CVPR 2026 论文列表

Single Image Super-Resolution (SISR) aims to reconstruct a high-resolution (HR) image from a low-resolution (LR) input, a task that becomes increasingly challenging under real-world computational constraints. However, most efficient SISR methods adopt lightweight, spatially uniform strategies that allocate equal computation and attention across all regions, ignoring the uneven distribution of visual complexity. From an information-theoretic perspective, textures and edges inherently carry more critical information, resulting in reconstruction errors that are disproportionately concentrated in these regions. This motivates the allocation of greater computational resources and attention to these informative areas. In this paper, we propose IAFMNet, an Information-Aware Feature Modulation network for efficient SR. At its core lies the Information Density Map (IDM), which is estimated in an unsupervised manner by minimizing the Information Entropy Loss, thereby highlighting informative regions with high estimated encoding costs. Guided by the IDM, IAFMNet adopts a synergistic dual-branch design: (1) a sparse convolution branch that dynamically allocates computation to informative areas while bypassing low-information regions, and (2) an implicit modulation branch that adaptively emphasizes complex regions through information-aware affine transformations. Extensive experiments demonstrate that IAFMNet effectively identifies informative regions and achieves superior visual fidelity with reduced computational overhead.

Tracking 3D human motion from egocentric, multi-camera devices is challenged by severe egomotion and partial visibility or occlusions. Existing methods are designed for monocular video often recorded from static or slowly-moving cameras and cannot easily leverage multi-view, calibrated and localized input. This makes them brittle and prone to fail on dynamic egocentric captures. We propose LAMP (Localization Aware Multi-camera People Tracking): a novel, simple framework to solve this via early disentanglement of observer and target motion. LAMP introduces a two-step process: First, we leverage the device's known 6-DoF pose and calibration to convert detected 2D body keypoints from all cameras over a temporal window into a unified 3D world reference frame. Second, an end-to-end-trained Transformer model fits 3D human motion directly to this spatio-temporal ray cloud in world coordinates. This "lift-then-fit" approach allows to learn and leverage a natural prior over world-space human motion, as well as providing an elegant framework to flexibly incorporate information from multiple, temporally asynchronous, partially observing, and moving cameras. LAMP achieves state-of-the-art results on monocular benchmarks, while significantly outperforming baselines for our targeted egocentric multi-camera setting.

Recent vision-language-action (VLA) models for multi-task robot manipulation often rely on fixed camera setups and shared visual encoders, which limit their performance under occlusions and during cross-task transfer. To address these challenges, we propose Task-aware Virtual View Exploration (TVVE), a framework that learns to select task-relevant virtual camera viewpoints and dynamically re-render observations from a reconstructed scene representation using the selected viewpoints. To enable efficient view selection, we train an exploration policy in a pseudo-environment. In addition, we introduce a Task-aware Mixture-of-Experts (TaskMoE) visual encoder that routes visual features to task-specialized experts, mitigating interference in multi-task learning. To evaluate robustness under distribution shifts, we construct RLBench-OG, an out-of-distribution benchmark with visual perturbations and camera pose variations. Experiments on RLBench and RLBench-OG demonstrate that TVVE achieves higher success rates than strong baselines, while real-robot experiments further confirm its robustness to visual disturbances and unseen instructions. Code and visualizations are available at: https://hcplab-sysu.github.io/TAVP.

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

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 head importance-aware visual token 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.

Concept erasure in text-to-image diffusion models seeks to remove undesired concepts while preserving overall generative capability. Localized erasure methods aim to restrict edits to the spatial region occupied by the target concept. However, we observe that suppressing a concept can unintentionally weaken semantically related neighbor concepts, reducing fidelity in fine-grained domains. We propose Neighbor-Aware Localized Concept Erasure (NLCE), a training-free framework designed to better preserve neighboring concepts while removing target concepts. It operates in three stages: (1) a spectrally-weighted embedding modulation that attenuates target concept directions while stabilizing neighbor concept representations, (2) an attention-guided spatial gate that identifies regions exhibiting residual concept activation, and (3) a spatially-gated hard erasure that eliminates remaining traces only where necessary. This neighbor-aware pipeline enables localized concept removal while maintaining the surrounding concept neighborhood structure. Experiments on fine-grained datasets (Oxford Flowers, Stanford Dogs) show that our method effectively removes target concepts while better preserving closely related categories. Additional results on celebrity identity, explicit content and artistic style demonstrate robustness and generalization to broader erasure scenarios.

Complex degradations like noise, blur, and low resolution are typical challenges in real-world image fusion tasks, limiting the performance and practicality of existing methods. End-to-end neural network-based approaches are generally simple to design and highly efficient in inference, but their black-box nature leads to limited interpretability. Diffusion-based methods alleviate this to some extent by providing powerful generative priors and a more structured inference process. However, they are trained to learn a single-domain target distribution, whereas fusion lacks natural fused data and relies on modeling complementary information from multiple sources, making diffusion hard to apply directly in practice. To address these challenges, this paper proposes an efficient degradation-aware diffusion framework for image fusion under arbitrary degradation scenarios. Specifically, instead of explicitly predicting noise as in conventional diffusion models, our method performs implicit denoising by directly regressing the fused image, enabling flexible adaptation to diverse fusion tasks under complex degradations with limited steps. Moreover, we design a joint observation model correction mechanism that simultaneously imposes degradation and fusion constraints during sampling to ensure high reconstruction accuracy. Experiments on diverse fusion tasks and degradation configurations demonstrate the superiority of the proposed method under complex degradation scenarios. Code: https://github.com/YShi-cool/DRFusion.

Automated 3D scene generation is pivotal for applications spanning virtual reality, digital content creation, and Embodied AI. While computer graphics prioritizes aesthetic layouts, vision and robotics demand scenes that mirror real-world complexity which current data-driven methods struggle to achieve due to limited unstructured training data and insufficient spatial and physical modeling. We propose SPREAD, a diffusion-based framework that jointly learns spatial and physical relationships through a graph transformer, explicitly conditioning on posed scene point clouds for geometric awareness. Moreover, our model integrates differentiable guidance for collision avoidance, relational constraint, and gravity, ensuring physically coherent scenes without sacrificing relational context. Our experiments on 3D-FRONT and ProcTHOR datasets demonstrate state-of-the-art performance in spatial-relational reasoning and physical metrics. Moreover, our method significantly outperforms baselines in scene consistency and stability during pre- and post-physics simulation, proving its capability to generate simulation-ready environments for embodied AI agents.

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.

This paper tackles the problem of physics-aware human motion synthesis in a dynamic scene. Unlike existing works which mainly tend to generate physically unrealistic motions due to limited contact modeling, typically restricted to hands, in this paper, we introduce a physics-aware human motion generation framework that explicitly models the full spectrum of human-related forces, including human-object, human-scene, and internal body dynamics. Our method imposes soft physical constraints to maintain force and torque balance, ensuring physically grounded motion synthesis. We further propose a novel continuous distance-based force model that generalizes contact modeling to arbitrary surfaces, capturing interactions not only with static environments but also with dynamic, moving objects. Extensive experiments show that our approach significantly improves physical plausibility and generalizes well to complex scenes, setting a new benchmark for physically consistent human motion generation.

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

Weakly-Supervised Dense Video Captioning aims to localize and describe events in videos trained only on caption annotations, without temporal boundaries. Prior work introduced an implicit supervision paradigm based on Gaussian masking and complementary captioning. However, existing method focus merely on generating non-overlapping masks without considering their semantic relationship to corresponding events, resulting in simplistic, uniformly distributed masks that fail to capture semantically meaningful regions. Moreover, relying solely on ground-truth captions leads to sub-optimal performance due to the inherent sparsity of existing datasets. In this work, we propose SAIL, which constructs semantically-aware masks through cross-modal alignment. Our similarity-aware training objective guides masks to emphasize video regions with high similarity to their corresponding event captions. Furthermore, to guide more accurate mask generation under sparse annotation settings, we introduce an LLM-based augmentation strategy that generates synthetic captions to provide additional alignment signals. These synthetic captions are incorporated through an inter-mask mechanism, providing auxiliary guidance for precise temporal localization without degrading the main objective. Experiments on ActivityNet Captions and YouCook2 demonstrate state-of-the-art performance on both captioning and localization metrics.

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

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 decoupled 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.

Large vision-language models (LVLMs) have demonstrated remarkable capabilities in multimodal understanding tasks. However, the increasing demand for high-resolution image and long-video understanding results in substantial token counts, consequently leading to reduced inference efficiency. Token compression offers a direct solution by reducing the number of tokens to be processed, thereby improving computational efficiency without architectural changes. Through extensive analysis, we identify two critical limitations in existing inner-LLM token compression methods: positional bias and incompatibility with efficient operators, which critically hinder their practical deployment for LVLM acceleration. This paper presents the first approach from a dynamic token variation perspective, revealing that visual token variations within LLMs exhibit task-agnostic properties. We propose Variation-aware Vision Token Dropping (i.e., V2Drop), which progressively removes visual tokens with minimal variation during LVLM inference, thereby enhancing computational efficiency. Extensive experiments across multiple models and benchmarks consistently demonstrate that V2Drop maintains 94.0% and 98.6% of the original performance for image and video understanding tasks respectively, while reducing LLM generation latency by 31.5% and 74.2%.

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

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.

Recent photo-realistic 3D talking head via 3D Gaussian Splatting still has significant shortcoming in emotional expression manipulation, especially for fine-grained and expansive dynamics emotional editing using multi-modal control. This paper introduces a new editable 3D Gaussian talking head, i.e. EmoDiffTalk. Our key idea is a novel Emotion-aware Gaussian Diffusion, which includes an action unit (AU) prompt Gaussian diffusion process for fine-grained facial animator, and moreover an accurate text-to-AU emotion controller to provide accurate and expansive dynamic emotional editing using text input. Experiments on public EmoTalk3D and RenderMe-360 datasets demonstrate superior emotional subtlety, lip-sync fidelity, and controllability of our EmoDiffTalk over previous works, establishing a principled pathway toward high-quality, diffusion-driven, multimodal editable 3D talking-head synthesis. To our best knowledge, our EmoDiffTalk is one of the first few 3D Gaussian Splatting talking-head generation framework, especially supporting continuous, multimodal emotional editing within the AU-based expression space.Please visit our website for more details and information:https://liuchang883.github.io/EmoDiffTalk/

Multimodal Continual Instruction Tuning aims to continually enhance Large Vision Language Models (LVLMs) by learning from new data without forgetting previously acquired knowledge. Mixture of Experts (MoE) architectures naturally facilitate this by incrementally adding new experts and expanding routers while keeping the existing ones frozen. However, despite expert isolation, MoE-based continual learners still suffer from forgetting due to routing-drift: old-task tokens become mistakenly attracted to newly added experts, degrading performance on prior tasks. We analyze the failure mode at the token level and reveal the token's dilemma: ambiguous and old tokens in new-task data offer minimal learning benefit yet induce forgetting when routed to new experts, due to their ambiguous routing assignment during training. Motivated by this, we propose LLaVA-DyMoE, a dynamic MoE framework that incrementally expands the MoE with drift-aware token assignment. We characterize token types via their routing score distributions and apply targeted regularization. Specifically, a token-level assignment guidance steers ambiguous and old tokens away from new experts to preserve established routing patterns and alleviate routing-drift, while complementary routing score regularizations enforce expert-group separation and promote new-expert specialization. Extensive experiments demonstrate that our LLaVA-DyMoE effectively mitigates routing-drift-induced forgetting, achieving over a 7% gain in mean final accuracy and a 12% reduction in forgetting compared to baselines. The project page is zhaoc5.github.io/DyMoE.