frames. With only 5K auto-generated SFT samples and 20K RL samples, EgoMind achieves competitive results on VSI-Bench, SPAR-Bench, STI-Bench and SPBench, demonstrating its effectiveness in reinforcing the spatial reasoning capabilities of MLLMs and highlighting the potential of linguistic reasoning for spatial cognition.

3D Referring Expression Segmentation (3D-RES) aims to segment objects in point clouds according to language descriptions. Unlike common practices in 2D that utilize learnable query embeddings, recent 3D-RES methods typically generate queries directly from 3D points. However, this direct coupling of queries to raw point clouds introduces new challenges: an impractically large number of queries derived from massive point cloud data and a reliance on non-deterministic sampling algorithms. In this paper, we propose a Semantic-based Adaptive Query Network (SAQN), which introduces a novel query strategy for 3D-RES. Instead of generating queries from points, SAQN employs a learnable query vector for each semantic class. This approach drastically reduces the number of queries while maintaining the advantage of avoiding Hungarian matching through implicit class alignment. Additionally, to address potential cross-object ambiguity within semantic classes, we introduce supplementary queries that are adaptively fused with each class query to disambiguate and enrich representations. Comprehensive experiments show that SAQN achieves state-of-the-art performance while reducing the number of queries.

Recent spatial intelligence approaches typically attach 3D cues to 2D reasoning pipelines or couple MLLMs with black-box reconstruction modules, leading to weak spatial consistency, limited viewpoint diversity, and evidence chains that cannot be traced back to supporting views. Frameworks for “thinking with images’’ (e.g., ChatGPT–o3 and DeepEyes) show that stepwise multimodal reasoning can emerge by interleaving hypothesis formation with active acquisition of visual evidence, but they do not address three key challenges in spatial Chain-of-Thought (CoT): building global space perception under strict token budgets, explicitly associating 3D hypotheses with video frames for verification, and designing spatially grounded rewards for reinforcement learning. To address these issues, we present \textbf{EagleVision}, a dual-stage framework for progressive spatial cognition through macro perception and micro verification. In the macro perception stage, EagleVision employs a semantics–perspective-fusion determinantal point process (SPF-DPP) to select a compact set of geometry- and semantics-aware keyframes from long videos under a fixed token budget. In the micro verification stage, we formalize spatial CoT as BEV-grounded pose querying: the agent iteratively predicts poses on a BEV plane, retrieves the nearest real frames, and is trained purely by reinforcement learning with a spatial grounding reward that scores the consistency between predicted poses and observed views. On VSI-Bench, EagleVision achieves state-of-the-art performance among open-source vision–language models, demonstrating strong and generalizable spatial understanding.

Vision-language models (VLMs) struggle with 3D-related tasks such as spatial cognition and physical understanding, which are crucial for real-world applications like robotics and embodied agents.We attribute this to a modality gap between the 3D tasks and the 2D training of VLM, which led to inefficient retrieval of 3D information from 2D input.To bridge this gap, we introduce SandboxVLM, a simple yet effective framework that leverages abstract bounding boxes to encode geometric structure and physical kinematics for VLM. Specifically, we design a 3D Sandbox reconstruction and perception pipeline comprising four stages: generating multi-view priors with abstract control, proxy elevation, multi-view voting and clustering, and 3D-aware reasoning.Evaluated in zero-shot settings across multiple benchmarks and VLM backbones, our approach consistently improves spatial intelligence, achieving an 8.3\% gain on SAT Real compared with baseline methods for instance.These results demonstrate that equipping VLMs with a 3D abstraction substantially enhances their 3D reasoning ability without additional training, suggesting new possibilities for general-purpose embodied intelligence.

3D visual grounding (VG) aims to localize target objects in 3D scenes based on free-form textual descriptions. Existing 3D VG models predominantly employ point-based backbones for point cloud feature extraction. Such methods require aggressive downsampling of the input point cloud, which sacrifices the fine-grained spatial details crucial for precise localization. This paper proposes PV-Ground, a novel 3D VG architecture based on effective text-guided point-voxel feature interaction. Our method leverages the complementary strengths of both voxels and keypoints: it employs a voxel-based feature extraction backbone to preserve high-resolution spatial details, while utilizing compact keypoints to aggregate these features for efficient, deep interaction with the textual query. Furthermore, we propose a text-guided keypoint sampling module to adaptively concentrate the keypoint distribution around the text-described object, enabling task-specific feature aggregation and significantly boosts model performance. Extensive qualitative and quantitative experiments demonstrate the superiority of our proposed method. Our method achieves a performance improvement of 5.1\% on the ScanRefer dataset and 5.6\% on the ReferIt3D dataset, while also achieves over 4\% improvement in the segmentation task. The code will be made publicly available.

Recent advances in 3D scene-language understanding have leveraged Large Language Models (LLMs) for 3D reasoning by transferring their general reasoning ability to 3D multi-modal contexts. However, existing methods typically adopt standard decoders from language modeling, which rely on a causal attention mask. This design introduces two fundamental conflicts in 3D scene understanding: sequential bias among order-agnostic 3D objects and restricted object-instruction attention, hindering task-specific reasoning. To overcome these limitations, we propose 3D Spatial Language Instruction Mask (3D-SLIM), an effective masking strategy that replaces the causal mask with an adaptive attention mask tailored to the spatial structure of 3D scenes. Our 3D-SLIM introduces two key components: a Geometry-adaptive Mask that constrains attention based on spatial density rather than token order, and an Instruction-aware Mask that enables object tokens to directly access instruction context. This design allows the model to process objects based on their spatial relationships while being guided by the user's task. 3D-SLIM is simple, requires no architectural modifications, and adds no extra parameters, yet it yields substantial performance improvements across diverse 3D scene-language tasks. Extensive experiments across multiple benchmarks and LLM baselines validate its effectiveness and underscore the critical role of decoder design in 3D multi-modal reasoning.

Large vision-language models (VLMs) still struggle with reliable 3D spatial reasoning, a core capability for embodied and physical AI systems. This limitation arises from their inability to capture fine-grained 3D geometry and spatial relationships. While recent efforts have introduced multi-view geometry transformers into VLMs, they typically fuse only the deep-layer features from vision and geometry encoders, discarding rich hierarchical signals and creating a fundamental bottleneck for spatial understanding. To overcome this, we propose SpatialStack, a general hierarchical fusion framework that progressively aligns vision, geometry, and language representations across the model hierarchy. Moving beyond conventional late-stage vision-geometry fusion, SpatialStack stacks and synchronizes multi-level geometric features with the language backbone, enabling the model to capture both local geometric precision and global contextual semantics. Building upon this framework, we develop VLM-SpatialStack, a model that achieves state-of-the-art performance on multiple 3D spatial reasoning benchmarks. Extensive experiments and ablations demonstrate that our multi-level fusion strategy consistently enhances 3D understanding and generalizes robustly across diverse spatial reasoning tasks, establishing SpatialStack as an effective and extensible design paradigm for vision-language-geometry integration in next-generation multimodal physcial AI systems.

Vision Language Models (VLMs) exhibit a fundamental semantic-to-geometric gap in spatial reasoning: they excel at qualitative semantic inference but their reasoning operates within a lossy semantic space, misaligned with high-fidelity geometry. Current paradigms fail to bridge this gap. Training-based methods suffer from an ``oracle paradox,'' learning flawed spatial logic from imperfect oracles. Tool-integrated methods constrain the final computation but critically leave the VLM's planning process unconstrained, resulting in geometrically flawed plans. In this work, we propose Geometrically-Constrained Agent (GCA), a training-free agentic paradigm that resolves this gap by introducing a formal task constraint. Specifically, we strategically decouples the VLM's role into two stages. First, acting as a semantic analyst, the VLM translates the user's ambiguous query into the formal, verifiable task constraint, which defines the reference frame and objective. Second, acting as a task solver, the VLM generates and executes tool calls strictly within the deterministic bounds defined by the constraint. This geometrically-constrained strategy successfully resolve the semantic-to-geometric gap, yielding a robust and verifiable reasoning pathway for complex spatial tasks. Comprehensive experiments demonstrate that GCA achieves SOTA performance on multiple spatial reasoning benchmarks, surpassing existing training-based and tool-integrated methods by ~27%.

Inter-object relations underpin spatial intelligence, yet existing representations—linguistic prepositions or object-level scene graphs—are too coarse to specify which regions actually support, contain, or contact one another, leading to ambiguous and physically inconsistent layouts. To address these ambiguities, a part-level formulation is needed; therefore, we introduce PARSE, a framework that explicitly models how object parts interact to determine feasible and spatial grounded scene configurations. PARSE centers on the Part-centric Assembly Graph (PAG), which encodes geometric relations between specific object parts, and a Part-Aware Spatial Configuration Solver that converts these relations into geometric constraints to assemble collision-free, physically valid scenes. Using PARSE, we build PARSE-10K, a dataset of 10,000 3D indoor scenes constructed from real-image layout priors and a curated part-annotated shape database, each with dense contact structures and a part-level contact graph. With this structured, spatially grounded supervision, fine-tuning Qwen3-VL on PARSE-10K yields stronger object-level layout reasoning and more accurate part-level relation understanding; furthermore, leveraging PAGs as structural priors in 3D generation models leads to scenes with substantially improved physical realism and structural complexity. Together, these results show that PARSE significantly advances geometry-grounded spatial reasoning and supports the generation of physically consistent 3D scenes.

Humans perceive and reason about their surroundings in four dimensions-three spatial and one temporal axis-by building persistent, structured internal representations that encode semantic meaning, spatial layout, and temporal dynamics. These multimodal memories enable them to recall past events, infer unobserved states, and integrate new information into context-dependent reasoning. Inspired by this capability, we introduce R4, a training-free framework for retrieval-augmented reasoning in 4D spatio-temporal space that equips vision-language models (VLMs) with structured, lifelong memory. R4 continuously constructs a 4D knowledge database by anchoring object-level semantic descriptions in metric space and time, yielding a persistent world model that can be shared across agents. At inference, natural language queries are decomposed into semantic, spatial, and temporal keys to retrieve relevant observations, which are integrated into the VLM's reasoning through an iterative retrieval-reasoning loop. Unlike classical retrieval-augmented generation methods, retrieval in R4 operates directly in structured 4D space, enabling episodic and collaborative reasoning without training. Experiments on embodied question answering and navigation benchmarks demonstrate that R4 substantially improves retrieval and reasoning over spatio-temporal information compared to baselines, advancing a new paradigm for embodied 4D reasoning in dynamic environments.

Chinese historical documents are essential carriers for the inheritance and dissemination of traditional Chinese culture.However, traditional manual digitization of different types of historical carriers is not only time-consuming and labor-intensive but also heavily reliant on experts with specialized knowledge of the specific carrier domains.In the past,experts read the Chinese historical documents relying on the recognition of the documents and consulted a large number of professional books for citation and correction .With the emergence of Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs), we see new opportunities for uniformly reading different types of carriers. Nevertheless, existing studies mainly focus on evaluating the OCR capabilities of MLLMs, without incorporating citation or retrieval functionalities, and are restricted to a single type of carrier.To address this,we introduce MCHDoc,a comprehensive benchmark for reading multi-carrier Chinese historical documents.This benchmark consists of 15,723 documents and covers six types of carriers, including Inscription, AncientBook, Calligraphy, Oracle Bone, Silk, and JianDu(bamboo slip).Based on this benchmark,we evaluate various MLLMs and LLMs to test their capacities of reading multi-carrier Chinese historical documents. The results reveal that the top MLLMs and LLMs achieve excellent performance on some type of carriers. but there is still some place for them to read the multi-type carriers perfectly.Overall,MCHDoc is a standardized and comprehensive benchmark for reading Chinese historical document, providing valuable insights for Chinese cultural study.

Text-aerial person retrieval aims to identify targets in UAV-captured images from eyewitness descriptions, supporting intelligent transportation and public security applications. Compared to ground-view text–image person retrieval, UAV-captured images often suffer from degraded visual information due to drastic variations in viewing angles and flight altitudes, making semantic alignment with textual descriptions very challenging. To address this issue, we propose a novel Cross-modal Fuzzy Alignment Network (CFANet), which quantifies the token-level reliability by fuzzy logic to achieve accurate fine-grained alignment and incorporates ground-view images as a bridge agent to further mitigate the gap between aerial images and text descriptions, for text–aerial person retrieval. In particular, we design the Fuzzy Token Alignment module that employs the fuzzy membership function to dynamically model token-level association strength and suppress the influence of unobservable or noisy tokens. It can alleviate the semantic inconsistencies caused by missing visual cues and significantly enhance the robustness of token-level semantic alignment. Moreover, to further mitigate the gap between aerial images and text descriptions, we design a Context-Aware Dynamic Alignment module to incorporates the ground-view agent as a bridge in text–aerial alignment and adaptively combine direct alignment and agent-assisted alignment to improve the robustness. In addition, we construct a large-scale benchmark dataset called AERI-PEDES by using a chain-of-thought to decompose text generation into attribute parsing, initial captioning, and refinement, thus boosting textual accuracy and semantic consistency. Experiments on AERI-PEDES and TBAPR demonstrate the superiority of our method. The code and dataset will be publicly released.

Code search, framed as information retrieval (IR), underpins modern software engineering and increasingly powers retrieval-augmented generation (RAG), improving code discovery, reuse, and the reliability of LLM-based coding.Yet existing code IR models remain largely text-centric and often overlook the visual and structural aspects inherent in programming artifacts such as web interfaces, data visualizations, SVGs, schematic diagrams, and UML.To bridge this gap, we introduce MMCoIR, the first comprehensive benchmark for evaluating multimodal code IR across five visual domains, and show through extensive evaluation the task is challenging.Therefore, we then propose CodeMMR, a unified retrieval model that jointly embeds natural language, code, and images into a shared semantic space through instruction-based multimodal alignment.CodeMMR achieves strong generalization across modalities and languages, outperforming competitive baselines (e.g., UniIR, GME, VLM2Vec) by an average of 10 points on nDCG@10.Moreover, integrating CodeMMR into RAG enhances code generation fidelity and visual grounding on unseen code generation tasks, underscoring the potential of multimodal retrieval as a core enabler for next-generation intelligent programming systems.

Open-set fine-grained retrieval~(OSFR) is a challenging task where models must generalize to unseen subcategories. Existing methods often fail this, as they embed category-specific semantics from closed-set training labels. Recently, diffusion transformers (DiT) have shown promise by encoding \textit{attribute-centric, generative curriculum knowledge} that is agnostic to these labels. However, the vanilla DiT is not optimized for fine-grained \textit{visual discrepancies} and its massive size makes \textit{deployment infeasible}. To solve this, we propose \textbf{DiT-Distill}, a framework to first refine and then distill this knowledge. We introduce a \textit{conditional discrepancy refinement} strategy to fine-tune the DiT, forcing it to focus on discrepancy-aware, attribute-centric details rather than holistic context. Subsequently, a \textit{generative curriculum distillation} mechanism transfers this refined, hierarchical knowledge from multiple diffusion timesteps of the DiT into a lightweight backbone using a generative infusion module and a curriculum alignment loss. This process results in an efficient retrieval model that enables \textit{DiT-free inference}. Extensive experiments show DiT-Distill achieves state-of-the-art performance on open-set fine-grained datasets.

Composed Image Retrieval (CIR) aims to retrieve target images based on a hybrid query comprising a reference image and a modification text. Early dual-tower Vision–Language Models (VLMs) struggle with cross-modality compositional reasoning required for this task. Recently, adapting generative Multimodal Large Language Models (MLLMs) for retrieval offers a promising direction. However, we identify that this adaptation strategy overlooks a fundamental issue: adapting a generative MLLM into a single-embedding discriminative retriever triggers a paradigm conflict, which leads to Capability Degradation—the deterioration of native fine-grained reasoning after retrieval adaptation. To address this challenge, we propose ReCALL (Recalibrating Capability Degradation), a model-agnostic framework that follows a diagnose–generate–refine pipeline: Firstly, we diagnose cognitive blind spots of the retriever via self-guided informative instance mining. Next, we generate corrective instructions and triplets by CoT prompting the foundation MLLM and conduct quality control with VQA-based consistency filtering. Finally, we refine the retriever through continual training on these triplets with a grouped contrastive scheme, thereby internalizing fine-grained visual–semantic distinctions and realigning the discriminative embedding space of retriever with intrinsic compositional reasoning within the MLLM. Extensive experiments on CIRR and FashionIQ show that ReCALL consistently recalibrates degraded capabilities and achieves state-of-the-art performance. Code will be released soon.

Visual in-context learning (VICL) enables visual foundation models to handle multiple tasks by steering them with demonstrative prompts. The choice of such prompts largely influences VICL performance, standing out as a key challenge.Prior work has made substantial progress on prompt retrieval and reranking strategies; however, they focused primarily on prompt images while often overlooking the labels. We reveal that these approaches sometimes get visually similar but label-inconsistent prompts, which potentially degrade VICL performance. On the other hand, higher label consistency between query and prompts preferably indicates stronger VICL results. Motivated by these findings, we develop a framework named LaPR (Label-aware Prompt Retrieval), which highlights the role of labels in prompt selection. Our framework first designs an image–label joint representation for prompts to incorporate label cues explicitly. Besides, to handle unavailable query labels at test time, we introduce a mixture-of-expert mechanism to the dual encoders with query-adaptive routing. Each expert is expected to capture a specific label mode, while the router infers query-adaptive mixture weights and helps to learn label-aware representation.We carefully design alternative optimization for experts and the router, with a VICL performance-guided contrastive loss and a label-guided contrastive loss, respectively. Extensive experiments show promising and consistent improvement of LaPR on in-context segmentation, detection, and colorization tasks. Moreover, LaPR generalizes well across feature extractors and cross-fold scenarios, suggesting the importance of label utilization in prompt retrieval for VICL. Faithful code will be released publicly.

Fine-tuning vision–language models such as CLIP has become the mainstream paradigm for multi-label image recognition, and prompt tuning is widely adopted due to its lightweight parameter cost and strong transferability. However, we find that when these methods use Binary Cross-entropy as the supervision loss, the model’s confidence structure becomes systematically distorted, leading to pronounced miscalibration. Existing calibration techniques, such as temperature scaling or regularization-based methods, largely fail in multi-label settings because they cannot capture inherent semantic dependencies between classes, nor can they correct the global structural shifts introduced during fine-tuning. To address this issue, we propose Class-wise Covariance Regularization, which aligns the predicted covariance structure of class confidences with the semantic correlations encoded in pretrained text embeddings. This alignment preserves the geometric consistency of the class space throughout fine-tuning, resulting in more stable and interpretable confidence distributions across categories. Experiments on multi-label benchmarks show that CCR significantly reduces calibration errors while maintaining or even improving classification accuracy.

Interactive Text-to-Image Retrieval (I-TIR) aims to refine image retrieval results through natural language dialogues, which allows users to progressively supplement or correct their search intention across multiple rounds, enabling a more precise and user-aligned visual search experience.However, existing methods perform cross-modal retrieval within a fixed multimodal feature space, mapping all dialogue text and images onto the same static embedding manifold.Such static formulation easily causes semantic vagueness, making it difficult to capture subtle embedding shifts in the user's updated intention for fine-grained retrieval.To address this limitation, we propose Context-Aware Latent Space Transformation (CAST), a lightweight framework that dynamically transforms the common latent space of both textual and visual representations according to the specific evolving user's search intention, enabling fine-grained and adaptive semantic alignment.The core of CAST is the Context-Aware Space Regulator (CASR), a crucial space transformation module composed of two key components: (1) the Context-Aware Low-Rank Projector (CLP), which learns to predict the projection direction of embedding space based on the intent's context;and (2) the Context-Guided Modulator (CGM), which adaptively determines appropriate projection strength. CASR is highly lightweight, adding negligible parameters and computational overhead, and can be seamlessly integrated into diverse I-TIR frameworks.Extensive experiments demonstrate the effectiveness of our proposed framework, indicating that it can serve as a general, plug-and-play solution for efficient and scalable interactive text-to-image retrieval. Our source code is provided in the supplementary material.

Non-human primates are our closest living relatives, and analyzing their behavior is central to research in cognition, evolution, and conservation. Computer vision could greatly aid this research, but existing methods often rely on human-centric pretrained models and focus on single datasets, which limits generalization. We address this limitation by shifting from a model-centric to a data-centric approach and introduce PriVi, a large-scale primate-centric video pretraining dataset. PriVi contains 424 hours of curated video, combining 174 hours from behavioral research across 11 settings with 250 hours of diverse web-sourced footage, assembled through a scalable data curation pipeline. We continue pretraining V-JEPA on PriVi to learn primate-specific representations and evaluate it using a lightweight frozen classifier. Across four benchmark datasets — ChimpACT, PanAf500, BaboonLand, and ChimpBehave — our approach consistently outperforms prior work, including fully finetuned baselines, and scales favorably with fewer labels. These results demonstrate that primate-centric pretraining substantially improves data efficiency and generalization, making it a promising approach for low-label applications. Code, models, and the majority of the dataset will be made available.

In everyday conversations, humans effortlessly recognize communication partners using visual cues such as gaze or head orientation. Replicating this social reasoning in computer vision is challenging, especially in dynamic, multi-person settings. We introduce Communication Context Identification (CCI) in egocentric vision: Given a first-person video sequence, determine which individuals are engaged in communication with the camera wearer. To support CCI, we collected a challenging large-scale dataset comprising 68.9 hours of egocentric video captured across diverse multi-person, multi-conversation scenarios.We propose CoCoNet, a temporal interaction model for CCI that tracks social dynamics via attention across individuals over long time scales. CoCoNet flexibly handles varying group sizes, maintains predictions through occlusions, and performs robustly even with limited temporal input. Leveraging long temporal contexts, it achieves 96\% balanced accuracy on CCI. Performance varies with group size and spatial scene layout, highlighting the importance of dataset diversity. Our work advances vision-based conversational awareness, enabling applications in assistive hearing that use egocentric video to enhance individuals in the user’s conversation group.

Human memory is often unreliable. We forget where we placed objects, overlook small details, and struggle to recall past events accurately. Episodic Memory with Natural Language Query (EM-NLQ) seeks to overcome these limitations by allowing users to search their past visual experiences, captured through egocentric videos, using natural language questions. While recent models focus on addressing challenges in EM-NLQ like noisy input videos and efficiency, they overlook a key aspect of this task: interactivity. In real scenarios, users have the ability to refine their queries and provide feedback when a model's response is off-target, yet current EM-NLQ methods cannot incorporate or benefit from such feedback. To address this gap, we introduce the first \textit{interactive} EM-NLQ framework, featuring a plug-and-play Feedback ALignment Module (FALM) that empowers existing models to efficiently incorporate user feedback and refine their predictions. Additionally, we introduce the Episodic Memory with Questions and Feedback task (EM-QnF) and new datasets tailored for feedback-based interaction and a lightweight training scheme that eliminates the need for expensive sequential optimization. Our approach, dubbed ReFocus, combines FALM with state-of-the-art EM-NLQ methods to achieve state-of-the-art results on three challenging benchmarks and demonstrates significant improvements in human-based feedback evaluation, bringing EM-NLQ closer to truly interactive and adaptive visual memory systems.

Can one perceive a video's content without seeing its pixels, just from the camera trajectory—the path it carves through space? This paper is the first to systematically investigate this seemingly implausible question. Towards this end, we propose a contrastive learning framework to train CamFormer, a dedicated encoder that projects camera pose trajectories into a joint embedding space, aligning them with natural language. We find that, contrary to its apparent simplicity, the camera trajectory is a remarkably informative signal to uncover video content. In other words, "how you move" can indeed reveal "what you are doing" (egocentric) or "observing" (exocentric). We demonstrate the versatility of our learned CamFormer embeddings on a diverse suite of downstream tasks, ranging from cross-modal alignment to classification and temporal analysis. Importantly, our representations are robust across diverse camera pose estimation methods, including both high-fidelity multi-sensored and standard RGB-only estimators. Our findings establish camera trajectory as a lightweight, robust, and versatile modality for perceiving video content.

Spatiotemporal predictive learning (STPL) aims to forecast future frames from past observations and is essential across a wide range of applications. Compared with recurrent or hybrid architectures, pure convolutional models offer superior efficiency and full parallelism, yet their fixed receptive fields limit their ability to adaptively capture spatially varying motion patterns. Inspired by biological center–surround organization and frequency-selective signal processing, we propose PFGNet, a fully convolutional framework that dynamically modulates receptive fields through pixel-wise frequency-guided gating. The core Peripheral Frequency Gating (PFG) block extracts localized spectral cues and adaptively fuses multi-scale large-kernel peripheral responses with learnable center suppression, effectively forming spatially adaptive band-pass filters. To maintain efficiency, all large kernels are decomposed into separable 1D convolutions ($1\times k$ followed by $k\times1$), reducing per-channel computational cost from $\mathcal{O}(k^2)$ to $\mathcal{O}(2k)$. PFGNet enables structure-aware spatiotemporal modeling without recurrence or attention. Experiments on Moving MNIST, TaxiBJ, Human3.6M, and KTH show that PFGNet delivers SOTA or near-SOTA forecasting performance with substantially fewer parameters and FLOPs.

Video reasoning models are a core component of egocentric and embodied agents. However, standard benchmarks for assessing models provide only evaluation of the output (e.g., the answer to a question), without evaluation of intermediate reasoning steps, and most provide answers only in the text domain. We introduce EgoSTAR, a benchmark for evaluating complex egocentric visual reasoning. We extend recent high-quality video data sources recorded from egocentric / embodied settings with a set of challenging, multi-step multimodal questions and spatiotemporally-dense human-annotated reasoning traces. Benchmarking experiments show that state-of-the-art models still have a large gap to human performance. To investigate this gap in detail, we annotate each reasoning trace in the dataset with the objects of interest required to solve the question, for which we also have spatio-temporal mask annotations. Through extensive evaluations, we identify that if frontier models are prompted with hints of where' andwhen' to look, we can get substantial improvements in performance. EgoSTAR will be released publicly to foster progress in egocentric reasoning.

The transition of Retrieval-Augmented Generation (RAG) from offline video analysis to online, streaming scenarios presents a set of critical, unexplored challenges. These include the need for on-the-fly semantic segmentation of continuous video, the inherent tension between low-latency processing and high-quality knowledge extraction, and the demand for query-specific temporal reasoning. We propose StreamRAG, a novel framework designed to overcome these hurdles. StreamRAG is built upon three core technical pillars: (1) a Stream Event Segmentation (SES) module that performs real-time boundary detection to chunk the stream into meaningful units; (2) a Token-Reusing Accelerator that drastically cuts down captioning latency by leveraging computational overlap between consecutive frames; and (3) a Dynamic Retrieval Gate that modulates the retrieval scope and strategy based on the query's temporal sensitivity and contextual similarity. Empirical evaluation confirms that StreamRAG establishes a new state-of-the-art, delivering superior accuracy with minimal latency in streaming video comprehension.

Recent advancements in Video Large Language Models (VideoLLMs) have enabled strong performance across diverse multimodal video tasks. To reduce the high computational cost of processing dense video frames, efficiency-oriented methods such as frame selection have been widely adopted. While effective at minimizing redundancy, these methods often cause notable performance drops on tasks requiring temporal reasoning. Unlike humans, who can infer event progression from sparse visual cues, VideoLLMs frequently misinterpret temporal relations when intermediate frames are omitted. To address this limitation, we explore visual prompting (VP) as a lightweight yet effective way to enhance temporal understanding in VideoLLMs. Our analysis reveals that simply annotating each frame with explicit ordinal information helps the model perceive temporal continuity. This visual cue also supports frame-level referencing and mitigates positional ambiguity within a sparsely sampled sequence. Building on these insights, we introduce ViKey, a training-free framework that combines VP with a lightweight Keyword–Frame Mapping (KFM) module. KFM leverages frame indices as dictionary-like keys to link textual cues to the most relevant frames, providing explicit temporal anchors during inference. Despite its simplicity, our approach substantially improves temporal reasoning and preserves dense-frame baseline performance with as few as 20% of frames.

Egocentric perception on smart glasses could transform how we learn new skills in the physical world, but automatic skill assessment remains a fundamental technical challenge. We introduce SkillSight for power-efficient skill assessment from first-person data. Central to our approach is the hypothesis that skill level is evident not only in how a person performs an activity (video), but also in how they directtheir attention when doing so (gaze). Our two-stage framework first learns to jointly model gaze and egocentric video when predicting skill level, then distills a gaze-only student model. At inference, the student model requires only gaze input, drastically reducing power consumption by eliminating continuous video processing. Experiments on three datasets spanning cooking, music, and sports establish, for the first time, the valuable role of gaze in skill understanding across diverse real-world settings. Our SkillSight teacher model achieves state-of-the-art performance, while our gaze-only student variant maintains high accuracy using 73× less power than competing methods. These results pave the way for in-the-wild AI-supported skill learning.

Action Quality Assessment (AQA) aims to score how well an action is performed and is widely used in sports analysis, rehabilitation assessment, and human skill evaluation. Multi-modal AQA has recently achieved strong progress by leveraging complementary visual and kinematic cues, yet real-world deployments often suffer from non-stationary modality imbalance, where certain modalities become missing or intermittently available due to sensor failures or annotation gaps. Existing continual AQA methods overlook this issue and assume that all modalities remain complete and stable throughout training, which restricts their practicality. To address this challenge, we introduce Bridged Modality Adaptation (BriMA), an innovative approach to multi-modal continual AQA under modality-missing conditions. BriMA consists of a memory-guided bridging imputation module that reconstructs missing modalities using both task-agnostic and task-specific representations, and a modality-aware replay mechanism that prioritizes informative samples based on modality distortion and distribution drift. Experiments on three representative multi-modal AQA datasets (RG, Fis-V, and FS1000) show that BriMA consistently improves performance under different modality-missing conditions, achieving 6--8\% higher correlation and 12--15\% lower error on average. These results demonstrate a step toward robust multi-modal AQA systems under real-world deployment constraints.

While language models have become impactful in many real-world applications, video generation remains largely confined to entertainment. Motivated by video's inherent capacity to demonstrate physical-world information that is difficult to convey through language alone (e.g., imagine teaching someone to tie a tie using only text), we identify an underutilized opportunity to extend video as a new answer modality for Next-Event Prediction (NEP), formalized as Video-Next-Event Prediction (VNEP). While the established NEP task takes a video with a procedural or predictive question as input to predict the next event in text, VNEP requires dynamic video responses. This shift from telling to showing unlocks more intuitive and customized answers for procedural learning and creative exploration. However, this task remains challenging for existing models, as it demands an understanding of multimodal input, instruction-conditioned reasoning, and the generation of video with visual and semantic consistency. To address this, we introduce VANS, a model that leverages reinforcement learning to align a Vision-Language Model (VLM) with a Video Diffusion Model (VDM) for VNEP. The core of VANS is our proposed Joint-GRPO that orchestrates the VLM and VDM to function as a unit. Driven by a shared reward on their respective output, it optimizes the VLM to produce captions that are both accurate and friendly to visualize, while guiding the VDM to generate videos that are faithful to these captions and the input visual context. To enable this learning, we craft VANS-Data-100K, a dedicated dataset for the VNEP task. Experiments on procedural and predictive benchmarks demonstrate that VANS achieves state-of-the-art performance in both video event prediction and visualization.

Saliency-based explainability methods are widely used to interpret deep learning models in medical imaging, yet many existing approaches rely on white box access of models, which is not always possible due to privacy concerns. In this work, we introduce MedLIME, a novel, model-agnostic explanation framework designed to enhance the robustness and fidelity of saliency maps for medical imaging abnormality localization. Building upon the Local Interpretable Model-agnostic Explanations (LIME) paradigm, MedLIME integrates three key components: (1) Generative Masking (GM), (2) Supervised Test-Time Adaptation (STT) and (3) a Evidence-based Regularization (EBR) to improve the saliency map generation accuracy of LIME. Extensive experiments on multiple medical datasets, across three model architectures demonstrate that MedLIME consistently outperforms gradient-based and perturbation-based baselines in abnormality localization as measured by AUPRC. Our results highlight that incorporating generative reconstruction, adaptive perturbation and data-driven regularization improves the reliability and interpretability of medical imaging models.

Reliable generalization metrics are fundamental to both the development and evaluation of machine learning models. Especially in high-stakes applications where labeled target data are scarce, evaluation of models' generalization performance under distribution shift is a pressing need. We focus on two practical scenarios: (1) Before deployment, how to select the best model for unlabeled target data? (2) After deployment, how to monitor model performance under distribution shift? The central need in both cases is a reliable, label-free proxy metric. Yet existing proxy metrics, such as model confidence or accuracy-on-the-line, are often unreliable as they only assess model outputs while ignoring the internal mechanisms that produce them. We address this limitation by introducing a new perspective: using a model’s inner working, i.e. circuits, as a predictive metric of generalization performance. Leveraging circuit discovery, we extract the causal interactions between internal representations as a circuit, from which we derive two metrics tailored to the two practical scenarios. (1) Before deployment, we introduce Dependency Depth Bias, which measures different models' generalization capability on target data. (2) After deployment, we propose Circuit Shift Score, which predicts a model's generalization under different distribution shifts. Across diverse tasks, both metrics demonstrate significantly improved correlation with generalization performance, outperforming existing proxies by an average of 11.0% and 45.3%, respectively.

Sparse Autoencoders uncover thousands of features in vision models, yet explaining these features without requiring human intervention remains an open challenge. While previous work has proposed generating correlation-based explanations based on top activating input examples, we present a fundamentally different alternative based on causal interventions. We leverage the structure of Vision-Language Models and steer individual SAE features in the vision encoder after providing an empty image. Then, we prompt the language model to explain what it "sees", effectively eliciting the visual concept represented by each feature. Results show that Steering offers an scalable alternative that complements traditional approaches based on input examples, serving as a new axis for automated interpretability in vision models. Moreover, the quality of explanations improves consistently with the scale of the language model, highlighting our method as a promising direction for future research. Finally, we propose Steering-informed Top-k, a hybrid approach that combines the strengths of causal interventions and input-based approaches to achieve state-of-the-art explanation quality without additional computational cost.

Deep Neural Networks (DNNs) have revolutionized numerous industries, yet their decision-making processes remain largely opaque. Most existing explanation methods visualize the importance of image regions that influence a classifier's decisions, but they predominantly focus on identifying regions with positive contributions, often overlooking those with negative impacts. In this paper, we introduce a novel black-box explanation method, the Metropolis-Hastings Explainer (MHE), designed to provide confidence-faithful explanations. MHE enhances the fidelity of explanations by ensuring that the explained regions closely align with the original confidence score, sampling instances that best match the classifier’s confidence. Furthermore, MHE improves sampling efficiency by utilizing existing valid samples to explore more potential valid ones, reducing computational overhead. To enhance the clarity of explanations, MHE prioritizes valid samples with smaller areas when other factors are equal, thereby reducing the explanation area. Building upon the MHE framework, we propose two extensions: MHE-e, which focuses exclusively on regions with positive contributions, and MHE-pro, which refines explanation quality by integrating multi-scale information. MHE-pro progressively regions, optimizing both sampling efficiency and explanation quality. Experimental results demonstrate that MHE delivers superior and stable explanation quality across various models, including ResNet50, VGG16, ViT DINO, and CLIP, on datasets such as ImageNet, CUB-200-2011, and VOC2012, providing explanations that closely approximate the original classification confidence.

Unsupervised Concept Extraction aims to extract concepts from a single image, yet existing methods suffer from the inability to extract composable intrinsic concepts. To address this, this paper introduces a new task called Compositional and Interpretable Intrinsic Concept Extraction (CI-ICE). The CI-ICE task aims to leverage diffusion-based text-to-image models to extract composable object-level and attribute-level concepts from a single image, such that the original concept can be reconstructed through the combination of these concepts. To achieve this goal, we propose a method called HyperExpress, which addresses the CI-ICE task through two core aspects. Specifically, first, we propose a concept learning approach that leverages the inherent hierarchical modeling capability of hyperbolic space to achieve accurate concept disentanglement while preserving the hierarchical structure and relational dependencies among concepts; second, we introduce a concept-wise optimization method that maps the concept embedding space to maintain complex inter-concept relationships while ensuring concept composability. Our method demonstrates outstanding performance in extracting compositionally interpretable intrinsic concepts from a single image.

The performance of neural network models deteriorates due to their unreliable behavior on non-robust features of corrupted samples. Owing to their opaque nature, rectifying models to address this problem often necessitates arduous data cleaning and model retraining, resulting in huge computational and manual overhead. In this work, we leverage rank-one model editing to establish an attribution-guided model rectification framework that effectively locates and corrects model unreliable behaviors. We first distinguish our rectification setting from existing model editing, yielding a formulation that corrects unreliable behavior while preserving model performance and reducing reliance on large budgets of cleansed samples. We further reveal a bottleneck of model rectifying arising from heterogeneous editability across layers. To target the primary source of misbehavior, we introduce an attribution-guided layer localization method that quantifies layer-wise editability and identifies the layer most responsible for unreliabilities. Extensive experiments demonstrate the effectiveness of our method in correcting unreliabilities observed for neural Trojans, spurious correlations and feature leakage. Our method shows remarkable performance by achieving its editing objective with as few as a single cleansed sample, which makes it appealing for practice.

Deep vision models perform input-output computations that are hard to interpret. Concept-based explanation methods (CBEMs) increase interpretability by re-expressing parts of the model with human-understandable semantic units, or concepts. Checking if the derived explanations are faithful---that is, they represent the model's internal computation---requires a surrogate that combines concepts to compute the output. Simplifications made for interpretability inevitably reduce faithfulness, resulting in a tradeoff between the two. State-of-the-art unsupervised CBEMs (U-CBEMs) have reported increasingly interpretable concepts, while also being more faithful to the model. However, we observe that the reported improvement in faithfulness artificially results from either (1) using overly complex surrogates, which introduces an unmeasured cost to the explanation's interpretability, or (2) relying on deletion-based approaches that, as we demonstrate, do not properly measure faithfulness. We propose Surrogate Faithfulness (SURF), which (1) replaces prior complex surrogates with a simple, linear surrogate that measures faithfulness without changing the explanation's interpretability and (2) introduces well-motivated metrics that assess loss across all output classes, not just the predicted class. We validate SURF with a measure-over-measure study by proposing a simple sanity check---explanations with random concepts should be less faithful---which prior surrogates fail. SURF enables the first reliable faithfulness benchmark of U-CBEMs, revealing that many visually compelling U-CBEMs are not faithful. Code to be released.

Recent advances in point cloud In-Context Learning (ICL) have demonstrated strong multitask capabilities. Existing approaches typically adopt a Masked Point Modeling (MPM)-based paradigm for point cloud ICL. However, MPM-based methods directly predict the target point cloud from masked tokens without leveraging geometric priors, requiring the model to infer spatial structure and geometric details solely from token-level correlations via transformers. Additionally, these methods suffer from a training–inference objective mismatch, as the model learns to predict the target point cloud using target-side information that is unavailable at inference time. To address these challenges, we propose DeformPIC, a deformation-based framework for point cloud ICL. Unlike existing approaches that rely on masked reconstruction, DeformPIC learns to deform the query point cloud under task-specific guidance from prompts, enabling explicit geometric reasoning and consistent objectives. Extensive experiments demonstrate that DeformPIC consistently outperforms previous state-of-the-art methods, achieving reductions of 1.6, 1.8, and 4.7 points in average Chamfer Distance on reconstruction, denoising, and registration tasks, respectively. Furthermore, we introduce a new out-of-domain benchmark to evaluate generalization across unseen data distributions, where DeformPIC achieves state-of-the-art performance.

Hierarchical LiDAR geometry compression encodes voxel occupancies from low to high bit-depths, yet prior methods treat each depth independently and re-estimate local context from coordinates at every level, limiting compression efficiency. We present ELiC, a real-time framework that combines cross-bit-depth feature propagation, a Bag-of-Encoders (BoE) selection scheme, and a Morton-order-preserving hierarchy. Cross-bit-depth propagation reuses features extracted at denser, lower depths to support prediction at sparser, higher depths. BoE selects, per depth, the most suitable coding network from a small pool, adapting capacity to observed occupancy statistics without training a separate model for each level. The Morton hierarchy maintains global Z-order across depth transitions, eliminating per-level sorting and reducing latency. Together these components improve entropy modeling and computation efficiency, yielding state-of-the-art compression at real-time throughput on Ford and SemanticKITTI.Code and models will be released upon publication.

Online 3D scene perception in real time is essential for robotics, AR/VR, and autonomous systems, particularly in edge computing scenarios where computational resources are limited and privacy is crucial. Recent state-of-the-art methods like EmbodiedSAM (ESAM) demonstrate the promise of online 3D perception by leveraging the Segment Anything Model (SAM) for real-time, fine-grained, and gen- eralized 3D instance segmentation. However, ESAM still relies on a computationally expensive 3D sparse UNet for point cloud feature extraction, which accounts for the majority of the 3D inference time, hindering its practicality on resource-constrained devices. In this paper, we propose ESAM++, a lightweight and scalable alternative for online 3D scene perception tailored to edge devices without GPU acceleration. Our method introduces a 3D Sparse Feature Pyramid Network (SFPN) that efficiently captures multi-scale geometric features from streaming 3D point clouds while significantly reducing computational overhead and model size. We evaluate our approach on four challenging segmentation benchmarks, namely ScanNet, ScanNet200, SceneNN, and 3RScan, demonstrating that our model achieves competitive accuracy with up to 3× faster inference with a 2× smaller model size compared to ESAM, enabling practical deployment in real-world edge scenarios. Code and models will be publicly available.

Noisy, partially overlapping data and the need for real-time processing pose major challenges for rigid registration. Considering that feature-based matching can handle large transformation differences but suffers from limited accuracy, while local geometry-based matching can achieve fine-grained local alignment but relies heavily on a good initial transformation, we propose a novel dual-space paradigm to fully leverage the strengths of both approaches. First, we introduce an efficient filtering mechanism that incorporates a computationally lightweight single-point RANSAC algorithm followed by a refinement module to eliminate unreliable feature-based correspondences. Subsequently, we treat filtered correspondences as anchor points, extract geometric proxies, and formulate an effective objective function with a tailored solver to estimate the transformation. Experiments verify our method's effectiveness, as shown by achieving up to a 32x CPU-time speedup over MAC on KITTI with comparable accuracy. The code will be made publicly available.

Image-to-point-cloud (I2P) registration aims to align 2D images with 3D point clouds by establishing reliable 2D-3D correspondences. The drastic modality gap between images and point clouds makes it challenging to learn features that are both discriminative and generalizable, leading to severe performance drops in unseen scenarios. We address this challenge by introducing a heterogeneous graph framework that jointly refines cross-modal features and correspondences within a unified architecture. The proposed graph represents a mapping between segmented 2D and 3D regions, which enhances cross-modal feature interaction and thus improves feature discriminability. In addition, modeling the consistency among vertices and edges within the graph enables pruning of unreliable correspondences. Building on these insights, we propose a heterogeneous graph embedded I2P registration method, termed Hg-I2P. It learns a heterogeneous graph by mining multi-path feature relationships, adapts features under the guidance of heterogeneous edges, and prunes correspondences using graph-based projection consistency. Experiments on six indoor and outdoor benchmarks under cross-domain setups demonstrate that Hg-I2P significantly outperforms existing methods in both generalization and accuracy. Code will be released upon publication.

Both detection-then-match and detection-free methods have been extensively studied for image-to-point cloud registration, yet they still face significant challenges. The detection-then-match approach emphasizes high-quality correspondences but is limited by the availability of repeatable keypoints, making it susceptible to errors from incorrect matches. In contrast, detection-free methods aim for dense correspondences using a coarse-to-fine strategy to mitigate matching errors. However, non-overlapping regions and low-quality matches still introduce inaccuracies, and the differences between image texture and point cloud structure cause inconsistent region representations, increasing the likelihood of incorrect matches.To address these challenges, we propose two innovative modules: the High-Value Zone Reinforced Selection Module (HZRS) and the Zone Representation Consistency Alignment Module (ZRCA). HZRS employs reinforcement learning to resolve the non-differentiable issue of selecting high-value matching regions, while ZRCA improves region alignment through three stages: understand, coordinate, and accelerate.Extensive experiments and ablation studies on RGB-D Scenes v2 and 7-Scenes demonstrate the superiority of our network, establishing it as the state-of-the-art for image-to-point cloud registration.

Detecting small aerial targets (SATs) from long-range LiDAR is challenging because point density changes dramatically with motion: fast flights produce ultra-sparse returns, while hovering or slow motion yields dense local clusters, breaking fixed-voxel and static-threshold assumptions in standard 3D detectors and trackers. We introduce A3PRL, an RL-driven adaptive perception framework that closes the loop between LiDAR sensing and tracking. A3PRL builds on a sparsity-aware proposal stage with Temporal Dispersion Signatures and velocity-change cues, and deploys a lightweight 5D policy that jointly adjusts voxel resolution, detection sensitivity, and association gating based on purely label-free statistics summarizing spatio–temporal sparsity, foreground acceptance, and tracking continuity. The policy is trained with privileged supervision from ground-truth trajectories to shape a reward that balances geometric accuracy, temporal stability, and regularized acceptance, but runs fully label-free at test time. On the public MMAUD benchmark, training on V1 and evaluating on unseen V2/V3 domains, A3PRL reduces 3D localization error by about 19\% compared to its non-RL counterpart and consistently outperforms LiDAR-only and multimodal baselines under both day and night conditions. We further show that the same policy transfers to an in-house LiDAR–RTK setup and a public multi-LiDAR SAT dataset with heterogeneous scan patterns, where it maintains accurate trajectories and stable tracks under varying sparsity, while adding less than 2 ms per frame on a 10 Hz LiDAR budget.

Despite recent progress in 3D self-supervised learning, collecting large-scale 3D scene scans remains expensive and labor-intensive. In this work, we investigate whether 3D representations can be learned from unlabeled videos recorded without any real 3D sensors. We present Laplacian-Aware Multi-level 3D Clustering with Sinkhorn-Knopp (LAM3C), a self-supervised framework that learns from video-generated point clouds from unlabeled videos. We first introduce \data, a video-generated point cloud dataset constructed by collecting room-walkthrough videos from the web (e.g., real-estate tours) and generating 49,219 scenes using an off-the-shelf feed-forward reconstruction model. We also propose a noise-regularized loss that stabilizes representation learning by enforcing local geometric smoothness and ensuring feature stability under noisy point clouds. Remarkably, without using any real 3D scans, LAM3C achieves higher performance than the previous self-supervised methods on indoor semantic and instance segmentation. These results suggest that unlabeled videos represent an abundant source of data for 3D self-supervised learning.

We propose StreamVLO, a streaming visual–LiDAR odometry framework that performs unified spatio-temporal correlation with Mamba models and tackles the long-standing cumulative drift problem via an online Cumulative Drift Compensation scheme for localization in 4D dynamic environments. Specifically, StreamVLO introduces a unified spatio-temporal correlation module built on Mamba to fuse heterogeneous visual and LiDAR cues across multi-frame clips, overcoming the limited temporal exploration of prior pairwise methods. Furthermore, a Cumulative Drift Compensation module minimizes cumulative drift by iteratively learning residual corrections from multiple historical frames in a causal manner. To strengthen spatial feature representation on salient regions, we adopt a Keypoint-Aware Auxiliary Loss with a winner-takes-all strategy. StreamVLO achieves state-of-the-art performance on two commonly used autonomous driving datasets, reducing errors by 19\% ($t_{\text{rel}}$) and 22\% ($r_{\text{rel}}$) on KITTI, and by 18\% ATE and 16\% RPE on Argoverse, while remaining suitable for real-time deployment.

While recent Transformer and Mamba architectures have advanced point cloud representation learning, they are typically developed for single-task or single-domain settings. Directly applying them to multi-task domain generalization (DG) leads to degraded performance. Transformers effectively model global dependencies but suffer from quadratic attention cost and lack explicit structural ordering, whereas Mamba offers linear-time recurrence yet often depends on coordinate-driven serialization, which is sensitive to viewpoint changes and missing regions, causing structural drift and unstable sequential modeling. In this paper, we propose Structure-Aware Domain Generalization (SADG), a Mamba-based In-Context Learning framework that preserves structural hierarchy across domains and tasks. We design structure-aware serialization (SAS) that generates transformation-invariant sequences using centroid-based topology and geodesic curvature continuity. We further devise hierarchical domain-aware modeling (HDM) that stabilizes cross-domain reasoning by consolidating intra-domain structure and fusing inter-domain relations. At test time, we introduce a lightweight spectral graph alignment (SGA) that shifts target features toward source prototypes in the spectral domain without updating model parameters, ensuring structure-preserving test-time feature shifting. In addition, we introduce MP3DObject, a real-scan object dataset for multi-task DG evaluation. Comprehensive experiments demonstrate that the proposed approach improves structural fidelity and consistently outperforms state-of-the-art methods across multiple tasks including reconstruction, denoising, and registration.

Point cloud denoising is a critical preprocessing step for enhancing the reliability and accuracy of 3D perception systems. Most existing progressive denoising methods rely on fixed iterative pipelines that process all regions uniformly, resulting in redundant computation and over-smoothing of geometric details when handling point clouds with non-uniform noise distributions. To overcome these limitations, we introduce Dynamic Skip Net (DSNet), a novel progressive denoising framework that adaptively determines the optimal denoising path for each local patch based on its noise characteristics. DSNet incorporates a noise discriminator that quantifies local noise intensity by analyzing normal similarity, and a reverse monotonic decision function that maps this measure to an appropriate denoising module. Furthermore, we propose a Path-Selective Iteration mechanism that dynamically re-evaluates the restoration state and re-plans the denoising route at each stage, enabling cross-stage skipping to minimize unnecessary computation. Extensive experiments on multiple benchmarks demonstrate that DSNet achieves state-of-the-art performance in noise suppression, geometric fidelity, and computational efficiency. Our code and models will be made publicly available at github.

Point cloud completion is an important yet challenging problem in 3D computer vision, which aims to reconstruct complete and dense 3D shapes from partial point clouds. Although transformer-based and geometry-based approaches have made significant progress, they often struggle to capture the complex, high-order correlations inherent in point clouds. To address this limitation, we propose Hyper-PCN, a point cloud completion framework that leverages hypergraphs to explicitly model complex, higher-order correlations within incomplete inputs for more accurate completion. It comprises two key modules: Hyper Refinement Stack, designed to progressively capture coarse-to-fine high-order correlations through a series of hypergraph learning stages, and Anchor-based Hypergraph Neural Network, which employs a two-stage sampling strategy to construct collaborative hypergraphs, ensuring robust modeling of global structures. Extensive experiments on multiple datasets demonstrate that our approach consistently outperforms state-of-the-art methods.

Prompt tuning of large-scale vision-language models such as CLIP enables efficienttask adaptation without updating model weights. However, it often leads to poorconfidence calibration and unreliable predictive uncertainty. We address thisproblem by proposing a calibration framework that enhances predictive reliabilitywhile preserving the geometry of the pretrained CLIP embedding space, which isrequired for robust generalization. Our approach extends the standard cross-entropyloss with two complementary regularizers: (1) a mean–variance margin penalty thatstabilizes inter-class logit margins by maximizing their average while minimizingdispersion, mitigating underconfidence and overconfidence spikes; and (2) a textmoment-matching loss that aligns the first and second moments of tuned textembeddings with their frozen CLIP counterparts, preserving semantic dispersioncrucial for generalization. Through extensive experiments across 7 prompt-tuningmethods and 11 diverse datasets, we demonstrate that our approach significantlyreduces the Expected Calibration Error (ECE) compared to competitive calibrationtechniques on both base and novel classes.

Fine-tuning large language models (LLMs) using reinforcement learning (RL) objectives has gained traction, especially in scenarios where labeled data is limited. Building on its success in the language domain, recent efforts have extended RL-based fine-tuning to multimodal tasks. Visual-RFT, for instance, applied Group Relative Policy Optimization (GRPO) to fine-tune multimodal LLMs (MLLMs) across various visual perception benchmarks, achieving notable improvements over standard supervised fine-tuning (SFT). However, its scope was limited by a narrow evaluation of RL adaptation strategies. In this work, we expand the landscape by introducing new RL-based baselines on the same benchmarks and conducting a deeper analysis of GRPO’s training dynamics. We identify key limitations—such as reduced generation diversity, constrained policy exploration, and suboptimal reward formulation and aggregation. To address these, we propose DEVA: a framework that enhances Diversity via a flow-based training objective, encourages broader policy Exploration through global entropic regularization, and leverages alignment Volume as a non-verifiable reward combined with harmonic Aggregation. Applied to GRPO and other RL methods, DEVA delivers consistent gains in both quantitative (+5 to +13 points) and qualitative metrics. We further provide visualizations, ablations, and analyses to unpack the contributions of each component in our framework.

Prompt Learning (PL) has emerged as a parameter-efficient technique for adapting Vision-Language Models (VLMs) to downstream tasks. However, almost all existing PL methods are primarily designed and evaluated on well-curated datasets, overlooking a critical post-deployment phenomenon, i.e., the intrinsic connection between input resolution and storage-memory consumption. Specifically, to satisfy the stringent storage-memory constraints on edge devices, models are often limited to low-resolution inputs (e.g., $\le$ 224$\times$224 for CLIP-ViT/B-16) and generate fewer tokens (with the position embedding resized), which poses a unique challenge in performance robustness. To tackle this issue, we propose LOREAL, an efficient prompt self-distillation framework that learns resolution-invariant representations by excavating attribute semantics. At the heart of LOREAL is a dual-student architecture, i.e., two student models fed with inputs at different resolutions synergistically learn from each other. Building upon this, we contextualize the students' prompt with resolution-invariant attributes queried from the LLM, then leverage cross-modality meta-nets to generate attribute semantics. These meta-nets are bridged between the different encoders of two students, wherein we introduce Low-Level Distillation (LLD) and High-Level Distillation (HLD) to facilitate the learning of more cross-resolution representations. Extensive experiments show that LOREAL significantly improves VLMs' performance and robustness under varied resolution settings, underscoring significant practical utilities.

This paper provides a simplification on OpenVision's architecture and loss design for enhancing its training efficiency. Following the prior vision-language pretraining works CapPa and AIMv2, as well as modern multimodal designs like LLaVA, our changes are straightforward: we remove the text encoder (and therefore the contrastive loss), retaining only the captioning loss as a purely generative training signal. We name this new version OpenVision 2. The initial results are promising: despite this simplification, OpenVision 2 competitively matches the original model's performance on a broad set of multimodal benchmarks while substantially cutting both training time and memory consumption. For example, with ViT-L/14, it reduces training time by about 1.5x (from 83h to 57h), and memory usage by about 1.8x (from 24.5GB to 13.8GB, equivalently allowing the maximum batch size to grow from 2k to 8k). This superior training efficiency also allows us to scale far beyond the largest vision encoder used in OpenVision, reaching more than 1 billion parameters. We hold a strong belief that this lightweight, generative-only paradigm is compelling for future vision encoder development in multimodal foundation models.

Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities in visual reasoning across diverse tasks. These tasks place different demands on visual representations: some prioritize high-level global context, while others emphasize fine-grained local details. However, most existing methods operate on visual representations primarily in the spatial domain, lacking an explicit mechanism for distinguishing between high-frequency local details and low-frequency global context. This limitation hinders fine-grained control of visual representations and complicates their hierarchical alignment with language. To address this issue, we introduce Language-guided Frequency Modulation (LFM), a plug-and-play approach that adaptively refines visual signals in the frequency domain under linguistic guidance. By selectively enhancing critical regions and details, LFM enables more structured and precise visual processing. Crucially, it adds no extra training parameters, relying solely on a lightweight learnable projector to refine visual tokens before integration into the LLM, thereby ensuring minimal computational overhead. Extensive experiments across diverse vision-language benchmarks highlight LFM’s scalability, effectiveness, and broad applicability to LVLMs. The code will be publicly available.

Fine-tuning large-scale Vision-Language Models (VLMs) is crucial for specialized tasks, but their performance is often undermined by the label noise prevalent in real-world datasets. Traditional approaches to learning with noisy labels typically rely on a self-referential loop, using a model's own predictions to correct errors. While recent VLM-specific methods have begun to leverage cross-modal information to aid in noise detection, we explore an alternative direction: using the text modality not just to identify noise, but to establish a source of ground truth that is fully independent of the training data's potentially corrupt labels.To this end, we propose \textbf{T}ext-\textbf{AN}chored \textbf{G}uided \textbf{O}ptimization (TANGO), a framework centered on ``semantic anchors"—a set of pure, immutable reference points generated from diverse text descriptions. Building upon these anchors, our approach reframes two key aspects of learning with noisy labels. First, we replace the conventional linear classifier with a parameter-free Text-Anchored Classifier, making predictions a direct, weighted consensus of the clean anchors. Second, we introduce an Anchor-Guided Refinement mechanism that validates each sample's given label against this external ground truth, providing a more robust signal for sample selection and label correction. Extensive experiments demonstrate that this approach achieves competitive and often state-of-the-art performance. Our code will be publicly available.

Large-scale video-language pretraining enables strong generalization across multimodal tasks but often incurs prohibitive computational costs. Although recent advances in masked visual modeling help mitigate this issue, they still suffer from two fundamental limitations: severe visual information loss under high masking ratios and temporal information leakage caused by inter-frame correlations. To address these challenges, we propose ClusterSTM, a Cluster-Wise Spatio-Temporal Masking strategy for efficient video-language pretraining. ClusterSTM first performs intra-frame clustering to partition visual tokens into multiple semantically independent clusters, then conducts cluster-wise masking by retaining the token with the highest temporal density within each cluster. Our masking strategy ensure that the retained tokens capture holistic video content while exhibit strong temporal correlation. Additionally, we introduce a video–text relevance reconstruction objective that aligns high-level multimodal semantics beyond conventional visual reconstruction. Extensive experiments across multiple benchmarks demonstrate that ClusterSTM achieves superior performance on video-text retrieval, video question answering, and video captioning tasks, establishing a new state-of-the-art among efficient video-language models.

Large-scale vision–language models (VLMs) such as CLIP have excelled in zero-shot image classification, yet they struggle to achieve the dense cross-modal alignment required by open-vocabulary dense perception (OVDP). While recent self-distillation methods address this by aligning dense features with the generalizable global semantics, a key question remains: how should such dense features be constructed to achieve optimal alignment? To address this, we propose DenseRC, a principled $\textbf{Dense}$ $\textbf{R}$epresentations $\textbf{C}$onstruction framework that reconstructs CLIP for OVDP based on two key insights.First, by analyzing the internal semantics encoded in the global $\textit{cls}$ token, we identify that multi-layer value embeddings serve as an informative basis for dense features. Second, we reveal that spatial aggregation tends to amplify semantic misalignment. Motivated by this, we design a lightweight Head-Selective Gating (HSG) module that adaptively reweights feature heads according to their intrinsic heterogeneity, enabling discriminative and alignment-friendly dense representations construction. Extensive experiments demonstrate that DenseRC delivers consistent and substantial gains across OVDP tasks including object detection and semantic segmentation, setting new state-of-the-art performance on multiple benchmarks.

The performance of Vision–Language Transformers drops sharply when an input modality (e.g., image) is missing, because the model is forced to make predictions using incomplete information. Existing missing-aware prompt methods help reduce this degradation, but they still rely on conventional prediction heads (e.g., a Fully-Connected layer) that compute class scores in the same way regardless of which modality is present or absent. We introduce Decoupled Prototype Learning (DPL), a new prediction head architecture that explicitly adjusts its decision process to the observed input modalities. For each class, DPL selects a set of prototypes specific to the current missing-modality cases (image-missing, text-missing, or mixed-missing). Each prototype is then decomposed into image-specific and text-specific components, enabling the head to make decisions that depend on the information actually present. This adaptive design allows DPL to handle inputs with missing modalities more effectively while remaining fully compatible with existing prompt-based frameworks. Extensive experiments on MM-IMDb, UPMC Food-101, and Hateful Memes demonstrate that DPL outperforms state-of-the-art approaches across all widely used multimodal image–text datasets and various missing cases.

Due to the heterogeneity of faces and edges in B-rep, conventional graph-based representations is incapable of establishing a unified formulation for faces and edges, thereby constraining the capabilities of B-rep generative models. We propose a B-rep Variational Graph Auto Encoding (BrepVGAE), the first variational graph autoencoder framework capable of holistically encoding and decoding boundary representations of B-rep models.Firstly, we novelly represent both geometry faces and edges as nodes in a graph representation. We then design a sparse graph autoencoder to aggregate the complete B-rep structure into a compact global latent vector. We then construct a decoder that employs set-based generation, which uses bilinear layers to reconstruct adjacency relationships, i.e., topology, with a single latent vector. Afterwards, the same decoder generates node features for all faces and edges through learnable query vectors and cross-attention mechanisms. Finally, a two-stage training strategy ensures effective coupling of geometry and topology throughout. Comprehensive experiments demonstrate that BrepVGAE significantly outperforms existing methods in reconstruction accuracy, topological validity, and generative diversity. This validates the feasibility and efficacy of decoding complete CAD geometric-topological distributions from a unified latent representation, while also offering novel insights for CAD part retrieval and feature recognition domains.

Data-driven approaches have revolutionized 3D vision, enabling transformers to effectively reconstruct and generate static 3D objects. However, generating simulative 4D dynamics---realistic temporal deformations of static objects under various physical conditions---remains challenging and often ad hoc despite being critical for building comprehensive 3D world models. Most existing methods assume a predefined physical model and use system identification to estimate parameters, restricting these methods to specific categories and small-scale datasets. We propose that these restrictions can be overcome by learning a data-driven kinematic state parameterization for object-centric physical systems. Specifically, we learn both a latent space of all possible states of the object and a decoder that maps any sampled latent to a plausibly deformed shape of the object. We refer to this parameterization as Neural Object Kinematics (NeuROK), and learn a transformer-based encoder-decoder model on a curated large-scale 4D dataset. This formulation and the learned model significantly simplify the generation of simulative dynamics since we only need to consider the dynamics within a low-dimensional latent space from the Lagrangian mechanics' perspective in classical physics. We demonstrate the effectiveness and generality of this framework across diverse dynamic object types, showing clear advantages over prior works.

We train a language model to generate LEGO®-brick build sequences. While prior work has been restricted to discrete, voxel-like towers, we consider a much broader set of pieces, encompassing thousands of part types with diverse connection semantics. To enable this, we first collect a large-scale dataset of over 100,000 human-designed LDraw brick objects and scenes. The complexity of our setting makes it challenging to autoregressively assemble structures that satisfy physical constraints. When predicting block pose directly, build sequences quickly become invalid after a small number of steps. Although pieces are placed in 3D space, it is the spatial relationships of the parts which define the whole. With this in mind, we design a graph-based program representation that parametrizes structure through connectivity, improving the physical grounding of generated sequences. To enable future applications, we make our dataset and models available for research purposes. https://kulits.github.io/BrickNet

Automatic residential floorplan generation has long been a central challenge bridging architecture and computer graphics, aiming to make spatial design more efficient and accessible. While early methods based on constraint satisfaction or combinatorial optimization ensure feasibility, they lack diversity and flexibility. Recent generative models achieve promising results but struggle to generalize across heterogeneous conditional tasks, such as generation from site boundaries, room adjacency graphs, or partial layouts, due to their suboptimal representations. To address this gap, we introduce Floorplan Markup Language (FML), a general representation that encodes floorplan information within a single structured grammar, which casts the entire floorplan generation problem into a next token prediction task. Leveraging FML, we develop a transformer-based generative model, Floorplan Markup Language Model (FMLM), capable of producing high-fidelity and functional floorplans under diverse conditions. Comprehensive experiments on the RPLAN dataset demonstrate that FMLM, despite being a single model, surpasses the previous task-specific state-of-the-art methods.

Computer-Aided Design (CAD) is essential in industrial design, but the complexity of traditional CAD modeling and workflows presents significant challenges for automating the generation of high-precision, editable CAD models. Existing methods, such as 3D reconstruction from sketches, often produce non-editable, approximate models that fall short of meeting the stringent requirements for precision and editability in industrial design. Moreover, the reliance on text or image-based inputs often requires significant manual annotation, limiting their scalability and applicability in industrial settings. To overcome these challenges, we propose the Heterogeneous Collaborative Multi-Expert Reinforcement Learning (CME-CAD) paradigm, a novel training paradigm for CAD code generation. Our approach integrates the complementary strengths of these models, facilitating collaborative learning and improving the model’s ability to generate accurate, constraint-compatible, and fully editable CAD models. We introduce a two-stage training process: Multi-Expert Fine-Tuning (MEFT), and Multi-Expert Reinforcement Learning (MERL). Additionally, we present CADExpert, an open-source benchmark consisting of 17,299 instances, including orthographic projections with precise dimension annotations, expert-generated Chain-of-Thought (CoT) processes, executable CADQuery code, and rendered 3D models.

In this paper, we propose Robo-SGG, a plug-and-play module for robust scene graph generation (SGG). Unlike standard SGG, the robust scene graph generation aims to perform inference on a diverse range of corrupted images, with the core challenge being the domain shift between the clean and corrupted images. Existing SGG methods suffer from degraded performance due to shifted visual features (e.g., corruption interference or occlusions). To obtain robust visual features, we leverage layout information, representing the global structure of an image, which is robust to domain shift, to enhance the robustness of SGG methods under corruption. Specifically, we employ Instance Normalization (IN) to alleviate the domain-specific variations and recover the robust structural features (i.e., the positional and semantic relationships among objects) by the proposed Layout-Oriented Restitution. Furthermore, under corrupted images, we introduce a Layout-Embedded Encoder (LEE) that adaptively fuses layout and visual features via a gating mechanism, enhancing the robustness of positional and semantic representations for objects and predicates. Note that our proposed Robo-SGG module is designed as a plug-and-play component, which can be easily integrated into any baseline SGG model. Extensive experiments demonstrate that by integrating the state-of-the-art method into our proposed Robo-SGG, we achieve relative improvements of 6.3%, 11.1%, and 8.0% in mR@50 for PredCls, SGCls, and SGDet tasks on the VG-C benchmark, respectively, and achieve new state-of-the-art performance in the corruption scene graph generation benchmark (VG-C and GQA-C). We will release our source code and model.

OmniLottie is a versatile framework that generates high-quality vector animations from multi-modal instructions, including interleaved texts, images, and videos. To fully parameterize vector animations for flexible motion and visual content control, we seek help from the Lottie representation, which encodes both shapes and animated behaviors in a single JSON file. Building upon a pretrained vision–language model (VLM), OmniLottie produces vivid, semantically aligned vector animations that adhere closely to multi-modal conditions. To avoid the complexity and irregularity of raw JSON structures, we introduce a dedicated Lottie tokenizer that transforms Lottie files into structured sequences of function calls representing shapes, animation commands, and their parameters. This design enables the model to directly learn the underlying shape and animation priors from data, substantially improving generation stability and controllability. To further advance research in vector animation generation, we curate MMLottie-2M, a large-scale dataset of professionally designed vector animations paired with textual and visual annotations. Leveraging the well-designed tokenizer and our newly established dataset, OmniLottie demonstrates strong multi-modal conditional generation capabilities using a simple next-token prediction objective. For qualitative results, please refer to the generated animations rendered through standard Lottie players on the supplementary website.

Ancient inscriptions, as repositories of cultural memory, have suffered centuries of environmental and human-induced degradation. Restoring their intertwined visual and textual integrity poses one of the most demanding challenges in digital heritage preservation. However, existing AI-based approaches often rely on rigid pipelines, struggling to generalize across such complex and heterogeneous real-world degradations.Inspired by the skill-coordinated workflow of human epigraphers, we propose EpiAgent, an agent-centric system that formalizes inscription restoration as a hierarchical planning problem. Following an Observe–Conceive–Execute–Reevaluate paradigm, an LLM-based central planner orchestrates collaboration among multimodal analysis, historical experience, specialized restoration tools, and iterative self-refinement. This agent-centric coordination enables a flexible and adaptive restoration process beyond conventional single-pass methods.Across real-world degraded inscriptions, EpiAgent delivers superior restoration quality and stronger generalization compared to existing methods. Our work marks a pivotal step toward expert-level agent-driven restoration of cultural heritage. Code will be released.

Deep Neural Networks achieve high performance in vision tasks by learning features from regions of interest (ROI) within images, but their performance degrades when deployed on out-of-distribution (OOD) data that differs from training data. This challenge has led to OOD detection methods that aim to identify and reject unreliable predictions. Although prior work shows that OOD detection performance varies by artefact type, the underlying causes remain underexplored. To this end, we identify a previously unreported bias in OOD detection: for hard-to-detect artefacts (near-OOD), detection performance typically improves when the artefact shares visual similarity (e.g. colour) with the model’s ROI and drops when it does not - a phenomenon we term the Invisible Gorilla Effect. For example, in a skin lesion classifier with red lesion ROI, we show the method Mahalanobis Score achieves a 31.5% higher AUROC when detecting OOD red ink (similar to ROI) compared to black ink (dissimilar) annotations. We annotated artefacts by colour in 11,355 images from three public datasets (e.g. ISIC) and generated colour-swapped counterfactuals to rule out dataset bias. We then evaluated 40 OOD methods across 7 benchmarks and found significant performance drops for most methods when artefacts differed from the ROI. Our findings highlight an overlooked failure mode in OOD detection and provide guidance for more robust detectors. Code and annotations will be released upon acceptance.

The rapid advancement of Vision-Language models (VLMs) has raised growing concern that their black-box reasoning processes could lead to unintended forms of social bias. Current debiasing approaches focus on mitigating surface-level bias signals through post-hoc learning or test-time algorithms, while leaving the internal dynamics of the model largely unexplored. In this work, we introduce an interpretable, model-agnostic bias mitigation framework, DEBIASLENS, that localizes social attribute neurons in VLMs through sparse autoencoders (SAEs) applied to multimodal encoders. Building upon the disentanglement ability of SAEs, we train them on facial image or caption datasets without corresponding social attribute labels to uncover neurons highly responsive to specific demographics, including those that are less represented. By selectively deactivating the social neurons most strongly tied to bias for each group, we effectively mitigate socially biased behaviors of VLMs without degrading their semantic knowledge. Our research lays the groundwork for future auditing tools, prioritizing social fairness in emerging real-world AI systems.

Out-of-distribution (OOD) detection is critical to ensure the safe deployment of deep learning models in critical applications. Deep learning models can often misidentify OOD samples as in-distribution (ID) samples. This vulnerability worsens in the presence of spurious correlation in the training set. Likewise, in fine-grained classification settings, detection of fine-grained OOD samples becomes inherently challenging due to their high similarity to ID samples. However, current research on OOD detection has focused instead largely on relatively easier (conventional) cases. Even the few recent works addressing these challenging cases rely on carefully curated or synthesized outliers, ultimately requiring external data. This motivates our central research question: ``Can we innovate OOD detection training framework for fine-grained and spurious settings \textbf{without requiring any external data at all?}" In this work, we present a unified \textbf{A}pproach to \textbf{S}purious, fine-grained, and \textbf{C}onventional \textbf{OOD D}etection (\textbf{\ASCOOD}) that eliminates the reliance on external data. First, we synthesize virtual outliers from ID data by approximating the destruction of invariant features. Specifically, we propose to add gradient attribution values to ID inputs to disrupt invariant features while amplifying true-class logit, thereby synthesizing challenging near-manifold virtual outliers. Then, we simultaneously incentivize ID classification and predictive uncertainty towards virtual outliers. For this, we further propose to leverage standardized features with z-score normalization. ASCOOD effectively mitigates impact of spurious correlations and encourages capturing fine-grained attributes. Extensive experiments across \textbf{7} datasets and and comparisons with \textbf{30+} methods demonstrate merit of ASCOOD in spurious, fine-grained and conventional settings.

As multimodal models like CLIP become integral to downstream systems, the need to remove sensitive information is critical. However, machine unlearning for contrastively-trained encoders remains underexplored, and existing evaluations fail to diagnose fine-grained, association-level forgetting. We introduce SALMUBench (Sensitive Association-Level Multimodal Unlearning), a benchmark built upon a synthetic dataset of 60K persona-attribute associations and two foundational models: a \textit{Compromised} model polluted with this data, and a \textit{Clean} model without it. Both are trained from scratch on a 400M-pair \texttt{retain} set to isolate unlearning effects. We propose a novel evaluation protocol with structured holdout sets (\texttt{holdout_identity}, \texttt{holdout_association}) to precisely measure unlearning efficacy and collateral damage. Our benchmark reveals that while utility-efficient deletion is feasible, current methods exhibit distinct failure modes: they either fail to forget effectively or over-generalize by erasing more than intended. SALMUBench sets a new standard for comprehensive unlearning evaluation, and we publicly release our dataset, models, evaluation scripts, and leaderboards to foster future research.

The emergence of Large Language Models (LLMs) has driven rapid progress in multi-modal learning, particularly in the development of Large Vision-Language Models (LVLMs). However, existing LVLM training paradigms place excessive reliance on the LLM component, giving rise to two critical robustness challenges: language bias and language sensitivity. To address both issues simultaneously, we propose a novel Self-Critical Inference (SCI) framework that extends Visual Contrastive Decoding by conducting multi-round counterfactual reasoning through both textual and visual perturbations. This process further introduces a new strategy for improving robustness by scaling the number of counterfactual rounds. Moreover, we also observe that failure cases of LVLMs differ significantly across models, indicating that fixed robustness benchmarks may not be able to capture the true reliability of LVLMs. To this end, we propose the Dynamic Robustness Benchmark (DRBench), a model-specific evaluation framework targeting both language bias and sensitivity issues. Extensive experiments show that SCI consistently outperforms baseline methods on DRBench, and that increasing the number of inference rounds further boosts robustness beyond existing single-step counterfactual reasoning methods.

Recently, multimodal large language models (MLLMs) have emerged as a unified paradigm for language and image generation. Our work shows that MLLMs pair usability with higher risks, highlighting the need for adaptive safeguards to mitigate real-world harms.Compared with diffusion models, MLLMs possess a much stronger capability for semantic understanding, enabling them to process more complex textual inputs and comprehend richer contextual meanings. However, this enhanced semantic ability may also introduce new and potentially greater safety risks.Taking diffusion models as a reference point, we systematically analyze and compare the safety risks of emerging MLLMs along two dimensions: unsafe content generation and fake image synthesis.Across multiple unsafe generation benchmark datasets, we observe that MLLMs tend to generate more unsafe images than diffusion models. This difference partly arises because diffusion models often fail to interpret abstract prompts, producing corrupted outputs, whereas MLLMs can comprehend these prompts and generate unsafe content.For current advanced fake image detectors, MLLM-generated images are also notably harder to identify. Even when detectors are retrained with MLLMs-specific data, they can still be bypassed by simply providing MLLMs with longer and more descriptive inputs.Our measurements indicate that the emerging safety risks of the cutting-edge generative paradigm, MLLMs, have not been sufficiently recognized, posing new challenges to real-world safety.

We propose IrisFP, a novel adversarial-example-based model fingerprinting framework that enhances both uniqueness and robustness by leveraging multi-boundary characteristics, multi-sample behaviors, and fingerprint discriminative power assessment to generate composite-sample fingerprints. Three key innovations make IrisFP outstanding: 1) It positions fingerprints near the intersection of all decision boundaries—unlike prior methods that target a single boundary—thus increasing the prediction margin without placing fingerprints deep inside target class regions, enhancing both robustness and uniqueness; 2) It constructs composite-sample fingerprints, each comprising multiple samples close to the multi-boundary intersection, to exploit collective behavior patterns and further boost uniqueness; and 3) It assesses the discriminative power of generated fingerprints using statistical separability metrics developed based on two reference model sets respectively for pirated and independently-trained models, and assigns fingerprint-specific thresholds to retained fingerprints. Extensive experiments show that IrisFP consistently outperforms state-of-the-art methods, achieving reliable ownership verification by enhancing both robustness and uniqueness.

Embedding invisible and temporally consistent watermarks into dynamic 4D Gaussian Splatting (4DGS) models poses unique challenges due to continuous spatio-temporal deformation of Gaussians and diverse motion dynamics.Existing 3DGS watermarking methods, which directly fine-tune parameters within Gaussian splats, fail to preserve geometric fidelity and temporal coherence when extended to dynamic 4D settings. In this regime, we propose Mark4D, a temporally consistent watermarking technique that achieves robust, imperceptible, and motion-aware watermark embedding for 4DGS. Mark4D comprises 1) a decoder for watermark recovery in the latent video–text space for robustness against pixel-level distortions, 2) trajectory-aligned offsets that embed watermark signals along Gaussian motion paths to preserve geometry, and 3) a motion-adaptive loss weighting strategy that balances supervision across frames with varying motion intensities. Extensive experiments on synthetic and real-world dynamic scene datasets demonstrate that Mark4D achieves superior bit accuracy, visual fidelity, and robustness under diverse distortions, establishing a foundation for secure and reliable protection of dynamic 4D scene assets. The implementation of Mark4D will be released upon publication.

Machine Unlearning (MU) focuses on removing the influence of training samples from pre-trained models without retraining the model entirely. Existing MU methods have made several efforts to enable complete forgetting while preserving the model’s performance on remaining data. However, they typically apply equal weights across different data, overlooking the ambiguous decision boundaries between similar samples or approximate classes. This leads to unnecessary consumption of shallowly memorized samples and significant performance degradation for approximate retention classes. Additionally, the inherent inconsistency between forgetting and retention objectives results in gradient conflict and domination problems during training, hindering model convergence and degrading overall performance. To address these, we introduce a novel adaptive gradient reweighting that assigns importance weights to individual forget samples or vulnerable retention classes, thereby enabling more efficient unlearning and preserving the performance of approximate classes. Subsequently, we propose a multi-stage objective optimization strategy, which comprises three optimization stages: Direction Rectification, Temporal Stabilization, and Adaptive Objective Combination. This strategy rectifies the direction of conflicting gradients and prevents one task (forgetting or retention) from dominating the model update. Comprehensive analyses and extensive experiments on multiple public datasets demonstrate that our method achieves considerable performance improvements in various tasks and scenarios.

Prototype-based Personalized Federated Learning (ProtoPFL) enables efficient cross-domain adaptation by communicating compact class prototypes, but directly sharing prototypes raises privacy risks. A common defense involves per-example $\ell_2$ clipping before prototype computation to limit sensitivity, followed by the addition of isotropic Gaussian noise during upload to enforce Local Differential Privacy (LDP). However, this Isotropic Gaussian Prototype Perturbation (IGPP) often over-perturbs key discriminative dimensions and struggles to balance the clipping threshold with representation fidelity. We propose VPDR, a client-side privacy plug-in that can be seamlessly integrated into existing ProtoPFL frameworks. Motivated by the statistical prior that dimension-wise class variance reflects discriminability, we introduce Variance-adaptive Prototype Perturbation (VPP), which uses groupwise calibration to apply less noise to discriminative subspaces, preserving semantic separability while ensuring privacy. We further design Distillation-guided Clipping Regularization (DCR), which enables feature norms to adaptively concentrate near the predefined clipping threshold while maintaining prediction consistency. Theoretical analysis shows that our groupwise noise provides privacy guarantees no weaker than those of the isotropic mechanism under the same privacy constraints. Extensive experiments on multiple cross-domain benchmarks demonstrate that VPDR achieves a superior privacy-utility trade-off, outperforming IGPP in personalized federated fine-tuning while maintaining strong privacy protection under realistic attack scenarios.

Federated Learning (FL) faces a fundamental dilemma: existing defenses against gradient leakage attacks (GLAs) invariably sacrifice model performance for privacy protection through noise injection or gradient clip. We introduce Federated Learning with Momentum-Based Orthogonal Projection (FedMOP), a method that simultaneously achieves strong privacy guarantees and superior model performance. The key insight is to leverage initialization-based offset mechanisms that operate on orthogonal dimensions. For performance enhancement, FedMOP employs gradient orthogonal projection to counteract local drift, effectively offsetting each client's round-training initial model using global statistical context. For privacy protection, it introduces momentum-based trajectory offset hiding, which makes the offset vector inherently unrecoverable by constructing information barriers through private initialization and randomized evolution. These two mechanisms are synergistic rather than antagonistic. Theoretically, we prove convergence preservation and characterize the computationally infeasible inverse problem faced by attackers. Extensive experiments on CIFAR-10/100 and Tiny-ImageNet demonstrate that FedMOP not only defends effectively against state-of-the-art GLAs but also surpasses existing FL methods in both accuracy and convergence speed, validating its ability to jointly enhance privacy and performance in FL.

Federated Learning (FL) is a decentralized approach where multiple clients collaboratively train a shared global model without sharing their raw data. Despite its effectiveness, conventional FL faces scalability challenges due to excessive computational and communication demands placed on a single central server as the number of participating devices grows. Hierarchical Federated Learning (HFL) addresses these issues by distributing model aggregation tasks across intermediate nodes (stations), thereby enhancing system scalability and robustness against single points of failure. However, HFL still suffers from a critical yet often overlooked limitation: domain shift, where data distributions vary significantly across different clients and stations, reducing model performance on unseen target domains. While Federated Domain Generalization (FedDG) methods have emerged to improve robustness to domain shifts, their integration into HFL frameworks remains largely unexplored. In this paper, we formally introduce Hierarchical Federated Domain Generalization (HFedDG), a novel scenario designed to investigate domain shift within hierarchical architectures. Specifically, we propose HFedATM, a hierarchical aggregation method that first aligns the convolutional filters of models from different stations through Filter-wise Optimal Transport Alignment and subsequently merges aligned models using a Shrinkage-aware Regularized Mean Aggregation. Our extensive experimental evaluations demonstrate that HFedATM significantly boosts the performance of existing FedDG baselines across multiple datasets and maintains computational and communication efficiency. Moreover, theoretical analyses indicate that HFedATM achieves tighter generalization error bounds compared to standard hierarchical averaging, resulting in faster convergence and stable training behavior.

Federated Learning (FL) is plagued by two key challenges: high communication overhead and performance collapse on heterogeneous (non-IID) data. Analytic FL (AFL) provides a single-round, data distribution invariant solution, but is limited to linear models. Subsequent non-linear approaches, like DeepAFL, regain accuracy but sacrifice the single-round benefit. In this work, we break this trade-off. We propose SAFLe, a framework that achieves scalable non-linear expressivity by introducing a structured head of bucketed features and sparse, grouped embeddings. We prove this non-linear architecture is mathematically equivalent to a high-dimensional linear regression. This key equivalence allows SAFLe to be solved with AFL's single-shot, invariant aggregation law. Empirically, SAFLe establishes a new state-of-the-art for analytic FL, significantly outperforming both linear AFL and multi-round DeepAFL in accuracy across all benchmarks, demonstrating a highly efficient and scalable solution for federated vision.

Federated Prompt Learning (FPL) has recently attracted increasing attention for its ability to leverage large-scale vision-language models such as CLIP within federated learning frameworks. While existing studies have advanced FPL through personalization strategies to enhance client-specific performance, personalized models often suffer severe degradation when deployed across unseen domains due to distribution shifts.In this paper, we take the first step toward exploring Test-Time FPL (TTFPL), aiming to bridge the cross-domain performance gap with minimal effort, requiring only unlabeled target-domain data. We propose COTE, a tri-prompt controllable TTFPL framework that dynamically balances three complementary prompts: the global prompt from standard FPL, the local prompt from personalized FPL, and the frozen CLIP prompt.Specifically, we introduce a novel confidence-guided Model-Data Alignment (MoDA) metric in COTE that quantifies alignment at both macro and micro levels, capturing the consistency between model predictions and data distributions. By integrating MoDA with model confidence, COTE adaptively adjusts the contribution of each prompt at test time, enabling robust generalization across heterogeneous clients and unseen domains without requiring labeled data.Extensive experiments on multiple benchmark datasets demonstrate that our method consistently improves target-domain performance, setting a new direction for adaptive FPL.

Federated learning (FL) enables collaborative training across clients without compromising privacy. While most existing FL methods assume homogeneous model architectures, client heterogeneity in data and resources renders this assumption impractical, motivating model-heterogeneous FL. To address this problem, we propose Federated Representation Entanglement (FedRE), a framework built upon a novel form of client knowledge termed entangled representation. In FedRE, each client aggregates its local representations into a single entangled representation using normalized random weights and applies the same weights to integrate the corresponding one-hot label encodings into the entangled-label encoding. Those are then uploaded to the server to train a global classifier. During training, each entangled representation is supervised across categories via its entangled-label encoding, while random weights are re-sampled each round to introduce diversity, mitigating the global classifier’s overconfidence and promoting smoother decision boundaries. Furthermore, each client uploads a single cross-category entangled representation along with its entangled-label encoding, mitigating the risk of representation inversion attacks and reducing communication overhead. Extensive experiments demonstrate that FedRE achieves an effective trade-off among model performance, privacy protection, and communication overhead.

Conversational image segmentation grounds abstract, intent-driven concepts into pixel-accurate masks. Prior work on referring image grounding focuses on categorical and spatial queries (\eg, “left-most apple”) and overlooks functional and physical reasoning (\eg, “where can I safely store the knife?”). We address this gap and introduce Conversational Image Segmentation (CIS) and ConvSeg, a benchmark spanning entities, spatial relations, intent, affordances, functions, safety, and physical reasoning. We also present ConvSeg-Net, which fuses strong segmentation priors with language understanding, and an AI-powered data engine that generates prompt–mask pairs without human supervision. We show that current language-guided segmentation models are inadequate for CIS, while ConvSeg-Net trained on our data engine achieves significant gains on ConvSeg and maintains strong performance on existing language-guided segmentation benchmarks.

3D Referring Expression segmentation (3D-RES) is an emerging field that segments 3D objects in point cloud scenes based on given referring expressions. Although existing methods have achieved substantial progress, they primarily focus on semantic cues and often overlook spatial relations, which are essential for segmenting the referred objects in complex 3D scenes, especially those containing multiple visually similar instances. In this paper, we propose Position3D, a novel approach that explicitly incorporates spatial relation modeling into 3D-RES. Specifically, we introduce a spatial-aware query generation module that constructs point proxies by aggregating local context and incorporating spatial relations, from which the most text-relevant are selected as queries. Furthermore, we design a position-guided deformable attention in the decoder, which progressively refines attention to concentrate on the target object under positional relationship guidance. Extensive experiments on two benchmark datasets, \ie, ScanRefer, and Multi3DRefer, validate the effectiveness of the proposed method Position3D.

Referring video object segmentation (RVOS) aims to identify, track and segment the objects in a video based on language descriptions, which has received great attention in recent years. However, existing datasets remain focus on short video clips within several seconds, with salient objects visible in most frames.To advance the task towards more practical scenarios, we introduce \textbf{Long-RVOS}, a large-scale benchmark for long-term referring video object segmentation. Long-RVOS contains 2,000+ videos of an average duration exceeding 60 seconds, covering a variety of objects that undergo occlusion, disappearance-reappearance and shot changing.The objects are manually annotated with three different types of descriptions to individually evaluate the understanding of static attributes, motion patterns and spatiotemporal relationships.Moreover, unlike previous benchmarks that rely solely on the per-frame spatial evaluation, we introduce two new metrics to assess the temporal and spatiotemporal consistency.We benchmark 7 state-of-the-art methods on Long-RVOS. The results show that current approaches struggle severely with the long-video challenges.To address this, we further propose ReferMo, a promising baseline method that integrates motion information to expand the temporal receptive field, and employs a local-to-global architecture to capture both short-term dynamics and long-term dependencies.Despite simplicity, ReferMo achieves significant improvements over current methods in long-term scenarios. We hope that Long-RVOS and our baseline can drive future RVOS research towards tackling more realistic and long-form videos. Our dataset and code will be released.

Referring Video Object Segmentation (RVOS) aims to segment objects in videos based on textual queries. Current methods mainly rely on large-scale supervised fine-tuning (SFT) of Multi-modal Large Language Models (MLLMs). However, this paradigm suffers from heavy data dependence and limited scalability against the rapid evolution of MLLMs. Although recent zero-shot approaches offer a flexible alternative, their performance remains significantly behind SFT-based methods, due to the straightforward workflow designs. To address these limitations, we propose \textbf{Refer-Agent}, a collaborative multi-agent system with alternating reasoning-reflection mechanisms. This system decomposes RVOS into step-by-step reasoning process. During reasoning, we introduce a Coarse-to-Fine frame selection strategy to ensure the frame diversity and textual relevance, along with a Dynamic Focus Layout that adaptively adjusts the agent’s visual focus.Furthermore, we propose a Chain-of-Reflection mechanism, which employs a Questioner-Responder pair to generate a self-reflection chain, enabling the system to verify intermediate results and generates feedback for next-round reasoning refinement. Extensive experiments on five challenging benchmarks demonstrate that Refer-Agent significantly outperforms state-of-the-art methods, including both SFT-based models and zero-shot approaches.Moreover, Refer-Agent is flexible and enables fast integration of new MLLMs without any additional fine-tuning costs. Code will be released.

Dense Video Object Captioning (DVOC) is the task of jointly detecting, tracking, and captioning object trajectories in a video, requiring the ability to understand spatio-temporal details and describe them in natural language.Due to the complexity of the task and the high cost associated with manual annotation, previous approaches resort to training strategies with limited data, potentially leading to suboptimal performance.To circumvent this issue, we propose to generate captions about spatio-temporally localized entities leveraging a state-of-the-art VLM, and extend the LVIS and LV-VIS datasets with our synthetic captions (LVISCap and LV-VISCap). Moreover, we introduce an end-to-end model, CaptionFormer, capable of jointly detecting, segmenting, tracking and captioning object trajectories.With pretraining on LVISCap and LV-VISCap, CaptionFormer achieves state-of-the-art DVOC results on three existing benchmarks, VidSTG, VLN and BenSMOT.

Large Vision-Language Models (LVLMs) have advanced multimodal learning but face high computational cost issues due to the input of large number of visual tokens, motivating token pruning to improve inference efficiency.The key challenge lies in identifying which tokens are truly important.Most existing approaches rely on attention- or similarity-based criteria to estimate token importance.However, they inherently suffer from certain limitations, such as being task-agnostic and exhibiting positional bias.In this work, we explore a new perspective on token importance assignment based on token transitions in LVLMs, where token transitions are defined as the changes in token representations occurring as they propagate through the model’s modules.We observe that the transition of token representations provides a meaningful signal of semantic information.Based on this insight, we propose TransPrune, a training-free and efficient token pruning method.Specifically, TransPrune progressively prunes tokens by assessing their importance through a combination of Token Transition Variation (TTV), which measures changes in both the magnitude and direction of token representations; as well as Instruction-Guided Attention (IGA), which measures how strongly the instruction attends to visual tokens via attention.Extensive experiments on various LVLM architectures, such as LLaVA-v1.5, LLaVA-Next and Qwen2.5-VL, demonstrate that TransPrune maintains comparable multimodal performance while reducing inference TFLOPs by more than half.

Vision–language–action (VLA) models unify perception, language, and control for embodied agents but face significant challenges in practical deployment due to rapidly increasing compute and memory demands, especially as models scale to longer horizons and larger backbones. To address these bottlenecks, we introduce QuantVLA, a training-free post-training quantization (PTQ) framework that, to our knowledge, is the first PTQ approach for VLA systems and the first to successfully quantize a diffusion transformer (DiT) action head. QuantVLA incorporates three scale-calibrated components: (1) a selective quantization layout that integerizes all linear layers in both the language backbone and the DiT while keeping attention projections in floating point to preserve the original operator schedule; (2) attention temperature matching, a lightweight per-head scaling mechanism that stabilizes attention logits and is folded into the dequantization scales at inference; and (3) output head balancing, a per-layer residual interface calibration that mitigates post-projection energy drift. The framework requires no additional training, uses only a small unlabeled calibration buffer, and supports integer kernels for low-bit weights and activations while leaving the architecture unchanged. Across representative VLA models, QuantVLA surpasses full-precision baselines, substantially reduces memory usage within quantized modules, and lowers inference latency, providing a practical pathway toward scalable, low-bit embodied intelligence under strict compute, memory, and power constraints.

Multimodal large language models suffer from substantial inference overhead since multimodal KV Cache grows proportionally with the visual input length. Existing multimodal KV cache compression methods mostly rely on attention score to reduce cache size, which makes them are incompatible with established efficient attention kernels (e.g., FlashAttention) and ignores the contribution of value vectors to the attention output. In this work, we revisit multimodal KV cache compression from the perspective of the KV matrices’ distribution. First, we observe that frequency-domain energy of multimodal KV matrices is predominantly concentrated in low-frequency and extract this principal energy via a low-pass filter. Further, we find that removing KV pairs that deviate substantially from this principal energy leads to a pronounced performance drop, which we define as Outlier KVs. Considering Outlier KVs are more likely to encode features critical for inference, we propose FlashCache, a frequency-domain–guided, Outlier-KV-aware KV Cache compression framework. First, we introduce an Outlier KV Recognition Module that models the principal component of multimodal KV matrices in the frequency domain and preferentially retains KV pairs that significantly deviate from it. Furthermore, Dynamic Budget Allocation Module is designed to adaptively determine the per-layer KV cache size to retain more Outlier KVs. Experiments on multiple MLLMs and benchmarks demonstrate that FlashCache outperforms state-of-the-art multimoal KV compression methods, achieving up to 1.69× faster decoding with 80\% lower KV memory usage while maintaining task performance.

Vision-Language Models (VLMs) have advanced rapidly within the unified Transformer architecture, yet their deployment on resource-constrained devices remains challenging due to high computational complexity. While pruning has emerged as an effective technique for compressing VLMs, existing approaches predominantly focus on a single mode by pruning either parameters or tokens, neglecting fully exploring the inherent redundancy in each mode, which leads to substantial performance degradation at high pruning ratios. To address the above limitations, we propose Collaborative Multi-Mode Pruning (CoMP), a novel framework tailored for VLMs by performing joint parameter and token pruning.Specifically, we first design a Collaborative Importance Metric (CIM) that investigates the mutual interference between the coupled parameters and tokens. It incorporates distinct significance of tokens into the computation of parameter importance scores, while simultaneously mitigating the affect of pruned parameters on token importance scores.Moreover, we develop a Multi-Mode Pruning Strategy (MPS) that decomposes the overall pruning process into a sequence of pruning stages, while in each stage we estimate the priory of different pruning modes based on their pruning cost and adaptively shift to the optimal one. Additionally, MPS integrates the historical cost and random exploration, in order to achieve a stable pruning process and avoid local optimum.Extensive experiments across various vision-language tasks and models demonstrate that our method effectively promotes the performance under high pruning ratios by comparing to the state-of-the-art approaches.The source code will be released upon acceptance.

Large Vision–Language Models (VLMs) enable strong multimodal reasoning but incur heavy inference costs from redundant visual tokens. Token pruning alleviates this issue, yet existing approaches face limitations. Attention-based methods rely on raw attention scores, which are often unstable across layers and heads and can lead to redundant selections. Diversity-based methods improve robustness by selecting tokens far apart in feature space, but risk dropping regions needed for accurate prediction. We propose ZOO-Prune, a training-free framework built on the intuition that highly sensitive tokens have a stronger influence on the model's output and capture complementary visual cues rather than redundant ones. To achieve this, we estimate token sensitivity using zeroth-order perturbations at the lightweight projection layer. This measures how small random perturbations affect the projected features and enables efficient approximation of each token’s influence without backpropagation. Extensive experiments across multiple VLMs and benchmarks show that ZOO-Prune consistently outperforms prior methods while pruning up to 94.4% of tokens without sacrificing accuracy. Our method also improves efficiency, reaching up to 2.30x faster end-to-end inference compared to the baseline.

In multimodal large language models (MLLMs), the surge of visual tokens significantly increases the inference time and computational overhead, making them impractical for real-time or resource-constrained applications.Visual token pruning is a promising strategy for reducing the cost of MLLM inference by removing redundant visual tokens.Existing researches usually assume that all attention heads contribute equally to the visual interpretation.However, our study reveals that different heads may capture distinct visual semantics and inherently play distinct roles in visual processing.In light of this observation, we propose HAWK, a \textbf{h}ead importance-\textbf{aw}are visual to\textbf{k}en pruning method that perceives the varying importance of attention heads in visual tasks to maximize the retention of crucial tokens.By leveraging head importance weights and text-guided attention to assess visual token significance, HAWK effectively retains task-relevant visual tokens while removing redundant ones.The proposed HAWK is entirely training-free and can be seamlessly applied to various MLLMs.Extensive experiments on multiple mainstream vision-language benchmarks demonstrate that HAWK achieves state-of-the-art accuracy. When applied to Qwen2.5-VL, HAWK retains 96.0\% of the original accuracy after pruning 80.2\% of the visual tokens. Additionally, it reduces end-to-end latency to 74.4\% of the original and further decreases GPU memory usage across the tested models.

Large Vision-Language Models (LVLMs) usually suffer from prohibitive computational and memory costs due to the quadratic growth of visual tokens with image resolution. Existing token compression methods, while varied, often lack a high-level semantic understanding, leading to suboptimal merges, information redundancy, or context loss. To address these limitations, we introduce CORE (Compact Object-centric REpresentations), a new paradigm for visual token compression. CORE leverages an efficient segmentation decoder to generate object masks, which serve as a high-level semantic prior to guide the merging of visual tokens into a compact set of object-centric representations. Furthermore, a novel centroid-guided sorting mechanism restores a coherent spatial order to the merged tokens, preserving vital positional information. Extensive experiments show that CORE not only establishes a new state-of-the-art on six authoritative multimodal benchmarks for fixed-rate compression, but also achieves dramatic efficiency gains in adaptive-rate settings. Even under extreme compression, after aggressively retaining with only 2.2% of all visual tokens, CORE still maintains 97.4% of baseline performance. Our work demonstrates the superiority of object-centric representations for efficient and effective LVLM processing.

In real-world applications, multi-view data often suffer from missing situations due to privacy protection and sensor failure factors. The Incomplete scenarios not only lead to partial information availability but also cause significant imbalance learning among views: certain “strong views” dominate the fusion process, while “weak views” contribute marginally, thereby undermining cross-view collaboration. Existing incomplete multi-view clustering methods mainly focus on "how to handle missing data", yet they largely overlook the imbalance view contribution induced by incompleteness and its profound impact on representation learning and clustering performance. To address these issues, our paper first analyzes the data imbalance caused by missing views and the resulting disparities in view learning quality. Then, we propose a collaborative evaluation and enhancement framework (\textbf{ICER}) for imbalanced incomplete multi-view clustering . Specifically, we employ shapley values to quantify the marginal contribution of each view, and incorporate imbalanced optimal transport to characterize distributional deviations across views. On this basis, we construct the view contribution imbalance metric to comprehensively evaluate cross-view collaboration and fusion quality, and design a collaboration enhancement module to explicitly reinforce inter-view cooperative optimization and feature fusion. Extensive experiments on multiple benchmark datasets demonstrate that the proposed method outperforms existing incomplete multi-view clustering approaches, validating the effectiveness and necessity of explicitly modeling and mitigating view imbalance in imbalanced incomplete scenarios.

Multi-view contrastive clustering has emerged as a powerful paradigm for learning comprehensive representations from heterogeneous data sources. However, prevailing approaches typically overlook the intrinsic geometric and clustering structures, rendering them structure-agnostic. In this paper, we propose a novel framework that performs Multi-Hierarchical Contrastive Spectral Fusion (MCSF) to address these limitations. MCSF integrates deep spectral embedding into the encoder to preserve local manifold structure, guiding the learned representations to be clustering-friendly. To enhance cross-view consistency, MCSF introduces a multi-hierarchical contrastive loss jointly optimizing (1) view-specific structure preservation, (2) view-consensus alignment, and (3) consensus structure refinement. This mechanism enables the construction of an accurate and semantically consistent consensus representation, effectively fusing multi-view information and uncovering authentic cluster structures. Extensive experiments on benchmarks validate the effectiveness of multi-hierarchical contrastive spectral fusion in clustering accuracy and representation quality.

In open-world semi-supervised learning (OWSSL), a model learns from labeled data and unlabeled data containing both known and novel classes. In practical OWSSL applications, models are expected to perform rigorous classification by directly selecting the most semantically relevant label from a candidate set for each sample. Existing OWSSL methods fail to achieve this because novel samples are trained without explicit supervision, and these methods lack mechanisms to extract latent semantic information, resulting in predicted labels that have no semantic correspondence to candidate textual labels. To address this, we introduce SEmantic Capture for Open-world Semi-supervised learning (SECOS), which directly predicts textual labels from the candidate set without post-processing, meeting the requirements of practical OWSSL applications. SECOS leverages external knowledge to extract and align semantic representations across modalities for both known and novel classes, providing explicit supervisory signals for training novel classes. Extensive experiments demonstrate that even when existing OWSSL methods are evaluated under the more lenient post-hoc matching setting, SECOS still surpasses them by up to 5.4\% without such assistance, highlighting its superior effectiveness. Code is available in the supplementary materials.

Generalized Category Discovery (GCD) aims to identify both known and unknown categories, with only partial labels given for the known categories, posing a challenging open-set recognition problem. State-of-the-art approaches for GCD task are usually built on multi-modality representation learning, which is heavily depend upon inter-modality alignment. However, few of them cast a proper intra-modality alignment to generate a desired underlying structure of representation distributions. In this paper, we propose a novel and effective multi-modal representation learning framework for GCD via Semi-Supervised Rate Reduction, called SSR$^2$-GCD, to learn cross-modality representations with desired structural properties based on emphasizing to properly align intra-modality relationships. Moreover, to boost knowledge transfer, we integrate prompt candidates by leveraging the inter-modal alignment offered by Vision Language Models. We conduct extensive experiments on generic and fine-grained benchmark datasets demonstrating superior performance of our approach.

Learning temporal transformations, that is, how visual objects evolve across frames, is a fundamental challenge in video representation learning. Frame-to-frame dynamics involve complex, non-linear, and non-local changes that go far beyond conventional spatial augmentations. We propose TimeBridge, a self-supervised method that combines the joint embedding for video representation with learning temporal transformations by reconstructing in-between frames from only the start and end frames. This formulation encourages the model to infer the temporal evolution bridging the two endpoints, rather than merely encoding static frame representations. Unlike joint-embedding methods that lack explicit transformation modelling or future-prediction objectives that rely on unconstrained extrapolation, TimeBridge learns concrete frame-to-frame dynamics by promoting temporal consistency. We realise this through cross-concatenated class tokens and lightweight decoders, which recombine features from the start and end frames to reconstruct intermediates. TimeBridge achieves new state-of-the-art performance on multiple dense video prediction benchmarks, including 73.5 J&amp;F on DAVIS 2017 video object segmentation, 47.5 mIoU on VIP part propagation.

Partial label learning is a prominent weakly supervised classification task, where each training instance is ambiguously labeled with a set of candidate labels. In real-world scenarios, candidate labels are often influenced by instance features, leading to the emergence of instance-dependent PLL (ID-PLL), a setting that more accurately reflects this relationship. A significant challenge in ID-PLL is instance entanglement, where instances from similar classes share overlapping features and candidate labels, resulting in increased class confusion. To address this issue, we propose a novel Class-specific Augmentation based Disentanglement (CAD) framework, which tackles instance entanglement by both intra- and inter-class regulations. For intra-class regulation, CAD amplifies class-specific features to generate class-wise augmentations and aligns same-class augmentations across instances. For inter-class regulation, CAD introduces a weighted penalty loss function that applies stronger penalties to more ambiguous labels, encouraging larger inter-class distances. By jointly applying intra- and inter-class regulations, CAD improves the clarity of class boundaries and reduces class confusion caused by entanglement. Extensive theoretical and experimental results demonstrate the effectiveness of CAD in mitigating the entanglement problem and enhancing ID-PLL performance.

We show that introducing a weighting factor to reduce the influence of identity shortcuts in residual networks significantly enhances semantic feature learning in generative representation learning frameworks, such as masked autoencoders (MAEs) and diffusion models. Our modification improves linear probing accuracy for both, notably increasing ImageNet accuracy from 67.8% to 72.7% for MAEs with a VIT-B/16 backbone, while also boosting generation quality for diffusion models. This significant gap suggests that, while residual connection structure serves an essential role in facilitating gradient propagation, it may have a harmful side effect of reducing capacity for abstract learning by virtue of injecting an echo of shallower representations into deeper layers. We ameliorate this downside via a fixed formula for monotonically decreasing the contribution of identity connections as layer depth increases. Our design promotes the gradual development of feature abstractions, without impacting network trainability. Analyzing the representations learned by our modified residual networks, we find correlation between low effective feature rank and downstream task performance.

The neural process (NP) is a probabilistic meta learning model that learns distributions over functions via a global latent variable.It enables fast adaptation in few-shot scenarios by leveraging past experience. However, the design of latent variable structures and conditioning mechanisms in NPs remains underexplored, despite their importance in capturing diverse functional distributions.This paper proposes a new variant of NPs via mixture density modeling, referred to as the neural mixture density process (NMDP).The NMDP decomposes model parameters into task-agnostic and task-specific components to represent function distributions more flexibly. We train the model via the Expectation–Maximization algorithm to construct expressive functional priors.Compared with existing work, our method maintains several advantages: (i) less overfitting by updating a small part of the network parameters, (ii) compact task representation via distributions in the simplex,(iii) an improvement guarantee of generative likelihoods over iteration. Experimental results show that our method can achieve competitive performance with adequate explainability.

Ensemble clustering aims to derive a consensus partition from multiple base clustering results. Anchor-based methods construct compact similarity representations via anchors, substantially improving computational efficiency. However, when outliers contaminate the data, reconstructing the base clustering results often yields biased anchors. These biased anchors degrade the quality of the anchor similarity matrix and lead to a decline in clustering accuracy. To address this issue, we propose a novel method called large-scale robust enhanced ensemble clustering via outlier decoupling (RANGE). Specifically, RANGE first converts the base clustering results into an initial bipartite graph. To enhance the reliability of this bipartite graph, RANGE designs a high-order fuzzy enhancement strategy (HFES) specifically for initial bipartite graphs. Next, a mapping matrix further filters redundant information from the enhanced bipartite graph. RANGE then reconstructs the mapped bipartite graph via matrix factorization. An anchor matrix is introduced to further enhance computational efficiency. To improve robustness, RANGE incorporates a decoupling term that separates the clean clustering structure and the outlier-contaminated structure in the anchor space. With this decoupling mechanism, RANGE is capable of performing robust ensemble clustering. Moreover, by applying outlier detectors to the decoupled outlier structure, RANGE can be extended to the outlier-detection task. Consequently, RANGE forms a cross-task general framework, and both tasks retain linear time complexity. Extensive cross-domain experiments indicate that RANGE delivers superior performance in both clustering validity and outlier detection. The code is available in the supplementary material.

World models have become crucial for autonomous driving, as they learn how scenarios evolve over time to address the long-tail challenges of the real world. However, current approaches relegate world models to limited roles: they operate within ostensibly unified architectures that still keep world prediction and motion planning as decoupled processes. To bridge this gap, we propose DriveLaW, a novel paradigm that unifies video generation and motion planning. By directly injecting the latent representation from its video generator into the planner, DriveLaW ensures inherent consistency between high-fidelity future generation and reliable trajectory planning. Specifically, DriveLaW consists of two core components: DriveLaW-Video, our powerful world model that generates high-fidelity forecasting with expressive latent representations, and DriveLaW-Act, a diffusion planner that generates consistent and reliable trajectories from the latent of DriveLaW-Video, with both components optimized by a three-stage progressive training strategy. The power of our unified paradigm is demonstrated by new state-of-the-art results across both tasks. DriveLaW not only advances video prediction significantly, surpassing best-performing work by 33.3% in FID and 1.8% in FVD, but also achieves a new record on the NAVSIM planning benchmark.

Vision-based autonomous driving has gained much attention due to its low costs and excellent performance.Compared with dense BEV (Bird’s Eye View) or sparse query models, Gaussian-centric method is a comprehensive yet sparse representation by describing scene with 3D semantic Gaussians. In this paper, we introduce DLWM, a novel paradigm with Dual Latent World Models specifically designed to enable holistic gaussian-centric pre-training in autonomous driving using two stages. In the first stage, DLWM predicts 3D Gaussians from queries by self-supervised reconstructing multi-view semantic and depth images. Equipped with fine-grained contextual features, in the second stage, two latent world models are trained separately for temporal feature learning, including Gaussian-flow-guided latent prediction for downstream occupancy perception and forecasting tasks, and ego-planning-guided latent prediction for motion planning. Extensive experiments in SurroundOcc and nuScenes benchmarks demonstrate that DLWM shows significant performance gains across Gaussian-centric 3D occupancy perception, 4D occupancy forecasting and motion planning tasks.

Recent Vision-Language-Action (VLA) models for autonomous driving explore inference-time reasoning as a way to improve driving performance and safety in challenging scenarios. Most prior work uses natural language to express chain-of-thought (CoT) reasoning before producing driving actions. However, text may not be the most efficient representation for reasoning. In this work, we present Latent-CoT-Drive (LDrive): a model that expresses CoT in a latent language that captures possible outcomes of the driving actions being considered. Our approach unifies CoT reasoning and decision making by representing both in an action-aligned latent space. Instead of natural language, the model reasons by interleaving (1) action-proposal tokens, which use the same vocabulary as the model’s output actions; and (2) world model tokens, which are grounded in a learned latent world model and express future outcomes of these actions. We cold start latent CoT by supervising the model’s action proposals and world model tokens based on ground-truth future rollouts of the scene. We then post-train with closed-loop reinforcement learning to strengthen reasoning capabilities. On a large-scale end-to-end driving benchmark, LDrive achieves faster inference, better trajectory quality, and larger improvements from interactive reinforcement learning compared to both non-reasoning and text-reasoning baselines.

Supervised open-loop training has been widely adopted for training traffic simulation models; however, it fails to capture the inherently dynamic, multi-agent interactions common in complex driving scenarios. We introduce RLFTSim, a reinforcement-learning-based fine-tuning framework that enhances scenario realism by aligning simulator rollouts with real-world data distributions and provides a method for distilling goal-conditioned controllability in scenario generation. We instantiate RLFTSim on top of a pre-trained simulation model, design a reward that balances fidelity and controllability, and perform comprehensive experiments on the Waymo Open Motion Dataset. Our results show improvements in realism enhancement, achieving state-of-the-art performance. Compared with other heuristic search-based fine-tuning methods, RLFTSim requires significantly fewer samples due to a proposed low-variance and dense reward signal, and it directly addresses the realism alignment issue by design. We also demonstrate the effectiveness of our approach for distilling traffic simulation controllability through goal conditioning.

Recent research has investigated the use of large language models (LLMs) to generate traffic scenarios for autonomous driving. However, pretrained LLMs often fail to align with real-world traffic distributions. In this work, we present TrafficAlign, an automated framework that synthesizes traffic scenarios based on real-world driving videos, performs data validation, and aligns LLMs with the synthesized scenarios. The evaluation shows that traffic scenarios generated by TrafficAlign are highly effective, revealing up to 10.8% more collisions on average across three autonomous driving models than state-of-the-art methods. Furthermore, fine-tuning these driving models with TrafficAlign-generated scenarios significantly reduced collision rates by 36.1% compared with the original models. A qualitative study using traffic datasets from six geographically diverse regions shows that TrafficAlign-generated scenarios exhibit strong alignment with corresponding traffic distributions in these regions.

Deep learning-based online mapping has emerged as a cornerstone of autonomous driving, yet these models frequently fail to generalize beyond familiar environments. We propose a framework to identify and measure the underlying failure modes by disentangling two effects: Memorization of input features and overfitting to known map topologies. We propose metrics based on evaluation subsets that control for geographical proximity and topological similarity between training and validation scenes. We introduce Fréchet distance–based reconstruction statistics that capture per-element shape fidelity without threshold tuning, and define complementary failure-mode scores: an input-feature overfitting score quantifying the performance drop when geographic cues disappear, and a topology overfitting score measuring degradation as scenes become topologically novel. Beyond models, we analyze dataset biases and contribute topology-aware diagnostics: A minimum-spanning-tree (MST) diversity metric for training sets and a symmetric coverage metric to quantify topological similarity between splits. Leveraging these, we formulate an MST-based sparsification strategy that reduces redundancy and improves balancing and performance while shrinking training size. Experiments on nuScenes and Argoverse 2 across multiple state-of-the-art models yield more trustworthy assessment of generalization and show that topology-diverse and balanced training sets lead to improved performance. Our results motivate failure-mode-aware protocols and topology-centric dataset design for deployable online mapping.

Collaborative perception leverages data exchange among multiple agents to enhance overall perception capabilities. However, heterogeneity across agents introduces domain gaps that hinder collaboration, and this is further exacerbated by an underexplored issue: modality isolation. It arises when multiple agents with different modalities never co-occur in any training data frame, enlarging cross-modal domain gaps. Existing alignment methods rely on supervision from spatially overlapping observations, thus fail to handle modality isolation. To address this challenge, we propose CodeAlign, the first efficient, co-occurrence-free alignment framework that smoothly aligns modalities via cross-modal feature-code-feature(FCF) translation. The key idea is to explicitly identify the representation consistency through codebook, and directly learn mappings between modality-specific feature spaces, thereby eliminating the need for spatial correspondence. Codebooks regularize feature spaces into code spaces, providing compact yet expressive representations. With a prepared code space for each modality, CodeAlign learns FCF translations that map features to the corresponding codes of other modalities, which are then decoded back into features in the target code space, enabling effective alignment. Experiments show that, when integrating three modalities, CodeAlign requires only 8% of the training parameters of prior alignment methods, reduces communication load by 1024x, and achieves state-of-the-art perception performance on both OPV2V and DAIR-V2X dataset. Code will be released.

Simulation-generated datasets for autonomous driving rely on omniscient data collection 'expert' policies, which use unobservable scene information (e.g., from occluded regions) to make driving decisions.When such data is used for end-to-end policy training, it results in an information asymmetry between the expert and the 'learner' policy, which has limited sensor coverage and navigational intent information compared to the expert. We show that this asymmetry leads to a significant drop in the performance of the learner. To combat this, we present LEAD, a new high-quality synthetic dataset collected in the CARLA simulator with three key improvements.(1) The expert minimizes its use of unobservable information by removing entities from its input state that would be occluded in the learner's field of view.By providing the learner with (2) detailed driver intent information and (3) rich sensor modalities (cameras, LiDARs, radars, and odometry), the dataset narrows down the information gap between the learner and expert. We then propose TransFuser v6 (TFv6), a simple end-to-end learner policy trained on LEAD.As a result of our improvements, TFv6 substantially advances the state of the art on all publicly available CARLA closed-loop driving benchmarks, reaching driving scores of 95 on Bench2Drive, 62 on Longest6 v2, and 15 on the Town13 validation routes.Finally, we aggregate the LEAD dataset with several public real-world datasets under a unified repository to enable cross-dataset evaluation.We show that pre-training TFv6 on synthetic data from LEAD leads to consistent performance gains when followed by fine-tuning with real data from the NAVSIM v1, NAVSIM v2, and WOD-E2E benchmarks.

Predicting driver attention is a critical problem for developing explainable autonomous driving systems and understanding driver behavior in mixed human-autonomous vehicle traffic scenarios. Although significant progress has been made through large-scale driver attention datasets and deep learning architectures, existing works are constrained by narrow frontal field-of-view and limited driving diversity. Consequently, they fail to capture the full spatial context of driving environments, especially during lane changes, turns, and interactions involving peripheral objects such as pedestrians or cyclists. In this paper, we introduce DriverGaze360, a large-scale 360$^\circ$ field of view driver attention dataset, containing $\sim$1 million gaze-labeled frames collected from 19 human drivers, enabling comprehensive omnidirectional modeling of driver gaze behavior. Moreover, our panoramic attention prediction approach, DriverGaze360-Net, jointly learns attention maps and attended objects by employing an auxiliary semantic segmentation head. This improves spatial awareness and attention prediction across wide panoramic inputs. Extensive experiments demonstrate that DriverGaze360-Net achieves state-of-the-art attention prediction performance on multiple metrics on panoramic driving images. Dataset and method will be made publicly available.

Learning-based motion planners, despite recent progress, often suffer from temporal inconsistency. Small perturbations across frames can accumulate into unstable trajectories, degrading comfort and safety in closed-loop driving. Several methods attempt to inject history as a static conditioning signal to stabilize outputs, only to induce the planner to copy historical patterns instead of adapting to environment contexts. To address this limitation, we propose Diffusion Forcing Planner (DFP), a diffusion-based planning framework driven by history-guided control. Specifically, DFP decomposes the full trajectory into history, current and future segments, and assign independent noise levels to each segment. The model jointly denoises the historical and the future segments, enforcing a heterogeneous joint diffusion process. At inference, classifier-free guidance (CFG) is applied to steer future sampling using annealed history in a controllable manner. Closed-loop evaluation and comprehensive ablations on nuPlan show that DFP achieves competitive performance while producing continuous, stable, and controllable motion plans in complex driving scenarios.

Moving instance segmentation (MIS) attracts increasing attention due to its broad applications in traffic surveillance, autonomous driving, and animal tracking. Event cameras record asynchronous brightness changes, providing high temporal resolution and dynamic range, which makes them highly sensitive to motion information. By fusing event and image features, motion cues from events can complement spatial details from images, enhancing the performance of MIS. However, current multimodal MIS methods still struggle to segment small moving instances, as event cameras often yield sparse features under limited resolution. Moreover, event features entangle appearance attributes with motion cues, which further restricts effective cross-modal fusion. To address these challenges, we first propose a dual-disentangling feature extraction framework that separates and extracts appearance and motion information within both image and event modalities, thereby improving feature density. Subsequently, a multi-granularity cross-modal alignment is introduced to align distributionally and semantically consistent features across modalities, enabling more effective fusion with rich spatial and temporal details. The experiment results demonstrate that our method achieves state-of-the-art performance in multimodal MIS, especially for small instances under challenging conditions such as fast motion and low-light settings.

We introduce EVObject for unsupervised 3D instance segmentation that bridges the geometric domain gap between synthetic pretraining data and real-world point clouds. Current methods suffer from structural discrepancies when transferring object priors from synthetic datasets (e.g., ShapeNet) to real scans (e.g., ScanNet), particularly due to morphological variations and occlusion artifacts. To address this, EVObject integrates two innovative modules: (1) An object discerning module that dynamically refines object candidates, enabling continuous adaptation of object priors to target domains; and (2) An object completion module that reconstructs partial geometries before discovering object. We conduct extensive experiments on two real-world datasets and one synthetic dataset, demonstrating superior 3D object segmentation performance over all baselines while achieving state-of-the-art results.

Interactive video segmentation often requires many user interventions for robust performance in challenging scenarios (e.g., occlusions, object separations, camouflage, etc.).Yet, even state-of-the-art models like SAM2 use corrections only for immediate fixes without *learning* from this feedback, leading to inefficient, repetitive user effort. To address this, we introduce Live Interactive Training (LIT), a novel framework for prompt-based visual systems where models also learn online from human corrections at inference time. Our primary instantiation, LIT-LoRA, implements this by continually updating a lightweight LoRA module on-the-fly. When a user provides a correction, this module is rapidly trained on that feedback, allowing the vision system to improve performance on subsequent frames of the same video.Leveraging the core principles of LIT, our LIT-LoRA implementation achieves an average 18-34\% reduction in total corrections on challenging video segmentation benchmarks, with a negligible training overhead of $\sim$0.5s per correction. We further demonstrate its generality by successfully adapting it to other segmentation models and extending it to CLIP-based fine-grained image classification. Our work highlights the promise of live adaptation to transform interactive tools and significantly reduce redundant human effort in complex visual tasks.

The performance of promptable video object segmentation (PVOS) models substantially degrades under input corruptions, which prevents PVOS deployment in safety-critical domains. This paper offers the first comprehensive study on robust PVOS (RobustPVOS). We first construct a new, comprehensive benchmark with two real-world evaluation datasets of 351 video clips and more than 2,500 object masks under real-world adverse conditions. At the same time, we generate synthetic training data by applying diverse and temporally varying corruptions to existing VOS datasets. Moreover, we present a new RobustPVOS method, dubbed Memory-object-conditioned Gated-rank Adaptation (MoGA). The key to successfully performing RobustPVOS is two-fold: effectively handling object-specific degradation and ensuring temporal consistency in predictions. MoGA leverages object-specific representations maintained in memory across frames to condition the robustification process, which allows the model to handle each tracked object differently in a temporally consistent way. Extensive experiments on our benchmark validate MoGA's efficacy, showing consistent and significant improvements across diverse corruption types on both synthetic and real-world datasets, establishing a strong baseline for future RobustPVOS research. Our benchmark will be made publicly available.

Segmenting long-form videos into semantically coherent scenes is a fundamental task in large-scale video understanding. Existing encoder-based methods are limited by visual-centric biases, classify each shot in isolation without leveraging sequential dependencies, and lack both narrative understanding and explainability. In this paper, we present Scene-VLM, the first fine-tuned vision-language model (VLM) framework for video scene segmentation. Scene-VLM jointly processes visual and textual cues including frames, transcriptions, and optional metadata to enable multimodal reasoning across consecutive shots. The model generates predictions sequentially with causal dependencies among shots and introduces a context–focus window mechanism to ensure sufficient temporal context for each shot-level decision. In addition, we propose a scheme to extract confidence scores from the token-level logits of the VLM, enabling controllable precision–recall trade-offs that were previously limited to encoder-based methods. Furthermore, we demonstrate that our model can be aligned to generate coherent natural-language rationales for its boundary decisions through minimal targeted supervision. Our approach achieves state-of-the-art performance on standard scene segmentation benchmarks. On MovieNet, for example, Scene-VLM yields significant improvements of +6 AP and +13.7 F1 over the previous leading method.

This paper addresses the challenge of data scarcity in semantic segmentation by generating datasets through text-to-image (T2I) generation models, reducing image acquisition and labeling costs. Segmentation dataset generation faces two key challenges: 1) aligning generated samples with the target domain and 2) producing informative samples beyond the training data. Fine-tuning T2I models can help generate samples aligned with the target domain. However, it often overfits and memorizes training data, limiting their ability to generate diverse and well-aligned samples. To overcome these issues, we propose Concept-Aware LoRA (CA-LoRA), a novel fine-tuning approach that selectively identifies and updates only the weights associated with necessary concepts (e.g., style or viewpoint) for domain alignment while preserving the pretrained knowledge of the T2I model to produce informative samples. We demonstrate its effectiveness in generating datasets for urban-scene segmentation, outperforming baseline and state-of-the-art methods in in-domain (few-shot and fully-supervised) settings, as well as in domain generalization tasks, especially under challenging conditions such as adverse weather and varying illumination, further highlighting its superiority.

Bird’s Eye View (BEV) semantic segmentation is essential for autonomous driving and mobile robotics, yet it still faces significant challenges on accurate segmentation of foreground object and efficient estimating of layout categories obscured by objects. To address these issues, we propose BEV-CAR, a Context-Aware Rasterization method that rasterizes the BEV representation without any coordinate transformations. By optimising each ray and incorporating depth features, BEV-CAR effectively addresses the challenges posed by object occlusions and varying environmental conditions. It ensures robust performance across diverse scenarios, particularly improving the accuracy of foreground object segmentation and layout estimation in occluded areas. And extensive experiments on the nuScenes and Argoverse datasets demonstrate that BEV-CAR achieves state-of-the-art (SOTA) performance. More importantly, the rasterization technique in this paper does not introduce additional computational overhead during the inference process, making it suitable for practical deployment in real-world scenarios. Code and technical appendix are available in supplementary material.

Accurate segmentation of marine organisms is vital for biodiversity monitoring and ecological assessment, yet existing datasets and models remain largely limited to terrestrial scenes. To bridge this gap, we introduce AquaOV255, the first large-scale and fine-grained underwater segmentation dataset containing 255 categories and over 20K images, covering diverse marine organisms and man-made objects for open-vocabulary evaluation. Furthermore, we establish the first underwater open-vocabulary segmentation benchmark, UOVSBench, by integrating AquaOV255 with five additional underwater datasets to enable comprehensive cross-domain evaluation. Alongside, we present Earth2Ocean, a training-free open-vocabulary segmentation framework that transfers terrestrial vision–language models (VLMs) to underwater domains without any additional underwater training. Earth2Ocean consists of two core components: a Geometric-guided Visual Mask Generator (GMG) that refines visual features via self-similarity geometric priors for local structure perception, and a Category-visual Semantic Alignment(CSA) module that enhances text embeddings through multimodal large language model reasoning and scene-aware template construction. Extensive experiments on the UOVSBench benchmark demonstrate that Earth2Ocean achieves over 6+ mIoU improvement on average while maintaining efficient inference.

Pre-trained weights have become a cornerstone of modern deep learning, enabling efficient knowledge transfer and improving downstream task performance, especially in data-scarce scenarios. However, a fundamental question remains: how can we obtain better pre-trained weights that encapsulate more knowledge beyond the given dataset? In this work, we introduce KNowledge-Overflowed Weights (KNOW) prediction, a novel strategy that leverages structured forgetting and its inversion to synthesize knowledge-enriched weights. Our key insight is that sequential fine-tuning on progressively downsized datasets induces a structured forgetting process, which can be modeled and reversed to recover knowledge as if trained on a larger dataset. We construct a dataset of weight transitions governed by this controlled forgetting and employ meta-learning to model weight prediction effectively. Specifically, our KNowledge-Overflowed Weights Nowcaster (KNOWN) acts as a hyper-model that learns the general evolution of weights and predicts enhanced weights with improved generalization. Extensive experiments across diverse datasets and architectures demonstrate that KNOW prediction consistently outperforms Na\"ive fine-tuning and simple weight prediction, leading to superior downstream performance. Our work provides a new perspective on reinterpreting forgetting dynamics to push the limits of knowledge transfer.

Cloud-Edge Continual Test-Time Adaptation (CTTA)—with edge devices processing real-time data and the cloud offering strong computing power—is a critical paradigm for models that adapt to dynamic data distributions in real-world scenarios. However, most existing frameworks assume architectural homogeneity between cloud and edge CNNs, which poses a significant performance bottleneck, particularly given the rapid emergence of Transformer-based models. Current methods fail to bridge this architectural gap, resulting in significant deficiencies in adaptation accuracy and practical applicability. To address this, we propose a novel Cross-Architecture Adaptation (CAA) framework for heterogeneous Cloud-Edge CTTA that enables effective adaptation to shifting data distributions. Specifically, CAA deploys a large Transformer-based teacher model on the cloud for robust feature extraction and prediction, and a lightweight CNN-based student model on edge devices to fit resource constraints. Based on such cloud-edge models, a synergistic edge-to-cloud communication strategy, Multi-criteria Dynamic Cross-domain Sampling, ensures only the most informative, class-balanced samples are uploaded, minimizing communication costs while guaranteeing stable, unbiased adaptation. Moreover, a Multi-level Adaptive Heterogeneous Distillation module is proposed to facilitate effective knowledge transfer across the architecturally disparate models, and improve the learning efficiency of the edge one. Experiments on several benchmarks demonstrate that CAA achieves state-of-the-art performance with low edge resource consumption and minimal edge-to-cloud communication overhead.

Multimodal Continual Learning (MMCL) aims to enable models to continuously accumulate knowledge across multiple tasks and modalities without forgetting prior information. MMCL presents more challenges than single-modal continual learning, as it requires effective cooperation and complementarity between modalities. Existing methods often treat modality alignment as a static process, assuming once alignment is established, it remains fixed. However, we argue that modality alignment is inherently dynamic, evolving with task learning and feature propagation across layers. To address this, we introduce Dynamic Alignment Graph Regularization (DAGR), a novel approach that explicitly models the evolving alignment across layers. By incorporating multi-level graph regularization, our method stabilizes the alignment process and mitigates catastrophic forgetting. Extensive experiments on benchmarks, such as MTIL, show that DAGR outperforms static alignment-based methods and other continual learning techniques, achieving superior stability.

Fairness in Continual Learning for Large Multimodal Models (LMMs) is an emerging yet underexplored challenge, particularly in the presence of imbalanced data distributions that can lead to biased model updates and suboptimal performance across tasks. While recent continual learning studies have made progress in addressing catastrophic forgetting, the problem of fairness caused the imbalanced data remains largely underexplored. This paper presents a novel Fairness Direct Preference Optimization (FaiDPO or $\phi$-DPO) framework for continual learning in LMMs. In particular, we first propose a new continual learning paradigm based on Direct Preference Optimization (DPO) to mitigate catastrophic forgetting by aligning learning with pairwise preference signals. Then, we identify the limitations of conventional DPO in imbalanced data and present a new $\phi$-DPO loss that explicitly addresses distributional biases. We provide a comprehensive theoretical analysis demonstrating that our approach addresses both forgetting and data imbalance. Additionally, to enable $\phi$-DPO-based continual learning, we construct pairwise preference annotations for existing benchmarks in the context of continual learning. Extensive experiments and ablation studies show the proposed $\phi$-DPO achieves State-of-the-Art performance across multiple benchmarks, outperforming prior continual learning methods of LMMs.

Incremental Object Detection (IOD) enables AI systems to continuously acquire new object classes while preserving knowledge of previously learned ones, an ability essential for deployment in dynamic, real-world environments. Existing IOD methods typically rely on knowledge distillation to mitigate catastrophic forgetting. However, the tight coupling between the student model’s detection head and backbone causes distillation gradients to conflict with new-class supervision at the head, injecting head-specific bias into the backbone and ultimately weakening distillation effectiveness. To address this issue, we propose a decoupled training mechanism for the model’s backbone and classification head. Specifically, we introduce the Future-aware Cross-head Distillation (FaCHD) method, which utilizes two frozen complementary teachers (historical and intermediate teachers) to decode the student’s ROI features for cross-head distillation. This strategy implicitly alleviates prediction conflicts caused by detection-head bias and provides richer task-relevant guidance, thereby improving distillation efficiency. To further address the detection head bias and model recency problem, we propose a Prototype Semantic Drift Compensation module, which recalibrates multi-granularity prototypes of old classes, effectively correcting semantic drift and enhancing the stability of the detection head. Extensive experiments on two standard IOD benchmarks demonstrate the effectiveness and superiority of the proposed method. Code is available in the supplementary materials.

Incremental learning (IL) arises from the need to continuously update models under limited data and computational resources. Most existing IL studies focus on data‑scarce settings. They often develop complex methods that rely on heavy computation, while overlooking the computational resource constraints common in real‑world scenarios. This motivates us to formalize the problem of Computational Resource‑Aware Incremental Learning, which explicitly considers the computational budget during model training. To tackle this problem, we propose Smart Replay, an efficient memory rehearsal algorithm that adaptively allocates resources by scheduling the replay ratio across mini‑batches. We cast replay‑ratio optimization into an optimal control formulation that jointly minimizes new‑task and memory losses. We further propose a heuristic Q-function to guide ratio adjustments, adaptively balancing short-term efficiency and long-term stability. Finally, we develop a practical algorithm that periodically updates the replay ratio during training. Experiments on multiple benchmarks validate that Smart Replay consistently outperforms fixed‑replay baselines, achieving higher accuracy and lower forgetting under the same computational budget.

Prompt learning has emerged as an effective technique for fine-tuning large-scale foundation models for downstream tasks. However, conventional prompt tuning methods are prone to overfitting and can struggle with out-of-distribution generalization. To address these limitations, Bayesian prompt learning has been proposed, which frames prompt optimization as a Bayesian inference problem to enhance robustness. This paper introduces Repulsive Bayesian Prompt Learning (ReBaPL), a novel method for Bayesian prompt learning, designed to efficiently explore the complex and often multimodal posterior landscape of prompts. Our method integrates a cyclical step-size schedule with a stochastic gradient Hamiltonian Monte Carlo (SGHMC) algorithm, enabling alternating phases of exploration to discover new modes, and exploitation to refine existing modes. Furthermore, we introduce a repulsive force derived from a potential function over probability metrics (including Maximum Mean Discrepancy and Wasserstein distance) computed on the distributions of representations produced by different prompts. This representation-space repulsion diversifies exploration and prevents premature collapse to a single mode. Our approach allows for a more comprehensive characterization of the prompt posterior distribution, leading to improved generalization. In contrast to prior Bayesian prompt learning methods, our method provides a modular plug-and-play Bayesian extension of any existing prompt learning method based on maximum likelihood estimation. We demonstrate the efficacy of ReBaPL on several benchmark datasets, showing superior performance over state-of-the-art methods for prompt learning.

While Parameter-Efficient Fine-Tuning using Mixture-of-Experts (MoE) is a promising solution for continual learning (CL), it suffers from two critical failure modes: structural interference, where expert updates interfere, and compositional forgetting, where the model’s routing policy drifts. To address these issues, we introduce Spectral MoE, a novel framework built for CL from three core components. First, Spectral Experts are parameterized using unique, disjoint spectral masks to confine their learnable parameters to distinct frequency subspaces, ensuring a priori orthogonal updates that prevent structural interference. Second, a Dual-Router mechanism decouples online routing that learns new tasks from an offline memory that archives historical expert importance. Finally, this offline memory enables a Dynamic Consistency Projection, a geometric constraint that suppresses router drift and adaptively shields experts based on their past contributions, mitigating compositional forgetting. Validated on a strict cross-domain CL benchmark, our framework significantly outperforms existing methods, demonstrating superior knowledge retention and plasticity for new tasks. Code will be released upon acceptance.

Despite significant advances in talking avatar generation, existing methods face critical challenges: insufficient text-following capability for diverse actions, lack of temporal alignment between actions and audio content, and dependency on additional control signals such as pose skeletons. We present ActAvatar, a framework that achieves phase-level precision in action control through textual guidance by capturing both action semantics and temporal context. Our approach introduces three core innovations: (1) Phase-Aware Cross-Attention (PACA), which decomposes prompts into a global base block and temporally-anchored phase blocks, enabling the model to concentrate on phase-relevant tokens for precise temporal-semantic alignment; (2) Progressive Audio-Visual Alignment, which aligns modality influence with the hierarchical feature learning process—early layers prioritize text for establishing action structure while deeper layers emphasize audio for refining lip movements, preventing modality interference; (3) A two-stage training strategy that first establishes robust audio-visual correspondence on diverse data, then injects action control through fine-tuning on structured annotations, maintaining both audio-visual alignment and the model’s text-following capabilities. Extensive experiments demonstrate that ActAvatar significantly outperforms state-of-the-art methods in both action control and visual quality. The code will be made publicly available.

Human communication is inherently multimodal and social: words, prosody, and body language jointly carry intent. Yet most prior systems model human behavior as a translation task—co-speech gesture or text-to-motion that maps a fixed utterance to motion clips—without requiring agentic decision-making about when to move, what to do, or how to adapt across multi-turn dialogue. This leads to brittle timing, weak social grounding, and fragmented stacks where speech, text, and motion are trained or inferred in isolation. We introduce ViBES (Voice in Behavioral Expression and Synchrony), a conversational 3D agent that jointly plans language and movement and executes dialogue-conditioned body actions. Concretely, ViBES is a speech-language-behavior (SLB) model with a mixture-of-modality-experts (MoME) backbone: modality-partitioned transformer experts for speech, facial expression, and body motion. The model processes interleaved multimodal token streams with hard routing by modality (parameters are split per expert), while sharing information through cross-expert attention. By leveraging strong pretrained speech-language models, the agent supports mixed-initiative interaction: users can speak, type, or issue body-action directives mid-conversation, and the system exposes controllable behavior hooks for streaming responses. We further benchmark on multi-turn conversation with automatic metrics of dialogue–motion alignment and behavior quality, and observe consistent gains over strong co-speech and text-to-motion baselines. ViBES goes beyond “speech-conditioned motion generation” toward agentic virtual bodies where language, prosody, and movement are jointly generated, enabling controllable, socially competent 3D interaction.

Portrait animation from a single source image and a driving video is a long-standing problem.Recent approaches tend to adopt diffusion-based image/video generation models for realistic and expressive animation.However, none of these diffusion models realizes high-fidelity disentangled control between the head pose and facial expression, hindering applications like expression-only or pose-only editing and animation.To address this, we propose DeX-Portrait, a novel approach capable of generating expressive portrait animation driven by disentangled pose and expression signals.Specifically, we represent the pose as an explicit global transformation and the expression as an implicit latent code.First, we design a powerful motion trainer to learn both pose and expression encoders for extracting precise and decomposed driving signals.Then we propose to inject the pose transformation into the diffusion model through a dual-branch conditioning mechanism, and the expression latent through cross attention.Finally, we design a progressive hybrid classifier-free guidance for more faithful identity consistency.Experiments show that our method outperforms state-of-the-art baselines on both animation quality and disentangled controllability.

3D Gaussian representations have emerged as a powerful paradigm for digital head modeling, achieving photorealistic quality with real-time rendering. However, intuitive and interactive creation or editing of 3D Gaussian head models remains challenging. Although 2D sketches provide an ideal interaction modality for fast, intuitive conceptual design, they are sparse, depth-ambiguous, and lack high-frequency appearance cues, making it difficult to infer dense, geometrically consistent 3D Gaussian structures from strokes—especially under real-time constraints. To address these challenges, we propose SketchFaceGS, the first sketch-driven framework for real-time generation and editing of photorealistic 3D Gaussian head models from 2D sketches. Our method uses a feed-forward, coarse-to-fine architecture. A Transformer-based UV feature-prediction module first reconstructs a coarse but geometrically consistent UV feature map from the input sketch, and a 3D UV feature enhancement module refines it with high-frequency, photorealistic detail to produce a high-fidelity 3D head. For editing, we introduce a UV Mask Fusion technique combined with a layer-by-layer feature-fusion strategy, enabling precise, real-time, free-viewpoint modifications. Extensive experiments show that SketchFaceGS outperforms existing methods in both generation fidelity and editing flexibility, producing high-quality, editable 3D heads from sketches in a single forward pass.

Embodied Conversational Agents (ECAs) aim to emulate human face-to-face interaction through speech, gestures, and facial expressions. Current large language model (LLM)–based conversational agents lack embodiment and the expressive gestures essential for natural interaction. Existing solutions for ECAs either produce rigid, low-diversity motions, that are unsuitable for human-like interaction. Alternatively, generative methods for co-speech gesture synthesis yield natural body gestures but depend on future speech context and require long run-times. To bridge this gap, we present MIBURI, an online, causal framework for generating expressive co-speech gestures and facial expressions synchronized with real-time spoken dialogue. We first employ body-part aware gesture codecs that encode hierarchical motion details into multi-level discrete tokens. These tokens are then autoregressively generated by a two-dimensional causal framework conditioned on LLM-based speech-text embeddings, modeling both temporal dynamics and part-level motion hierarchy in real time. Further, we introduce contrastive objectives to encourage expressive and diverse gestures while preventing convergence to static poses. Comparative evaluations demonstrate that our causal and real-time approach produces natural and contextually aligned gestures against recent baselines.

People can view the same image differently: they focus on different regions, objects, and details in varying orders and describe them in distinct linguistic styles. This leads to substantial variability in image descriptions. However, existing models for personalized image description focus on linguistic style alone, with no prior work leveraging individual viewing patterns. We address this gap by explicitly modeling personalized viewing behavior as a core factor in description generation. Our method, DEPER (DEscription–PERception persona encoder), learns a subject embedding that captures both linguistic style and viewing behavior, guided by an auxiliary attention-prediction task. A lightweight adapter aligns these embeddings with a frozen vision–language model, enabling few-shot personalization without retraining. Across four datasets spanning diverse viewing tasks and both short and detailed descriptions, DEPER achieves a 24% average improvement, showing that modeling personalized attention produces more human-aligned and high-quality descriptions. We posit that understanding how people see helps predict what they say; modeling human diversity in perception can improve both performance and human alignment in multi-modal systems.

Despite significant progress in Vision-Language Navigation (VLN), existing approaches still rely on dense RGB videos that produce excessive patch tokens and lack explicit spatial structure, resulting in substantial computational overhead and limited spatial reasoning. To address these issues, we introduce the Geometry-Aware BEV (GA-BEV) —a compact, 3D-grounded feature representation that integrates both explicit and implicit geometric cues into multimodal large language model (MLLM)–based navigation systems. We construct BEV spatial maps from RGB-D inputs by projecting visual features into 3D space and aggregating them into an agent-centric layout that preserves geometric consistency while reducing token redundancy. To further enrich geometric understanding, we incorporate features from a pretrained 3D foundation model into the BEV space, injecting structural priors learned from large-scale 3D reconstruction tasks. Together, these complementary cues—explicit depth-based projection and implicit learned priors—yield compact yet spatially expressive representations that substantially improve navigation efficiency and performance. Experiments show that our method achieves state-of-the-art results using only navigation data, without DAgger augmentation or mixed VQA training, demonstrating the robustness and data efficiency of the proposed GA-VLN framework.

Most mapping tools remain point-and-click, making it hard to ask spatial questions or relate what a camera sees to its surrounding geography in a view-aware way. We present IMAIA — the Interactive Maps AI Assistant — which enables natural-language interaction with both vector (street) maps and satellite imagery, while enriching camera inputs with geospatial intelligence to help users interpret the world around them.IMAIA consists of two complementary modules:* Maps Plus, which treats the map as primary context by converting tiled vector or satellite views into a grid-aligned format that language models can query to resolve deictic references (e.g., “the flower-shaped building next to the park in the top-right”).* Places AI Smart Assistant (PAISA), which performs camera-aware place reasoning by fusing image–place embeddings with geospatial signals such as location, heading, and distance to ground the scene, highlight key attributes, and produce concise explanations.A lightweight multi-agent design ensures low latency and transparent intermediate reasoning. Across map-centric question answering and camera-to-place grounding tasks, IMAIA consistently improves accuracy and responsiveness over strong baselines while remaining efficient for real-world use. By uniting language, maps, and geospatial cues, IMAIA advances from scripted interactions to conversational mapping that is both spatially grounded and widely accessible.

Embodied navigation stands as a foundation pillar within the pursuit of embodied intelligence. However, previous navigation research is divided into different tasks/capabilities, e.g., ObjNav, ImgNav and VLN, where they differ in task settings/objectives and modalities, making datasets and methods designed individually. In this work, we take steps toward generalist navigation, which can follow free-form instructions that include arbitrary compounds of modality and capability.To achieve this, we propose a large-scale benchmark and corresponding method, termed OctoNav-Bench and OctoNav-R1. Specifically, OctoNav-Bench is constructed via a designed automatic annotation pipeline. We thoroughly craft instruction-trajectory pairs, where instructions are diverse in free-form with arbitrary modality and capability. Also, we construct a Think-Before-Action (TBA-CoT) dataset within OctoNav-Bench to provide the thinking process behind actions.For OctoNav-R1, we build it upon MLLMs and adapt it to a VLA-type model, which can produce low-level actions solely based on 2D visual observations. Moreover, we design a Hybrid Training Paradigm (HTP) that consists of three stages, i.e., Action-/TBA-SFT, Nav-GRPO, and Online RL stages. Each stage contains designed learning policies and rewards. Specifically, inspired by the OpenAI-o1 and DeepSeek-R1, which show impressive reasoning ability via thinking-before-answer, we design TBA-SFT and Nav-GRPO to achieve thinking-before-action for embodied navigation, improving model's reasoning ability toward generalists.TBA-SFT utilizes the TBA-CoT dataset to fine-tune the model, and then we leverage Nav-GRPO to improve its thinking ability. Finally, OctoNav-R1 shows superior performance compared with the previous methods.

Ensuring accessible pedestrian navigation requires reasoning about both semantic and spatial aspects of complex urban scenes, a challenge that existing Large Vision–Language Models (LVLMs) struggle to meet. Although these models can describe visual content, their lack of explicit grounding leads to object hallucinations and unreliable depth reasoning, limiting their usefulness for accessibility guidance. We introduce WalkGPT, a pixel-grounded LVLM for the new task of Grounded Navigation Guide, unifying language reasoning and segmentation within a single architecture for depth-aware accessibility guidance. Given a pedestrian-view image and a navigation query, WalkGPT generates a conversational response with segmentation masks that delineate accessible and harmful features along with relative depth estimation. The model incorporates a Multi-Scale Query Projector (MSQP) that shapes the final image tokens by aggregating them along text tokens across spatial hierarchies, and a Calibrated Text Projector (CTP), guided by a Region Alignment Loss, that maps language embeddings into segmentation-aware representations. These components enable fine-grained grounding and depth inference without user-provided cues or anchor points, allowing the model to generate complete and realistic navigation guidance. We also introduce PAVE, a large-scale benchmark of 41k pedestrian-view images paired with accessibility-aware questions and depth-grounded answers. Experiments show that WalkGPT achieves strong grounded reasoning and segmentation performance. The source code and dataset are included in the supplementary materials.

End-to-end autonomous driving methods built on vision language models (VLMs) have undergone rapid development driven by their universal visual understanding and strong reasoning capabilities obtained from the large-scale pretraining. However, we find that current VLMs struggle to understand fine-grained 3D spatial relationships which is a fundamental requirement for systems interacting with the physical world. To address this issue, we propose SpaceDrive, a spatial-aware VLM-based driving framework that treats spatial information as explicit positional encodings (PEs) instead of textual digit tokens, enabling joint reasoning over semantic and spatial representations. SpaceDrive employs a universal positional encoder to all 3D coordinates derived from multi-view depth estimation, historical ego-states, and text prompts. These 3D PEs are first superimposed to augment the corresponding 2D visual tokens. Meanwhile, they serve as a task-agnostic coordinate representation, replacing the digit-wise numerical tokens as both inputs and outputs for the VLM. This mechanism enables the model to better index specific visual semantics in spatial reasoning and directly regress trajectory coordinates rather than generating digit-by-digit, thereby enhancing planning accuracy. Extensive experiments validate that SpaceDrive achieves state-of-the-art open-loop performance on the nuScenes dataset and the second-best Driving Score of 78.02 on the Bench2Drive closed-loop benchmark over existing VLM-based methods. Code will be released upon acceptance.

Semantic route planning involves generating itineraries that align with user intent while respecting real-world spatial constraints. However, text-only large language models (LLMs) often hallucinate geographically implausible routes due to poor spatial grounding. Inspired by how humans use maps for route planning, we propose the SMAP, which is the first multimodal framework combining user queries, POI metadata, and map tiles to produce spatially coherent, preference-aware routes. To enhance the spatial consistency, the SMAP features a two-stage anti-hallucination mechanism: (1) a map-grounded self-editing pipeline where a multimodal LLM (MLLM) drafts routes and a second MLLM verifies and refines them using geographic evidence; and (2) hallucination-penalized Direct Preference Optimization (HDPO) that steers the route generator toward spatially plausible routes by using verified routes as accepted responses and hallucinated drafts as rejected ones. Additionally, we introduce MM-Route, the first multimodal dataset for semantic route planning, with 3,000 diverse queries annotated with POI metadata and map tiles, covering a broad spectrum of geographic granularities and user intents. Experimental results demonstrate that SMAP significantly reduces geographical hallucinations and outperforms strong baselines in spatial plausibility and user alignment. The code and dataset will be made publicly available.

Synthetic data has emerged as a practical alternative to authentic face datasets for training face recognition (FR) systems, especially as privacy and legal concerns increasingly restrict the use of real biometric data. Recent advances in identity-conditional diffusion models have enabled the generation of photorealistic and identity-consistent face images. However, many of these models suffer from limited intra-class variation, an essential property for training robust and generalizable FR models. In this work, we propose IDperturb, a simple yet effective geometric-driven sampling strategy to enhance diversity in synthetic face generation. IDperturb perturbs identity embeddings within a constrained angular region of the unit hyper-sphere, producing a diverse set of embeddings without modifying the underlying generative model. Each perturbed embedding serves as a conditioning vector for a pre-trained diffusion model, enabling the synthesis of visually varied yet identity-coherent face images suitable for training generalizable FR systems. Empirical results show that training FR on datasets generated using IDperturb leads to improved performance across multiple FR benchmarks, compared to existing synthetic data generation approaches. Code and generated datasets will be publicly released.

We propose Fresco, a unified optimization paradigm designed to mitigate early over-sharpening, and cross-view drifting in head avatar reconstruction. Fresco combines a Laplacian-pyramid-based frequency curriculum with UV-space consistency regularization to progressively enhance reconstruction quality. The optimization begins by stabilizing low-frequency appearance in the image domain, which suppresses spurious details and promotes reliable convergence. As learning proceeds, consistency across different viewpoints is reinforced through pixel-level alignment on shared UV texture coordinates. Finally, high-frequency components are refined under explicit frequency-band constraints, and seam boundary regularization is applied to preserve local continuity. By optimizing in a frequency- and UV-aligned space, Fresco achieves robust convergence without pseudo high-frequency artifacts and yields consistent, high-fidelity results across views. Experiments on the NeRSemble dataset validate the effectiveness of our design. Our method outperforms previous state-of the-art methods while avoiding additional training overhead through frequency scheduling and UV-bake caching.

The creation of 3D human avatars from multi-view videos is a significant yet challenging task in computer vision. However, existing techniques rely on high-quality, sharp images as input, which are often impractical to obtain in real-world scenarios due to variations in human motion speed and intensity. This paper introduces a novel method for directly reconstructing sharp 3D human Gaussian avatars from blurry videos. The proposed approach incorporates a 3D-aware, physics-based model of blur formation caused by human motion, together with a 3D human motion model designed to resolve ambiguities in motion-induced blur. This framework enables the joint optimization of the avatar representation and motion parameters from a coarse initialization. Comprehensive benchmarks are established using both a synthetic dataset and a real-world dataset captured with a 360-degree synchronous hybrid-exposure camera system. Extensive evaluations demonstrate the effectiveness and robustness of the model across diverse conditions.

We introduce ELITE, an Efficient Gaussian head avatar synthesis from a monocular video via Learned Initialization and TEst-time generative adaptation. Prior works rely either on a 3D data prior or a 2D generative prior to compensate for missing visual cues in monocular videos. However, 3D data prior methods often struggle to generalize in-the-wild, while 2D generative prior methods are computationally heavy and prone to identity hallucination. We identify a complementary synergy between these two priors and design an efficient system that achieves high-fidelity animatable avatar synthesis with strong in-the-wild generalization. Specifically, we introduce a feed-forward Mesh2Gaussian Prior Model (MGPM) that enables fast initialization of a Gaussian avatar. To further bridge the domain gap at test time, we design a test-time generative adaptation stage, leveraging both real and synthetic images as supervision. Unlike previous full diffusion denoising strategies that are slow and hallucination-prone, we propose a rendering-guided single-step diffusion enhancer that restores missing visual details, grounded on Gaussian avatar renderings. Our experiments demonstrate that ELITE produces visually superior avatars to prior works, even for challenging expressions, while achieving 60x faster synthesis than the 2D generative prior method.

Generating large-scale 3D head avatars of non-existent identities with high-fidelity and strong multi-view consistency (MVC) is essential for applications such as synthetic crowds, digital twins, and large asset libraries. For high scalability, avatars must be generated from minimal resources, without costly MV studio captures or any 3D data. In this work, we target this challenging minimal-resource setting for 3D head generation. Second, we argue that the common strategy of enforcing MVC via intermediate MV image generation is both expensive and fundamentally fragile. Instead, we analyze how MVC can be induced by design, showing that intermediate view synthesis is unnecessary. To this end, we introduce MVCHead — a fast, single-shot state space model that directly predicts Gaussians, without intermediate generation. At its core, we propose a Hierarchical State Space (HiSS) block that enforces grid-aligned coherence while capturing long-range dependencies. We further modify Mamba's standard unidirectional scanning into a Hierarchical Bi-directional State Scan (HiBiSS), scanning the render grid to better propagate geometric and appearance cues. Finally, we design an SE(3) MV Critic that judges whether a set of self-renders arise from a single underlying 3D configuration, rewarding cross-view pixel alignment without real MV data. In this setting, MVCHead surpasses SOTA in perceptual quality and on all three MVC axes—shape, texture, and geometry. The code has been submitted and will be open-sourced with model weights upon acceptance.

Multimodal Large Language Models (MLLMs) suffer from cross-modal hallucinations, where one modality inappropriately influences generation about another, leading to fabricated output. This exposes a more fundamental deficiency in modality-interaction control. To address this, we propose Modality-Adaptive Decoding (MAD), a training-free method that adaptively weights modality-specific decoding branches based on task requirements. MAD leverages the model's inherent ability to self-assess modality relevance by querying which modalities are needed for each task. The extracted modality probabilities are then used to adaptively weight contrastive decoding branches, enabling the model to focus on relevant information while suppressing cross-modal interference. Extensive experiments on CMM and AVHBench demonstrate that MAD significantly reduces cross-modal hallucinations across multiple audio-visual language models (7.8\% and 2.0\% improvements for VideoLLaMA2-AV, 8.7\% and 4.7\% improvements for Qwen2.5-Omni). Our approach demonstrates that explicit modality awareness through self-assessment is crucial for robust multimodal reasoning, offering a principled extension to existing contrastive decoding methods.

Large vision-language models (LVLMs) have shown remarkable capabilities in visual-language understanding. Despite their success, LVLMs still suffer from generating hallucinations in complex generation tasks, leading to inconsistencies between visual inputs and generated content. To address this issue, some approaches have introduced inference-time interventions, such as contrastive decoding, to reduce overreliance on language priors. However, these approaches overlook hallucinations stemming from position bias and spurious inter-modality correlations. In this paper, we propose a Cross-Modal Attention Calibration (CMAC) method to mitigate hallucinations in LVLMs in a training-free manner. In this method, we design an Inter-Modality Decoding (IMD) module to alleviate hallucination by a novel contrastive decoding mechanism. IMD masks the value vectors associated with significant cross-modal attention weights as distortion, which addresses both uni-modality overreliance and misleading inter-modality correlations. Additionally, a Cross-Modal Position Calibration (CMPC) module shrinks the position gap of image tokens, alleviating the position bias in cross-modal attention. Experimental results on diverse hallucination benchmarks validate the superiority of our method over existing state-of-the-art techniques in reducing hallucinations for LVLM. Our code will be available.

Large Language Models are increasingly integrated as the cognitive core of 3D embodied agents to enable complex environmental reasoning. However, these agents tend to inherit the critical flaw of hallucination, often failing to ground their responses to their 3D view. While Visual Contrastive Decoding (VCD) is a powerful training-free method for mitigating hallucinations in 2D image-based models, it has not been adapted to the complex 3D embodied environment. In this paper, we embarked on the ambitious goal of being the very first to bridge this gap by introducing a VCD framework for 3D embodied agents. Our method operates at inference time by generating a "negative" 3D context, not by blurring an image, but by applying novel distortions directly to a 3D scene graph, such as swapping object category labels or noising positional coordinates. We evaluate our approach on standard evaluation benchmarks and find that it consistently outperforms existing models. For example, in the random category of 3D-POPE, our 3D-VCD method reduced Yes-rate from 99.9\% to 75.1\% while simultaneously increasing precision from 50.0\% to 62.2\%. These results demonstrate that our training-free approach effectively curbs hallucination, yielding 3D agents that are significantly more reliable and grounded.

Existing GUI benchmarks primarily focus on evaluating models’ comprehensive capabilities but largely overlook hallucination phenomena in grounding tasks, which are crucial to the reliability of GUI understanding. In this work, we expose two major types of hallucinations in GUI grounding: 1) Confusion Hallucination, where distractor elements are mistakenly selected, and 2) Fabricated Hallucination, where nonexistent elements are hallucinated with plausible coordinates. To systematically investigate their origins, we introduce GUI-HalluBench, a benchmark comprising two complementary subsets: a parsing subset for assessing structural representation of GUI elements and a hallucination subset for measuring grounding robustness under challenging conditions. This design allows us to associate hallucination patterns with deficiencies in prerequisite abilities: parsing errors are closely tied to both fabricated and confusion hallucinations. Experiments on state-of-the-art models confirm these connections, offering new insights into the root causes of hallucinations and guiding the development of more reliable GUI understanding tools.

Multimodal Chain-of-Thought (MCoT) models have demonstrated impressive capability in complex visual reasoning tasks. Unfortunately, recent studies reveal that they suffer from severe hallucination problems due to diminished visual attention during the generation process.However, visual attention decay is a well-studied problem in Large Vision-Language Models (LVLMs). Considering the fundamental differences in reasoning processes between MCoT models and traditional LVLMs, we raise a basic question: Whether MCoT models have unique causes of hallucinations? To answer this question, we systematically investigate the hallucination patterns of MCoT models and find that fabricated texts are primarily generated in associative reasoning steps, which we term divergent thinking. Leveraging these insights, we introduce a simple yet effective strategy that can effectively localize divergent thinking steps and intervene in the decoding process to mitigate hallucinations. Extensive experiments show that our method outperforms existing methods by a large margin. More importantly, our proposed method can conveniently integrate with other hallucination mitigation methods and further boost their performance. The code will be released.

Large vision-language models (LVLMs) achieve strong performance on visual reasoning tasks but remain highly susceptible to hallucination. Existing detection methods predominantly rely on coarse, whole-image measures of how an object token relates to the input image. This global strategy is limited: hallucinated tokens may exhibit weak but widely scattered correlations across many local regions, which aggregate into deceptively high overall relevance, thus evading the current global hallucination detectors. We begin with a simple yet critical observation: a faithful object token must be strongly grounded in a specific image region. Building on this insight, we introduce a patch-level hallucination detection framework that examines fine-grained token-level interactions across model layers.Our analysis uncovers two characteristic signatures of hallucinated tokens: (i) they yield diffuse, non-localized attention patterns, in contrast to the compact, well-focused attention and (ii) they fail to exhibit meaningful semantic alignment with any visual region.Guided by these findings, we develop a lightweight and interpretable detection method that leverages patch-level statistical features, combined with hidden-layer representations. Our approach achieves up to 90% accuracy in token-level hallucination detection, demonstrating the superiority of fine-grained structural analysis for detecting hallucinations.

The growing adoption of XR devices has fueled strong demand for high-quality stereo video, yet its production remains costly and artifact-prone.To address this challenge, we present StereoWorld, an end-to-end framework that repurposes a pretrained video generator for high-fidelity monocular-to-stereo video generation. Our framework jointly conditions the model on the monocular video input while explicitly supervising the generation with a geometry-aware regularization to ensure 3D structural fidelity.A spatio-temporal tiling scheme is further integrated to enable efficient, high-resolution synthesis.To enable large-scale training and evaluation, we curate a high-definition stereo video dataset containing over 11M frames aligned to natural human interpupillary distance (IPD).Extensive experiments demonstrate that StereoWorld substantially outperforms prior methods, generating stereo videos with superior visual fidelity and geometric consistency.

Current autoregressive video diffusion models are constrained by three core bottlenecks: (i) the finite temporal horizon imposed by the base model's 3D Rotary Positional Embedding (3D-RoPE), (ii) slow prompt responsiveness in maintaining fine-grained action control during long-form rollouts, and (iii) the inability to realize discontinuous cinematic transitions within a single generation stream. We introduce $\infty$-RoPE, a unified inference-time framework that addresses all three limitations through three interconnected components: Block-Relativistic RoPE, KV Flush, and RoPE Cut. Block-Relativistic RoPE reformulates temporal encoding as a moving local reference frame, where each newly generated latent block is rotated relative to the base model's maximum frame horizon while earlier blocks are rotated backward to preserve relative temporal geometry. This relativistic formulation eliminates fixed temporal positions, enabling continuous video generation far beyond the base positional limits. To obtain fine-grained action control without re-encoding, KV Flush renews the KV cache by retaining only two anchor tokens, the global sink and the last generated latent frame, thereby ensuring immediate prompt responsiveness. Finally, RoPE Cut introduces controlled discontinuities in temporal RoPE coordinates, enabling multi-cut scene transitions within a single continuous rollout. Together, these components establish $\infty$-RoPE as a training-free foundation for infinite-horizon, controllable, and cinematic video diffusion. Comprehensive experiments show that $\infty$-RoPE consistently surpasses previous autoregressive models in overall VBench scores while requiring only a fraction of their KV cache budget.

Creating realistic 3D animation remains a time-consuming and expertise-dependent process, requiring manual rigging, keyframing, and fine-tuning of complex motions. Meanwhile, video diffusion models have recently demonstrated remarkable motion imagination in 2D, generating dynamic and visually coherent motion from text or image prompts. However, their results lack explicit 3D structure and cannot be directly used for animation or simulation. We present AnimaMimic, a framework that animates static 3D meshes using motion priors learned from video diffusion models. Starting from an input mesh, AnimaMimic synthesizes a monocular animation video, automatically constructs a skeleton with skinning weights, and refines joint parameters through differentiable rendering and video-based supervision. To further enhance realism, we integrate a differentiable simulation module that refines mesh deformation through physically grounded soft-tissue dynamics. Our method bridges the creativity of video diffusion and the structural control of 3D rigged animation, producing physically plausible, temporally coherent, and artist-editable motion sequences that integrate seamlessly into standard animation pipelines.

Video world models aim to simulate dynamic, real-world environments, yet existing methods struggle to provide unified and precise control over camera and multi-object motion, as videos inherently operate dynamics in the projected 2D image plane. To bridge this gap, we introduce VerseCrafter, a 4D-aware video world model that enables explicit and coherent control over both camera and object dynamics within a unified 4D geometric world state. Our approach is centered on a novel 4D Geometric Control representation, which encodes the world state through a static background point cloud and per-object 3D Gaussian trajectories. This representation captures not only an object's path but also its probabilistic 3D occupancy over time, offering a flexible, category-agnostic alternative to rigid bounding boxes or parametric models. These 4D controls are rendered into conditioning signals for a pretrained video diffusion model, enabling the generation of high-fidelity, view-consistent videos that precisely adhere to the specified dynamics. Unfortunately, another major challenge lies in the scarcity of large-scale training data with explicit 4D annotations. We address this by developing an automatic data engine that extracts the required 4D controls from in-the-wild videos, allowing us to train our model on a massive and diverse dataset. We will release our code, model, and dataset to support research in realistic and controllable video world models.

Recent advancements in 3D object generation using diffusion models have achieved remarkable success, but generating realistic 3D urban scenes remains challenging. Existing methods relying solely on 3D diffusion models tend to suffer a degradation in appearance details, while those utilizing only 2D diffusion models typically compromise camera controllability. To overcome this limitation, we propose ScenDi, a method for urban scene generation that integrates both 3D and 2D diffusion models. We first train a 3D latent diffusion model to generate 3D Gaussians, enabling the rendering of images at a relatively low resolution. To enable controllable synthesis, this 3DGS generation process can be optionally conditioned by specifying inputs such as 3d bounding boxes, road maps, or text prompts. Then, we train a 2D video diffusion model to enhance appearance details conditioned on rendered images from the 3D Gaussians. By leveraging the coarse 3D scene as guidance for 2D video diffusion, ScenDi generates desired scenes based on input conditions and successfully adheres to accurate camera trajectories. Experiments on two challenging real-world datasets, Waymo and KITTI-360, demonstrate the effectiveness of our approach.

We introduce MotionCrafter, the first video diffusion-based framework that jointly reconstructs 4D geometry and estimates dense motion from a monocular video. To represent them effectively in latent space, we propose a 4D VAE that encodes point maps and scene flows as a unified latent compatible with pretrained video generators. Unlike prior work that forces the 3D value and latents to align strictly with RGB VAE latents—despite their fundamentally different distributions—we show that such alignment is unnecessary and leads to suboptimal performance. Instead, we introduce a new data normalization and VAE training strategy that better transfers diffusion priors and greatly improves reconstruction quality. Extensive experiments across multiple datasets demonstrate that MotionCrafter achieves state-of-the-art performance in joint 4D geometry reconstruction and dense scene flow estimation, delivering 38.64\% and 25.0\% improvements in geometry and motion reconstruction, respectively, all without any post-optimization.

Recent advances in generative modeling have substantially enhanced novel view synthesis, yet maintaining consistency across viewpoints remains challenging. Diffusion-based models rely on stochastic noise-to-data transitions, which obscure deterministic structures and yield inconsistent view predictions.We propose a Data-to-Data Flow Matching framework that learns deterministic transformations directly between paired views, enhancing view-consistent synthesis through explicit data coupling.To further enhance geometric coherence, we introduce Probability Density Geodesic Flow Matching (PDG-FM), which constrains flow trajectories using geodesic interpolants derived from probability density metrics of pretrained diffusion models. Such alignment with high-density regions of the data manifold promotes more realistic interpolants between samples.Empirically, our method surpasses diffusion-based NVS baselines, demonstrating improved structural coherence and smoother transitions across views. These results highlight the advantages of incorporating data-dependent geometric regularization into deterministic flow matching for consistent novel view generation.

Recent advances in foundational Video Diffusion Models (VDMs) have yielded significant progress. Yet, despite the remarkable visual quality of generated videos, reconstructing consistent 3D scenes from these outputs remains challenging, due to limited camera controllability and inconsistent generated content when viewed from distinct camera trajectories.In this paper, we propose WorldStereo, a novel framework that bridges camera-guided video generation and 3D reconstruction via two dedicated geometric memory modules. Formally, the global-geometric memory enables precise camera control while injecting coarse structural priors through incrementally updated point clouds.Moreover, the spatial-stereo memory constrains the model's attention receptive fields with 3D correspondence to focus on fine-grained details from the memory bank.These components enable WorldStereo to generate multi-view-consistent videos under precise camera control, facilitating high-quality 3D reconstruction.Furthermore, the flexible control branch-based WorldStereo shows impressive efficiency, benefiting from the distribution matching distilled VDM backbone without joint training.Extensive experiments across both camera-guided video generation and 3D reconstruction benchmarks demonstrate the effectiveness of our approach. Notably, we show that WorldStereo acts as a powerful world model, tackling diverse scene generation tasks (whether starting from perspective or panoramic images) with high-fidelity 3D results. Models will be released.

In this paper, we propose NeoVerse, a versatile 4D world model that is capable of 4D reconstruction, novel-trajectory video generation, and rich downstream applications. We first identify a common limitation of scalability in current 4D world modeling methods, caused either by expensive and specialized multi-view 4D data or by cumbersome training pre-processing. In contrast, our NeoVerse is built upon a core philosophy that makes the full pipeline scalable to diverse in-the-wild monocular videos. Specifically, NeoVerse features pose-free feed-forward 4D reconstruction, online monocular degradation pattern simulation, and other well-aligned techniques. These designs empower NeoVerse with versatility and generalization to various domains. Meanwhile, NeoVerse achieves state-of-the-art performance in standard reconstruction and generation benchmarks.

Video diffusion models have rich world priors, but their use in spatial tasks is limited by poor control, spatial-temporal inconsistent results, and entangled scene-camera dynamics. Current approaches, such as per-task fine-tuning or post-process warping strategies, are insufficient, often introducing visual artifacts, failing to generalize, or incurring high computational costs. We introduce a novel, training-free framework that operates purely at inference time to resolve these issues. Our method is comprised of three synergistic components. First, an intra-step refinement loop injects fine-grained motion guidance during the denoising process, iteratively correcting the output to ensure strict adherence to the target camera path. Second, an optical flow-based analysis identifies and isolates motion-related channels within the latent space. This allows our framework to selectively apply guidance, thereby decoupling motion from appearance and preserving visual fidelity. Third, a dual-path guidance strategy adaptively corrects for drift by comparing the guided generation against an unguided, reference denoising path, effectively neutralizing artifacts caused by misaligned structural inputs. These components work in concert to inject precise, trajectory-aligned control without any model retraining, achieving both accurate motion guidance and photorealistic synthesis. Our plug-and-play, model-agnostic solution demonstrates broad applicability for 3D/4D tasks. Extensive experiments confirm state-of-the-art performance in trajectory adherence and perceptual quality, outperforming both training-dependent and other inference-only methods.

Recent text-to-video (T2V) models have demonstrated strong capabilities in producing high-quality, dynamic videos. To improve the visual controllability, recent works have considered fine-tuning pre-trained T2V models to support image-to-video (I2V) generation. However, such adaptation frequently suppresses motion dynamics of generated outputs, resulting in more static videos compared to their T2V counterparts. In this work, we analyze this phenomenon and identify that it stems from the premature exposure to high-frequency details in the input image, which biases the sampling process toward a shortcut trajectory that overfits to the static appearance of the reference image. To address this, we propose adaptive low-pass guidance (ALG), a simple training-free fix to the I2V model sampling procedure to generate more dynamic videos without compromising per-frame image quality. Specifically, ALG adaptively modulates the frequency content of the conditioning image by applying a low-pass filter at the early stage of denoising. Extensive experiments show ALG significantly improves the temporal dynamics of generated videos, while preserving or even improving image fidelity and text alignment. For instance, on the VBench test suite, ALG achieves a 33\% average improvement across models in dynamic degree while maintaining the original video quality.

Person De-Reidentification (De-ReID) is an emerging and safety-critical task that aims to selectively forget specific individuals in surveillance systems while preserving the recognition capability for others. Existing methods typically learn both forgetting and retaining objectives within a unified feature space, which leads to conflicting optimization goals and may cause unexpected performance degradation on novel retaining identities. Although decoupling pre-trained feature space for forgetting or retaining purpose is a promising solution, discriminating which feature space should be used for the given novel query remain unsolved. To alleviate these challenges, we propose SANER, advancing De-ReID with switchable adapter (SA) and test-time non-parametric enhanced routing (NER) algorithm. SA decouples the pre-trained feature space into two task-specific subspaces with forgetting adapter (F-Adapter) and retaining adapter (R-Adapter). The former suppresses identity-specific semantics for de-identification, while the latter preserves discriminative cues for accurate re-ID. Moreover, SA is further enhanced with NER to adaptively analyze optimal feature space routing for the given novel query at test-time. Specifically, NER compares queries with pre-computed prototypes in the original feature space, mitigating the potential training–testing gap and thus ensures accurate routing for De-ReID. Extensive experiments on multiple De-ReID benchmarks demonstrate SANER's efficacy, providing a new perspective for privacy-preserving visual perception.

Visible-Infrared Person Re-Identification (VI-ReID) is a challenging retrieval task due to the substantial modality gap between visible and infrared images. While existing methods attempt to bridge this gap by learning modality-invariant features within a shared embedding space, they often overlook the complex and implicit correlations between modalities. This limitation becomes more severe under distribution shifts, where infrared samples are often far fewer than visible ones. To address these challenges, we propose a novel network termed Bi-directional Interaction Transformation (BIT). Instead of relying on rigid feature alignment, BIT adopts a matching-based strategy that explicitly models the interaction between visible and infrared image pairs. Specifically, BIT employs an encoder-decoder architecture where the encoder extracts preliminary feature representations, and the decoder performs bi-directional feature integration and query aware scoring to enhance cross-modality correspondence. To our best knowledge, BIT is the first to introduce such pairwise matching-driven interaction in VI-ReID. Extensive experiments on several benchmarks demonstrate that our BIT achieves state-of-the-art performance, highlighting its effectiveness in the VI-ReID task.

Lifelong person re-identification (LReID) aims to learn from varying domains to obtain a unified person retrieval model. Existing LReID approaches typically focus on learning from scratch or a visual classification-pretrained model, while the Vision-Language Model (VLM) has shown generalizable knowledge in a variety of tasks. Although existing methods can be directly adapted to the VLM, since they only consider global-aware learning, the fine-grained attribute knowledge is underleveraged, leading to limited acquisition and anti-forgetting capacity. To address this problem, we introduce a novel VLM-driven LReID approach named Vision-Language Attribute Disentanglement and Reinforcement (VLADR). Our key idea is to explicitly model the universally shared human attributes to improve inter-domain knowledge transfer, thereby effectively utilizing historical knowledge to reinforce new knowledge learning and alleviate forgetting. Specifically, VLADR includes a Multi-grain Text Attribute Disentanglement mechanism that mines the global and diverse local text attributes of an image. Then, an Inter-domain Cross-modal Attribute Reinforcement scheme is developed, which introduces cross-modal attribute alignment to guide visual attribute extraction and adopts inter-domain attribute alignment to achieve fine-grained knowledge transfer. Experimental results demonstrate that our VLADR outperforms the state-of-the-art methods by 1.9%-2.2% and 2.1%-2.5% on anti-forgetting and generalization capacity. Our code will be released.

With the rapid advancement of generative models, generated image detection has become an important task in visual forensics. Although existing methods have achieved remarkable progress, they often rely, after training, on only a small subset of highly salient forgery cues, which limits their ability to generalize to unseen generative mechanisms. We argue that reliably generated image detection should not depend on a single decision path but should preserve multiple judgment perspectives, enabling the model to understand the differences between real and generated images from diverse viewpoints. Based on this idea, we propose an anti-feature-collapse learning framework that filters task-irrelevant components and suppresses excessive overlap among different forgery cues in the representation space, preventing discriminative information from collapsing into a few dominant feature directions. This design maintains diverse and complementary evidence within the model, reduces reliance on a small set of salient cues, and enhances robustness under unseen generative settings. Extensive experiments on multiple public benchmarks demonstrate that the proposed method significantly outperforms the state-of-the-art approaches in cross-model scenarios, achieving an accuracy improvement of 5.02\% and exhibiting superior generalization and detection reliability. The source code will be publicly available upon publication.

Human–Object Interaction (HOI) detection aims to localize human–object pairs and classify their interactions from a single image, a task that demands strong visual understanding and nuanced contextual reasoning. Recent approaches have leveraged Vision–Language Models (VLMs) to introduce semantic priors, significantly improving HOI detection performance. However, existing methods often fail to fully capitalize on the diverse contextual cues distributed across the entire scene. To overcome these limitations, we propose the Instance-centric Context Mining Network (InCoM-Net)—a novel framework that effectively integrates rich semantic knowledge extracted from VLMs with instance-specific features produced by an object detector. This design enables deeper interaction reasoning by modeling relationships not only within each detected instance but also across instances and their surrounding scene context. InCoM-Net comprises two core components: Instance-centric Context Refinement (ICR), which separately extracts intra-instance, inter-instance, and global contextual cues from VLM-derived features and Progressive Context Aggregation (ProCA), which iteratively fuses these multi-context features with instance-level detector features to support high-level HOI reasoning. Extensive experiments on the HICO-DET and V-COCO benchmarks show that InCoM-Net achieves state-of-the-art performance, surpassing previous HOI detection methods.

Person search, which aims to detect and re-identify individuals in unconstrained scenes, faces an inherent conflict in one-stage models: pedestrian detection focuses on shared human features, while person re-identification requires identity-specific representations. Existing approaches, such as feature decoupling and loss re-weighting, primarily address this issue in later network stages but fail to resolve early-stage feature entanglement. To overcome this limitation, we propose FSLoRA, a Freq-Spatial Low-Rank Adapter that progressively decouples task-specific features at the backbone level. FSLoRA consists of a Spatial-Level Module (SLM), which employs LoRA and a mixture-of-experts to dynamically activate task-relevant spatial features, and a Frequency-Level Module (FLM), which transforms features into the frequency domain to selectively enhance task-relevant frequency components while suppressing task-irrelevant noise. By integrating both spatial and frequency-based adaptations, FSLoRA reduces feature interference, enabling more effective joint optimization. Extensive experiments on CUHK-SYSU, PRW, and Posetrack21 demonstrate that FSLoRA not only achieves state-of-the-art performance but also serves as a plug-and-play module adaptable to various person search frameworks, offering a unified and generalizable solution for one-stage person search.

Decoding visual stimuli from electroencephalography (EEG) signals is a crucial step toward practical brain–computer interfaces (BCIs). However, this task requires large-scale and high-quality EEG–image paired datasets. Compared with abundant image data, the limited EEG recordings restrict the decoding models’ performance. To address this challenge, we propose EEGiT, a framework that converts sequential EEG signals into image-like EEG patches and enables the direct use of a pretrained Vision Transformer (ViT) as the EEG encoder. To preserve the spatial topology of brain regions and minimize distributional differences across channels, we group EEG electrodes according to anatomical structures and apply linear interpolation along the spatial dimension. We then resample the EEG signals to align the structure of EEG patches with that of image patches in ViT. This design encourages effective transfer of visual priors learned from large-scale image datasets to EEG representation learning. Experiments on the THINGS-EEG and EEG-3D datasets show that fine-tuning pretrained ViTs improves EEG-to-image retrieval and EEG-based visual classification, while maintaining robustness and strong cross-subject generalization. These results demonstrate a promising direction for leveraging powerful vision models to mitigate data scarcity in EEG decoding.

Federated Domain Generalization for Person Re-Identification (FedDG-ReID) aims to learn domain-invariant representations from decentralized data. Although Vision Transformers (ViTs) are widely adopted, their global attention often fails to distinguish pedestrians from high similarity backgrounds or diverse viewpoints---a challenge further amplified by cross-client distribution shifts in FedDG-ReID. To address this, we propose Federated Body Distribution Aware Visual Prompt (FedBPrompt), which introduces learnable visual prompts to explicitly guide Transformer attention toward pedestrian-centric regions. FedBPrompt employs a Body Distribution Aware Visual Prompts Mechanism (BAPM) that divides prompts into two groups: Holistic Full Body Prompts suppress cross-client background noise, while Body Part Alignment Prompts capture fine-grained details robust to pose and viewpoint variations. To mitigate the high communication cost of large Transformer models, we further design a Prompt-based Fine-Tuning Strategy (PFTS) that freezes the ViT backbone and updates only lightweight prompts, significantly reducing communication overhead while maintaining adaptability. Extensive experiments demonstrate that BAPM effectively enhances feature discrimination and cross-domain generalization, while PFTS}achieves notable performance gains within only a few aggregation rounds. Moreover, both BAPM and PFTS can be easily integrated into existing ViT-based FedDG-ReID frameworks, making FedBPrompt a flexible and effective solution for federated person re-identification. The code will be released.

Extending person re-identification (ReID) to a federated scenario has recently drawn attention due to privacy concerns of individuals, but existing methods mostly assume sufficient diversity in pose variations even within a decentralized client. We focus on a more realistic federated-by-camera scenario, where each client corresponds to a single camera and thus captures only a sparse set of poses. To enrich pose variety, we propose Pose-guided Enriched Feature Learning (PEFL) that explicitly augments pose-diverse samples in the federated ReID scenario. Specifically, a Pose-Extraction Module (PEM) disentangles pose-relevant and pose-irrelevant feature components, where Pose-Relationship Knowledge Distillation (PKD) method helps identify the correct pose and Semantic Consistency Maintenance (SCM) method preserves semantics even with pose changes. In addition, a Compatibility Regularization method ensures the PEM to be compatible with the feature space of the global model. By recombining pose-relevant and -irrelevant components across identities via PEM, our PEFL synthesizes pose-swapped features, thereby largely facilitating contrastive learning of ReID models. Extensive experiments on Market1501 and MSMT17 under the federated-by-camera setting demonstrate that PEFL consistently outperforms federated ReID baselines and their conjunctions with the existing feature augmentation methods; thus achieving state-of-the-art federated ReID performance.

Detecting Unmanned Aerial Vehicles (UAVs) in low-altitude environments is essential for perception and defense systems but remains highly challenging due to complex backgrounds, camouflage, and multimodal interference. In real-world scenarios, UAVs are frequently visually blended with surrounding structures such as buildings, vegetation, and power lines, resulting in low contrast, weak boundaries, and strong confusion with cluttered background textures. Existing UAV detection datasets, though diverse, are not specifically designed to capture these camouflage and complex-background challenges, which limits progress toward robust real-world perception. To fill this gap, we construct UAV-CB, a new RGB–T UAV detection dataset deliberately curated to emphasize complex low-altitude backgrounds and camouflage characteristics. Furthermore, we propose the Local Frequency Bridge Network (LFBNet), which models features in localized frequency space to bridge both the frequency–spatial fusion gap and the cross-modality discrepancy gap in RGB–T fusion. Extensive experiments on UAV-CB and public benchmarks demonstrate that LFBNet achieves state-of-the-art detection performance and strong robustness under camouflaged and cluttered conditions, offering a frequency-aware perspective on multimodal UAV perception in real-world applications. The UAV-CB dataset will be publicly released to support future research.

We introduce TimeViper, a hybrid vision-language model designed to tackle challenges of long video understanding.Processing hour-long videos demands both an efficient model architecture and an effective mechanism for handling extended temporal contexts. To this end, TimeViper adopts a hybrid Mamba-Transformer backbone that combines the efficiency of state-space models with the expressivity of attention mechanisms. Through this hybrid design, we reveal the vision-to-text information aggregation phenomenon, where information progressively flows from visual tokens to textual tokens across increasing LLM depth, resulting in severe vision token redundancy. Motivated by this observation, we propose TransV, a token information transfer module that transfers and compresses vision tokens into instruction tokens while maintaining multimodal understanding capabilities. This design enables TimeViper to process hour-long videos exceeding 10,000 frames. Extensive experiments across multiple benchmarks demonstrate that TimeViper competes with state-of-the-art models while extending input length. We further analyze attention behaviors of both Mamba and Transformer layers, offering new insights into hybrid model interpretability. This work represents an initial step towards developing, interpreting, and compressing hybrid Mamba-Transformer architectures. All code and model weights will be released.

Streaming video understanding often involves time-sensitive scenarios where models need to answer exactly when the supporting visual evidence appears: answering before the evidence reflects speculation, answering after it has passed reduces real-time utility. To capture this behavior, we introduce a readiness-aware formulation of streaming video understanding with the Answer Readiness Score (ARS), a timing-aware objective with asymmetric early and late penalties. When combined with correctness, ARS defines an effective accuracy that measures not just whether a model is right, but whether it answers at the appropriate moment. Building on this formulation, we introduce StreamReady, a framework to unify temporal reasoning with on-time answering through a lightweight readiness mechanism that decides if sufficient evidence has been observed before responding. To evaluate this capability, we further introduce ProReady-QA, a benchmark with annotated answer evidence windows and proactive multi-turn questions across local and global contexts. StreamReady achieves superior performance on ProReady-QA, and consistently outperforms prior methods across eight additional streaming and offline long-video benchmarks, demonstrating robust and broadly generalizable video understanding capability.

This paper addresses the critical and underexplored challenge of long video understanding with low computational budgets.We propose LongVideo-R1, an active, reasoning-equipped multimodal large language model (MLLM) agent designed for efficient video context navigation, avoiding the redundancy of exhaustive search.At the core of LongVideo-R1 lies a reasoning module that leverages high-level visual cues to infer the most informative video clip for subsequent processing.During inference, the agent initiates traversal from top-level visual summaries and iteratively refines its focus, immediately halting the exploration process upon acquiring sufficient knowledge to answer the query.To facilitate training, we first extract hierarchical video captions from CGBench, a video corpus with grounding annotations, and guide GPT-5 to generate 33K high-quality chain-of-thought-with-tool trajectories. The LongVideo-R1 agent is fine-tuned upon the Qwen-3-8B model through a two-stage paradigm: supervised fine-tuning (SFT) followed by reinforcement learning (RL), where RL employs a specifically designed reward function to maximize selective and efficient clip navigation.Experiments on multiple long video benchmarks validate the effectiveness of name, which enjoys superior tradeoff between QA accuracy and efficiency. All curated data and source code are provided in the supplementary material and will be made publicly available.

The rise of AI agents powered by large language models (LLMs) has transformed intelligent systems by enabling autonomous tool utilizing, reasoning, and action across diverse tasks. Despite this rapid progress, existing video summarization approaches primarily focus on feature extraction or frame-level importance regression but lack the autonomous reasoning, self-correction, and decision-making capabilities that define true agent-based intelligence. To bridge this gap, we propose AgenticVS—the first agentic workflow for video summarization that leverages multimodal large language models (MLLMs) to complete the summarization–verify–reflection loop in a fully autonomous manner. Rather than designing new architectures for feature extraction or regression, we exploit the understanding and reflective reasoning abilities of MLLMs to build an adaptive summarization framework with a self-reflecting workflow. Experiments on SumMe and TVSum demonstrate that our agentic workflow outperforms state-of-the-art methods, enhancing interpretability, adaptability, and paving the way for agent-based multimodal video understanding.

Video-based commonsense captioning aims to generate captions for the video content while providing multiple commonsense about the underlying events. Existing approaches rely on constructing a "video $\rightarrow$ content caption $\rightarrow$ commonsense" reasoning chain, which generates visually ungrounded commonsense and neglects inter-category commonsense correlations. Firstly, the existing reasoning chain induces the model's excessive reliance on content caption when generating commonsense, resulting in generic outputs with limited visual relevance. Secondly, the reasoning chain adopts multiple isolated decoders for commonsense generation, which fails to leverage the correlations between different categories of commonsense. To address these limitations, we introduce a novel self-critical distillation network (SCD-Net), which optimizes the reasoning chain by enhancing visual reasoning and establishing inter-category commonsense correlations. Specifically, on the one hand, we introduce self-critical learning and design a reward function to allow the model to refine its output. This mechanism incentivizes the model to maximize the utilization of visual information, thus improving the model's capacity for visual comprehension. On the other hand, we propose a joint reasoning distillation framework that fosters mutual inference among diverse commonsense categories. In this framework, we incorporate the cascaded decoder and knowledge distillation strategy to facilitate inter-category commonsense knowledge transfer while maintaining the fairness of the testing. Our experiments on the large-scale Video-to-Commonsense dataset demonstrate that our approach performs favorably against state-of-the-art methods. The code will be released soon.

We present the first systematic analysis of multimodal large language models (MLLMs) in personalized question-answering requiring ego-grounding - the ability to understand the camera-wearer in egocentric videos. To this end, we introduce MyEgo, the first egocentric VideoQA dataset designed to evaluate MLLMs' ability to understand, remember, and reason about the camera wearer. MyEgo comprises 541 videos and 5K personalized questions asking about "my things", "my activities", and "my past". Benchmarking reveals that competitive MLLMs across variants, including open-source vs. proprietary, thinking vs. non-thinking, small vs. large scales all struggle on MyEgo. Top models such as GPT-5 achieve only 46% accuracy, trailing human performance (85%) by a whopping 39%. Surprisingly, neither explicit reasoning nor model scaling yield consistent improvements. Models improve when relevant evidence is explicitly provided, but gains drop over time, indicating limitations in tracking and remembering "me" and "my past". These findings collectively highlight the crucial role of ego-grounding and long-range memory in enabling personalized QA in egocentric videos. We hope MyEgo and our analyses catalyze further progress in these areas for egocentric personalized assistance.

Processing long-form videos with Video Large Language Models (Video-LLMs) is computationally prohibitive. Current efficiency methods often compromise fine-grained perception through irreversible information disposal or inhibit long-range temporal modeling via rigid, predefined sparse patterns. This paper introduces AdaSpark, an adaptive sparsity framework designed to address these limitations. AdaSpark first partitions video inputs into 3D spatio-temporal cubes. It then employs two co-designed, context-aware components: (1) Adaptive Cube-Selective Attention (AdaS-Attn), which dynamically selects a subset of relevant video cubes to attend for each query token, and (2) Adaptive Token-Selective FFN (AdaS-FFN), which selectively processes only the most salient tokens within each cube. An entropy-based (Top-p) selection mechanism adaptively allocates computational resources based on input complexity. Experiments demonstrate that AdaSpark significantly reduces computational load by up to 57% FLOPs while maintaining comparable performance to dense models and preserving fine-grained, long-range dependencies, as validated on challenging hour-scale video benchmarks.

Video large language models (Video-LLMs) have demonstrated strong capabilities in video understanding tasks. However, their practical deployment is still hindered by the inefficiency introduced by processing massive amounts of visual tokens. Although recent approaches achieve extremely low token retention ratios while maintaining accuracy comparable to full-token baselines, most of them perform compression only at the late stage of prefilling, leaving the efficiency of the vision encoder unoptimized. In this paper, we first show that vision encoding contributes a large portion to the time-to-first-token (TTFT). Therefore, instead of compressing visual tokens only after the vision encoder, performing compression inside the encoder still leaves substantial room for exploration. Based on this insight, we propose EarlyTom, a training-free token compression framework that performs early-stage visual token compression inside the vision encoder, enabling significantly better TTFT reduction and higher throughput. In addition, we introduce a decoupled spatial token selection strategy that improves the overall compression effectiveness. EarlyTom reduces TTFT by up to 2.65$\times$ and FLOPs by up to 61\% on a single NVIDIA A100 GPU for the LLaVA-OneVision-7B model, while maintaining accuracy comparable to the full-token baseline. These improvements substantially enhance the practicality of deploying Video-LLMs in real-world production scenarios.

Learning transferable knowledge from unlabeled video data and applying it in new environments is a hallmark of advanced artificial intelligence. We present VideoWorld 2, which extends VideoWorld and offers the first investigation into learning transferable knowledge directly from raw real-world videos. At its core, VideoWorld 2 introduces a disentangled Latent Dynamics Model (dLDM) that decouples action dynamics from visual appearance: a pretrained video diffusion model handles appearance modeling, enabling the dLDM to learn latent codes that focus on compact and meaningful task-related changes. These latent codes are then modeled autoregressively as a sequence to learn task policies and support long-horizon reasoning. We evaluate VideoWorld 2 on real-world video handcraft making tasks, where prior video generation and latent-dynamics models struggle to operate reliably. VideoWorld 2 achieves over a 70% improvement in task success rate and produces coherent long execution videos. In robotics, we show that VideoWorld 2 can acquire effective manipulation knowledge from the Open-X dataset, which substantially improves task performance on CALVIN. This study reveals the potential of learning transferable world knowledge directly from raw videos, with all code, data, and models to be open-sourced for further research.

Recent Vision-Language Models (VLMs) have made remarkable progress in multimodal understanding tasks, yet their evaluation on long video understanding remains unreliable. Due to limited frame inputs, key frames necessary for answering the question may be missing from the model’s input. However, models that truthfully refuse to answer under such uncertainty are marked as incorrect, while those that guess may coincidentally produce the correct answer and thus obtain deceptively higher accuracy, leading to misleading evaluation results and encouraging models to guess rather than respond honestly.To address this issue, we introduce VirtueBench, a benchmark explicitly designed to assess model trustworthiness under uncertainty. VirtueBench constructs multiple frame-sampling levels for each video and provides ground truths that distinguish between answerable and unanswerable cases. Evaluations on 25 open-source and commercial VLMs reveal distinct refusal behaviors across different model families, with refusal accuracy ranging from over 70\% in the best models to nearly 0\% in the worst. Moreover, most models exhibit a substantial drop in refusal when the prompt does not explicitly require them to do so. These findings highlight the need for developing trustworthy VLMs for multimodal understanding, guided by benchmarks and leaderboards that emphasize reliability and trustworthiness.

Recent breakthroughs in Diffusion Transformers (DiTs) have revolutionized the field of visual synthesis due to their superior scalability. To facilitate DiTs' capability of capturing meaningful internal representations, recent works such as REPA incorporate external pretrained encoders for representation alignment. However, the underlying mechanisms governing representation learning within DiTs are not well understood. To this end, we first systematically investigate the representation dynamics of DiTs. Through analyzing the evolution and influence of internal representations under various settings, we reveal that representation diversity across blocks is a crucial factor for effective learning. Based on this key insight, we propose DiverseDiT, a novel framework that explicitly promotes representation diversity. DiverseDiT incorporates long residual connections to diversify input representations across blocks and a representation diversity loss to encourage blocks to learn distinct features. Extensive experiments on ImageNet 256 $\times$ 256 and 512 $\times$ 512 demonstrate that our DiverseDiT yields consistent performance gains and convergence acceleration when applied to different backbones with various sizes, even when tested on the challenging one-step generation setting. Furthermore, we show that DiverseDiT is complementary to existing representation learning techniques, leading to further performance gains. Our work provides valuable insights into the representation learning dynamics of DiTs and offers a practical approach for enhancing their performance. Our code and models will be released.

Conventional physically-based rendering (PBR) pipelines generate photorealistic images through computationally expensive light transport simulations. Although recent deep learning approaches leverage diffusion model priors with geometry buffers (G-buffers) to produce visually compelling results without explicit scene geometry or light simulation, they remain constrained by two major limitations. First, the iterative nature of the diffusion process introduces substantial latency. Second, the inherent stochasticity of these generative models compromises physical accuracy and temporal consistency. In response to these challenges, we propose a novel, end-to-end, deterministic single-step neural rendering framework \textit{RenderFlow} built upon a flow matching paradigm. To further strengthen both rendering quality and generalization, we propose an efficient and effective module for sparse keyframe guidance. Our method significantly accelerates the rendering process and, by optionally incorporating sparsely rendered keyframes as guidance, enhances both the physical plausibility and overall visual quality of the output. The resulting pipeline achieves near real-time performance with photorealistic rendering quality, effectively bridging the gap between the efficiency of modern generative models and the precision of traditional physically based rendering. Furthermore, we demonstrate the versatility of our framework by introducing a lightweight, adapter-based module that efficiently repurposes the pretrained forward model for the inverse rendering task of intrinsic decomposition.

Leveraging pre-trained Diffusion Transformers (DiTs) for high-resolution (HR) image synthesis often leads to spatial layout collapse and degraded texture fidelity. Prior work mitigates these issues with complex pipelines that first perform a base-resolution (i.e., training-resolution) denoising process to guide HR generation. We instead explore the intrinsic generative mechanisms of DiTs and propose ResDiT, a training-free method that scales resolution efficiently. We identify the core factor governing spatial layout, position embeddings (PEs), and show that the original PEs encode incorrect positional information when extrapolated to HR, which triggers layout collapse. To address this, we introduce a PE scaling technique that rectifies positional encoding under resolution changes. To further remedy low-fidelity details, we develop a local-enhancement mechanism grounded in base-resolution local attention. We design a patch-level fusion module that aggregates global and local cues, together with a Gaussian-weighted splicing strategy that eliminates grid artifacts. Comprehensive evaluations demonstrate that ResDiT consistently delivers high-fidelity, high-resolution image synthesis and integrates seamlessly with downstream tasks, including spatially controlled generation.

Layered image generation and editing is a fundamental capability that enables layer-wise reuse, editing, and composition of the generated visual content, analogous to word-level editing in natural language. Despite its importance, this remains an underexplored area at scale. To address this gap, we present the Masked Region Transformer, a 20B-parameter diffusion model tailored for multi-layer transparent image generation and editing, trained on over 10M multilingual design samples spanning diverse aspect ratios and textual prompts. To fully leverage this scale, we make three key technical contributions. First, we unify three complementary tasks---text-to-layers, image-to-layers, and layers-to-layers---within a shared masked region diffusion framework, where selective token masking enables flexible cross-modal generation and fine-grained layer-wise editing. Second, we design an efficient conditional diffusion decoder that incorporates Gated DeltaNet and gated attention mechanisms, enhancing visual fidelity while maintaining computational efficiency. Third, we introduce an overflow-aware canvas layer to handle boundary inconsistencies and support semi-transparent background synthesis, enabling complete editable layer generation beyond visible canvas boundaries. Additionally, we apply distribution matching distillation to achieve one-step, real-time multi-layer generation with minimal quality degradation. Extensive experiments demonstrate that our framework substantially outperforms prior state-of-the-art approaches across all three tasks, establishing a new benchmark for region-aware transparent image generation.

Diffusion transformers have demonstrated remarkable generation quality, albeit requiring longer training iterations and numerous inference steps. In each denoising step, diffusion transformers encode the noisy inputs to extract the lower-frequency semantic component and then decode the higher frequency with identical modules. This scheme creates an inherent optimization dilemma: encoding low-frequency semantics necessitates reducing high-frequency components, creating tension between semantic encoding and high-frequency decoding. To resolve this challenge, we propose a new \textbf{\color{ddt}D}ecoupled \textbf{\color{ddt}D}iffusion \textbf{\color{ddt}T}ransformer(\textbf{\color{ddt}DDT}), with a decoupled design of a dedicated condition encoder for semantic extraction alongside a specialized velocity decoder. Our experiments reveal that a more substantial encoder yields performance improvements as model size increases. For ImageNet $256\times256$, Our DDT-XL/2 achieves a new state-of-the-art performance of {1.31 FID}~(nearly $4\times$ faster training convergence compared to previous diffusion transformers). For ImageNet $512\times512$, Our DDT-XL/2 achieves a new state-of-the-art FID of 1.28. Additionally, as a beneficial by-product, our decoupled architecture enhances inference speed by enabling the sharing self-condition between adjacent denoising steps. To minimize performance degradation, we propose a novel statistical dynamic programming approach to identify optimal sharing strategies.