CVPR 2026 论文列表

Visual attention serves as the primary mechanism through which MLLMs interpret visual information; however, its limited localization capability often leads to hallucinations. We observe that although MLLMs can accurately extract visual semantics from visual tokens, they fail to fully leverage this advantage during subsequent inference. To address this limitation, we propose Vision-Guided Attention (VGA), a training-free method that first constructs precise visual grounding by exploiting the semantic content of visual tokens, and then uses this grounding to guide the model's focus toward relevant visual regions. In image captioning, VGA further refines this guidance dynamically during generation by suppressing regions that have already been described. In VGA, each token undergoes only a single forward pass, introducing a negligible latency overhead. In addition, VGA is fully compatible with efficient attention implementations such as FlashAttention. Extensive experiments across diverse MLLMs and multiple hallucination benchmarks demonstrate that VGA achieves state-of-the-art dehallucination performance. Further analysis confirms that explicit visual guidance plays a crucial role in enhancing the visual understanding capabilities of MLLMs.

Estimating human gaze targets from images in-the-wild is an important and formidable task. Existing approaches primarily employ brittle, multi-stage pipelines that require explicit inputs, like head bounding boxes and human pose, in order to identify the subject of gaze analysis. As a result, detection errors can cascade and lead to failure. Moreover, these prior works lack the flexibility of specifying the gaze analysis task via natural language prompting, an approach which has been shown to have significant benefits in convenience and scalability for other image analysis tasks. To overcome these limitations, we introduce the Promptable Gaze Target Estimation (PGE) task, a new end-to-end, concept-driven paradigm for gaze analysis. PGE conditions gaze prediction on flexible user text or visual prompts (e.g., "the boy in the red shirt" or "person in point [0.52, 0.48]") to identify a specific subject for gaze analysis. This approach integrates subject localization with gaze estimation, and eliminates the rigid dependency on intermediate analysis stages. We develop a scalable data engine to generate Gaze-Co (Gaze Estimation with Concepts), a dataset and benchmark of 120K high-quality, prompt-annotated image pairs. We also propose GazeAnywhere, the first model designed for PGE. GazeAnywhere uses a transformer-based detector to fuse features from frozen encoders and simultaneously solves subject localization, in/out-of-frame presence, and gaze target heatmap estimation. GazeAnywhere achieves state-of-the-art performance on multiple PGE benchmarks, setting a strong baseline for this new problem even on a difficult out-of-domain, real-world clinical dataset. GazeAnywhere is open-sourced in github.com/IrohXu/GazeAnywhere.

We introduce RefTon, a flux-based person-to-person virtual try-on framework that enhances garment realism through unpaired visual references. Unlike conventional approaches that rely on complex auxiliary inputs such as body parsing and warped mask or require finely designed extract branches to process various input conditions, RefTon streamlines the process by directly generating try-on results from a source image and a target garment, without the need for structural guidance or auxiliary components to handle diverse inputs. Moreover, inspired by human clothing selection behavior, RefTon leverages additional reference images (the target garment worn on different individuals) to provide powerful guidance for refining texture alignment and maintaining the garment details. To enable this capability, we built a dataset containing unpaired reference images for training. Extensive experiments on public benchmarks demonstrate that RefTon achieves competitive or superior performance compared to state-of-the-art methods, while maintaining a simple and efficient person-to-person design. \thanks Code and dataset is available at: https://github.com/360CVGroup/RefTon.

Continual Visual Question Answering (Continual VQA) poses unique challenges for multimodal continual learning, requiring models to incrementally acquire new knowledge while preserving visual-semantic grounding across tasks. However, existing benchmarks hinder fair and robust evaluation of such capabilities, as they allow models to exploit dataset biases rather than demonstrate genuine continual reasoning. We identify two structural flaws in current benchmark design. First, shared answer vocabularies across tasks encourage answer memorization, inflating performance and underestimating forgetting. Second, static answer priors within each task make the training and test answer distributions nearly identical, obscuring robustness under distribution shifts. To address these issues, we introduce UCo-VQA, an Unbiased benchmark suite that enforces token-level disjoint answer spaces across tasks and introduces intra-task train-test distribution shifts, enabling fairer assessment of forgetting and generalization in multimodal continual learning. We further provide a parameter-efficient baseline that mitigates forgetting and enhances grounding through question-only replay and dual-level distillation, offering a lightweight and memory-efficient framework for continual adaptation. Extensive experiments on UCo-VQA reveal that prior methods substantially overestimate performance under biased setups, while our approach achieves state-of-the-art results, improving robustness and retention by up to 4.18% and 2.21%, respectively.

Embodied Question Answering (EQA) connects perception, reasoning, and interaction within embodied environments. However, existing datasets and benchmarks remain fragmented, each focusing on a limited subset of reasoning skills such as spatial understanding or procedural reasoning, without offering a unified large-scale framework for comprehensive evaluation. We present EQA-Decision, a large-scale embodied QA dataset that systematically covers four complementary dimensions of embodied reasoning: static scene construction, spatial understanding, task dynamics reasoning, and instant decision. The dataset contains over four million question-answer pairs with hierarchical annotations across diverse embodied scenarios. In addition, we develop RoboDecision, a strong baseline model aligned with the EQA-Decision Benchmark, providing a unified framework that jointly evaluates perception, reasoning, and action-level decision-making in embodied environments. Results demonstrate that EQA-Decision effectively benchmarks and enhances VLM capabilities in spatial and interaction reasoning, providing a solid foundation for advancing embodied intelligence research.

We study instruction-guided editing of egocentric videos for interactive AR applications. While recent AI video editors perform well on third-person footage, egocentric views present unique challenges -- including rapid egomotion, and frequent hand-object interactions -- that create a significant domain gap. Moreover, existing offline editing pipelines suffer from high latency, limiting real-time interaction.To address these issues, we present a complete ecosystem for egocentric video editing. First, we construct EgoEditData, a carefully designed and manually curated dataset specifically designed for egocentric editing scenarios, featuring rich hand-object interactions, while explicitly preserving hands. Second, we develop EgoEdit, an instruction-following egocentric video editor that supports real-time streaming inference on a single GPU. Finally, we introduce EgoEditBench, an evaluation suite targeting instruction faithfulness, hand and interaction preservation, and temporal stability under egomotion.Across both egocentric and general editing tasks, EgoEdit produces temporally stable, instruction-faithful results with interactive latency. It achieves clear gains on egocentric editing benchmarks--where existing methods struggle--while maintaining performance comparable to the strongest baselines on general editing tasks. EgoEditData and EgoEditBench will be made public for the research community.

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

Existing dynamic scene rendering methods often adopt rigid-body or direction-limited assumptions, yet real-world motion and contact routinely violate these, producing artifacts near occlusion boundaries. To address this, we introduce a unified, source-aware framework for dynamic rendering that enforces the consistency of Gaussian primitives under explicit manifold constraints. We project predicted velocities onto physically grounded priors via efficient, parallel inner solves: (i) a Helmholtz parameterization that separates divergence-free and potential-flow motion components; (ii) an anisotropic, compressible directional prior; and (iii) an affine family that disentangles rotation from isotropic scaling. Experiments on extensive benchmarks show consistent improvements over state-of-the-art methods in reconstruction fidelity and temporal coherence. Our approach ensures physically realistic rendering, especially near contacts, and substantially reduces motion-boundary artifacts.

Multimodal Emotion Analysis (MEA) is crucial for human-centric AI, yet current methods struggle with two core challenges: the sparse nature of emotional cues across modalities and their inherent temporal asynchrony. Existing approaches, which often rely on implicit fusion, consequently suffer from diluted salient features and entangled representations. To address this issue, we propose EmoThinker, a new framework that advances MEA through explicit, structured reasoning. Our method introduces a structural token selection mechanism to concentrate on pivotal facial regions while refining background context, enhancing visual saliency and efficiency. For audio, an audio evidence extractor aggregates critical paralinguistic features into compact, emotion-rich tokens. More importantly, we enable step-by-step reasoning by constructing a Chain-of-Emotion-Thought dataset, which provides fine-grained annotations for disentangling asynchronous cues and resolving inter-modal conflicts. By decoupling evidence acquisition from reasoning, EmoThinker achieves a more interpretable and robust emotion analysis. Extensive experiments on multiple benchmarks demonstrate that our framework achieves new state-of-the-art performance.

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

Visual Autoregressive (VAR) modeling introduces a new paradigm for image generation by extending autoregressive mechanisms from next-token prediction to next-scale prediction, achieving remarkable performance. However, as the number of tokens increases rapidly with scale, processing full token maps at high resolution becomes computationally expensive. In addition, the inherently sequential nature of autoregressive modeling prevents parallel inference across scales, which further increases latency.To address these challenges, we propose LazyVAR, a training-free and plug-and-play acceleration method for VAR models. Our key observation is that the similarity of aggregated latent features between adjacent scales progressively increases with the scale index, reaching particularly higher values at larger scales. We treat this similarity as a Scale-wise Update Index, which serves as the pruning criterion. Consequently, more tokens can be pruned at larger scales to improve efficiency. Furthermore, we propose Parallel Group Decoding, which leverages this high similarity at larger scales to decode tokens from different scales in parallel, further accelerating inference. Experimental results show that the proposed LazyVAR achieves up to a 2.94x speedup of recent VAR models with negligible performance degradation, allowing the Infinity-2B text-to-image model to generate 1024x1024 resolution images within 0.5 seconds on a single RTX 4090 GPU.

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

Solving partial differential equations (PDEs) on shapes underpins many shape analysis and engineering tasks; yet, prevailing PDE solvers operate on polygonal/triangle meshes while modern 3D assets increasingly live as neural representations. This mismatch leaves no suitable method to solve surface PDEs directly within the neural domain, forcing explicit mesh extraction or per-instance residual training, preventing end-to-end workflows. We present a novel, mesh-free formulation that learns a local update operator conditioned on neural (local) shape attributes, enabling surface PDEs to be solved directly where the (neural) data lives. The operator integrates naturally with prevalent neural surface representations, is trained once on a single representative shape, and generalizes across shape and topology variations, enabling accurate, fast inference without explicit meshing or per-instance optimization while preserving differentiability. Across analytic benchmarks (heat diffusion and Poisson equations on the sphere) and on diverse shapes and neural surface representations, our method achieves accuracy comparable to classical solvers while enabling a unified, end-to-end pipeline across neural and traditional surface representations. Our source code and project page: https://welschinger.github.io/Learning-to-Solve-PDEs-on-Neural-Shape-Representations/.

Reconstructing and semantically interpreting 3D scenes from sparse 2D views remains a fundamental challenge in computer vision. Conventional methods often decouple semantic understanding from reconstruction or necessitate costly per-scene optimization, thereby restricting their scalability and generalizability. In this paper, we introduce Uni3R, a novel feed-forward framework that jointly reconstructs a unified 3D scene representation enriched with open-vocabulary semantics, directly from unposed multi-view images. Our approach leverages a Cross-View Transformer to robustly integrate information across arbitrary multi-view inputs, which then regresses a set of 3D Gaussian primitives endowed with semantic feature fields. This unified representation facilitates high-fidelity novel view synthesis, open-vocabulary 3D semantic segmentation, and depth prediction--all within a single, feed-forward pass. Extensive experiments demonstrate that Uni3R sets a new state of the art across multiple benchmarks, including in-domain datasets such as RE10K and ScanNet, as well as the out-of-domain dataset Mip-NeRF360. This work represents a new paradigm toward generalizable and unified 3D scene reconstruction and understanding.

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.

Masked video modeling (MVM) has emerged as a simple and scalable self-supervised pretraining paradigm, but only encodes motion information implicitly, limiting the encoding of temporal dynamics in the learned representations. As a result, such models struggle on motion-centric tasks that require fine-grained motion awareness. To address this, we propose TrackMAE, a simple masked video modeling paradigm that explicitly uses motion information as a reconstruction signal. In TrackMAE, we use an off-the-shelf point tracker to sparsely track points in the input videos generating motion trajectories. Furthermore, we exploit the extracted trajectories to improve the random tube masking with a motion-aware masking strategy. We enhance video representations learned in both pixel and feature semantic reconstruction space by providing a complementary supervision signal in the form of motion targets. We evaluate on six datasets across diverse downstream settings and find that TrackMAE consistently outperforms the state-of-the-art video SSL baselines, therefore learning more discriminative and generalizable representations Code available at https://github.com/rvandeghen/TrackMAE

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.

We present MOSAIC-GS, a novel, fully explicit, and computationally efficient approach for high-fidelity dynamic scene reconstruction from monocular videos using Gaussian Splatting.Monocular reconstruction is inherently ill-posed due to the lack of sufficient multiview constraints, making accurate recovery of object geometry and temporal coherence particularly challenging. To address this, we leverage multiple geometric cues, such as depth, optical flow, dynamic object segmentation, and point tracking. Combined with rigidity-based motion constraints, these cues allow us to estimate preliminary 3D scene dynamics during an initialization stage.Recovering scene dynamics prior to the photometric optimization reduces reliance on motion inference from visual appearance alone, which is often ambiguous in monocular settings.To enable compact representations, fast training, and real-time rendering while supporting non-rigid deformations, the scene is decomposed into static and dynamic components. Each Gaussian in the dynamic part of the scene is assigned a trajectory represented as time-dependent Poly-Fourier curve for parameter-efficient motion encoding.We demonstrate that MOSAIC-GS achieves substantially faster optimization and rendering compared to existing methods,while maintaining reconstruction quality on par with state-of-the-art approaches across standard monocular dynamic scene benchmarks.

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.