CVPR 2026 Papers

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 2.65x 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.

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

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.

Diffusion Transformers (DiTs) achieve strong video generation performance but suffer from prohibitive computation cost due to dense spatiotemporal tokenization. Most existing works rely on uniform patchification, tokenizing non-overlapping spatiotemporal with a fixed patch size regardless of the underlying content. This content-agnostic tokenization results in substantial redundant computation, especially in visually simple or static areas. To address this inefficiency while preserving the video generation quality, we propose DynaPatch, a fine-grained dynamic patchification framework that adaptively selects patch sizes for each spatiotemporal region based on content complexity. A lightweight router predicts patch sizes directly from the latents encoded by 3D Variational Autoencoder (VAE), and is jointly optimized with the diffusion model through diffusion loss, an attention-guided saliency alignment loss, and a token-budget regularizer. Learnable patchify/unpatchify layers integrate seamlessly with standard DiT backbones, allowing flexible tokenization without architectural changes. Experiments demonstrate that DynaPatch can effectively reduce redundant computations while preserving fine details, achieving 1.3-1.8x acceleration with minimal quality degradation. On VBench, DynaPatch attains a Total Score of 83.42 at 30% token reduction, significantly outperforming prior patchification and token pruning approaches. These results indicate that content-aware patchification offers an effective direction for efficient and scalable video diffusion. Project page: https://shengli99.github.io/DynaPatch/.

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

The miniaturization of thermal sensors for mobile platforms inherently limits their spatial resolution and textural fidelity, leading to blurry and less informative images. Existing thermal super-resolution (SR) methods can be grouped into single-image and RGB-guided approaches: the former struggles to recover fine structures from limited information, while the latter relies on accurate and laborious cross-camera calibration, which hinders practical deployment and robustness. Here, we propose 3M-TI, a calibration-free Multi-camera cross-Modality diffusion framework for Mobile Thermal Imaging. At its core, 3M-TI integrates a cross-modal self-attention module (CSM) into the diffusion UNet, replacing the original self-attention layers to adaptively align thermal and RGB features throughout the denoising process, without requiring explicit camera calibration. This design enables the diffusion network to leverage its generative prior to enhance spatial resolution, structural fidelity, and texture detail in the super-resolved thermal images. Extensive evaluations on real-world mobile thermal cameras and public benchmarks validate our superior performance, achieving state-of-the-art results in both visual quality and quantitative metrics. More importantly, the thermal images enhanced by 3M-TI lead to substantial gains in critical downstream tasks like object detection and segmentation, underscoring its practical value for robust mobile thermal perception systems.

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. Project code: https://github.com/jeremyJJB/Differentiable-Laplacian-Matrix-Guided-Superpixel-Segmentation.

In large-scale industrial documents with scanned images, complex layouts, and multiple pages, the effectiveness of retrieval-augmented generation (RAG) is highly dependent on chunking quality. However, existing text-centric chunkers overlook the visual and structural cues present in real-world documents, leading to redundant or ambiguous chunks that impair retrieval and answer accuracy. To address this problem, we propose \ours which integrates (i) SharedDet for normalizing document parsing and OCR outputs into a document-level frame, (ii) Multi-modal block embeddings with boundary-aware SoftROI, (iii) global document-tree reconstruction via biaffine scoring, and (iv) structure-aware dependency chunking that preserves boundaries and reduces redundancy. \ours achieves consistent gains across both Document Hierarchical Parsing (DHP) and corpus-level RAG evaluations, improving STEDS by +28.5--39.6%, retrieval nDCG by +1.1--15.3%, and QA ANLS by +4.5--15.3%. These results demonstrate that modeling document-level dependencies with Multi-modal, structure-aware chunking improves RAG performance on long, multi-page industrial documents.

Reconstructing dynamic visual experiences from brain activity provides a compelling avenue for exploring the neural mechanisms of human visual perception. While recent progress in fMRI-based image reconstruction has been notable, extending this success to video reconstruction remains a significant challenge. Current fMRI-to-video reconstruction approaches consistently encounter two major shortcomings: (i) inconsistent visual representations of salient objects across frames, leading to appearance mismatches; (ii) poor temporal coherence, resulting in motion misalignment or abrupt frame transitions.To address these limitations, we introduce SemVideo, a novel fMRI-to-video reconstruction framework guided by hierarchical semantic information. At the core of SemVideo is SemMiner, a hierarchical guidance module that constructs three levels of semantic cues from the original video stimulus: static anchor descriptions, motion-oriented narratives, and holistic summaries. Leveraging this semantic guidance, SemVideo comprises three key components: a Semantic Alignment Decoder that aligns fMRI signals with CLIP-style embeddings derived from SemMiner, a Motion Adaptation Decoder that reconstructs dynamic motion patterns using a novel tripartite attention fusion architecture, and a Conditional Video Render that leverages hierarchical semantic guidance for video reconstruction. Experiments conducted on the CC2017 and HCP datasets demonstrate that SemVideo achieves superior performance in both semantic alignment and temporal consistency, setting a new state-of-the-art in fMRI-to-video reconstruction.

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.

While spatial transcriptomics (ST) has advanced our understanding of gene expression in tissue context, its high experimental cost limits its large-scale application. Predicting ST from pathology images is a promising, cost-effective alternative, but existing methods struggle to capture complex cross-slide spatial relationships. To address the challenge, we propose SpaHGC, a multi-modal heterogeneous graph-based model that captures both intra-slice and inter-slice spot-spot relationships from histology images. It integrates local spatial context within the target slide and cross-slide similarities computed from image embeddings extracted by a pathology foundation model. These embeddings enable inter-slice knowledge transfer, and SpaHGC further incorporates Masked Graph Contrastive Learning to enhance feature representation and transfer spatial gene expression knowledge from reference to target slides, enabling it to model complex spatial dependencies and significantly improve prediction accuracy.We conducted comprehensive benchmarking on seven matched histology-ST datasets from different platforms, tissues, and cancer subtypes. The results demonstrate that SpaHGC significantly outperforms the existing nine state-of-the-art methods across all evaluation metrics. Additionally, the predictions are significantly enriched in multiple cancer-related pathways, thereby highlighting its strong biological relevance and application potential. Code and data are available at https://github.com/wenwenmin/SpaHGC.

Edge-based representations are fundamental cues for visual understanding, a principle rooted in early vision research and still central today. We extend this principle to vision-language alignment, showing that isolating and aligning structural cues across modalities can greatly benefit fine-tuning on long, detail-rich captions, with a specific focus on improving cross modal retrieval. We introduce StructXLIP, a fine-tuning alignment paradigm that extracts edge maps (e.g., Canny), treating them as proxies for the visual structure of an image, and filters the corresponding captions to emphasize structural cues, making them "structure-centric". Fine-tuning augments the standard alignment loss with three structure-centric losses: (i) aligning edge maps with structural text, (ii) matching local edge regions to textual chunks, and (iii) connecting edge maps to color images to prevent representation drift. From a theoretical standpoint, while standard CLIP maximizes the mutual information between visual and textual embeddings, StructXLIP additionally maximizes the mutual information between multimodal structural representations. This auxiliary optimization is intrinsically harder, guiding the model toward more robust and semantically stable minima, enhancing vision-language alignment. Beyond outperforming current competitors on cross-modal retrieval on both general and specialized domains, our method serves as a general boosting recipe that can be integrated into future approaches in a plug-and-play manner. Code and pretrained models are available at: https://github.com/intelligolabs/StructXLIP.

Video diffusion models achieve impressive visual fidelity but remain computationally prohibitive for real-time or interactive generation due to their sequential denoising process. Recent caching methods accelerate inference by reusing outputs across timesteps, typically estimating each new output from the first-order residual, which is the difference between adjacent model predictions.To mitigate the accumulated error in caching methods, we propose D2Cache, a training-free method that leverages the smoothness of second-order residual delta, which is temporal differences between consecutive first-order residuals, to predict future timesteps more accurately. We theoretically show that this second-order correction improves prediction accuracy and effectively suppresses cumulative errors. Moreover, D2Cache adaptively scales second-order deltas using error estimates derived from timestep embeddings, maintaining accuracy across varying cache intervals.Empirically, D2Cache outperforms the state-of-the-art TeaCache across four video diffusion models (Latte, Open-Sora, LTX-video, and Wan2.1) at comparable acceleration rates, showing even larger gains under higher acceleration settings.

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.

Incremental Object Detection (IOD) aims to equip detectors with the ability to handle dynamic environments and emerging object categories, and the rise of vision-language models has substantially advanced this goal. However, existing studies often oversimplify real-world scenarios by assuming the incremental tasks come from a single general domain. To better investigate vision-language models under IOD, it is necessary to explore more generalized scenarios that encompass both novel categories and domains. To this end, we propose Cross-Domain Incremental Object Detection (CDIOD), a new benchmark that assesses the ability to continuously adapt to diverse object detection tasks across domains. CDIOD reveals that existing methods struggle to balance between adaptivity and stability under substantial domain shifts. To tackle this challenge, we propose Dynamic Group Subspace (DGS), a novel framework that dynamically groups tasks by distribution to promote knowledge sharing and prevent task collisions; progressively consolidates adapters to build shared subspaces and control parameter growth; and implements a dynamic training pipeline to maintain a proper stability-adaptivity balance. DGS enables vision-language models to effectively handle task streams of various distribution shifts. Extensive experiments across three benchmarks demonstrate that DGS achieves SOTA performance, highlighting its robustness in diverse incremental learning scenarios. Code is available at https://github.com/Never-wx/dgs.

Source-free domain adaptation (SFDA) adapts a pre-trained source model to an unlabeled target domain using only the model itself, typically relying on pseudo labeling augmented with auxiliary knowledge and consistency regularization (CR) mechanisms to alleviate noise in the generated pseudo labels. However, existing approaches overlook the geometric structure of the target embedding manifold when assigning pseudo labels, resulting in unreliable distance measurements and consequently severe mislabeling. Moreover, existing CR is applied solely to output logits, making it insensitive to feature-level reliability. To solve these issues, we propose a novel pseudo labeling scheme based on Feature universe, which is an expanded embedding space that models class-wise target distributions and Gravity consistency (GV) regularization, which modulates consistency strength according to feature-level similarity. Our pseudo labeling strategy first models the embedding space with virtual features to construct a feature universe. On this space, pseudo labels are generated through feature traversal, which propagates labels only from statistically reliable regions. In addition, GV jointly encourages logit- and feature-level consistency, aligning predictions for augmented images while preserving the geometric structure of the embedding space. It further modulates the strength of CR for each sample, preventing the confirmation of noisy pseudo labels through a gravity-based force defined between two input embeddings. Experiments on Office-Home, DomainNet-126, and VisDA-C demonstrate consistent improvements over prior SFDA methods, and incorporating GV into baselines yields additional gains.

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.

Achieving accurate garment grasping under dynamically changing illumination is crucial for all-day operation of service robots. However, the reduced illumination in low-light scenes severely degrades garment structural features, leading to a significant drop in grasping robustness. Existing methods typically enhance RGB features by exploiting the illumination-invariant properties of non-RGB modalities, yet they overlook the varying dependence on non-RGB features under varying lighting conditions, which can introduce misaligned non-RGB cues and thereby weaken the model's adaptability to illumination changes when utilizing multimodal information. To address this problem, we propose GraspALL, an illumination-structure interactive compensation model. The innovation of GraspALL lies in encoding continuous illumination changes into quantitative references to guide adaptive feature fusion between RGB and non-RGB modalities according to varying lighting intensities, thereby generating illumination-consistent grasping representations. Experiments on the self-built garment grasping dataset demonstrate that GraspALL improves grasping accuracy by 32-44% over baselines under diverse illumination conditions. The code is available at https://github.com/Zhonghaifeng6/GraspALL

Controlling both camera motion and object dynamics is essential for coherent and expressive video generation, yet current methods typically handle only one motion type or rely on ambiguous 2D cues that entangle camera-induced parallax with true object movement. We present SymphoMotion, a unified motion-control framework that jointly governs camera trajectories and object dynamics within a single model. SymphoMotion features a Camera Trajectory Control mechanism that integrates explicit camera paths with geometry-aware cues to ensure stable, structurally consistent viewpoint transitions, and an Object Dynamics Control mechanism that combines 2D visual guidance with 3D trajectory embeddings to enable depth-aware, spatially coherent object manipulation. To support large-scale training and evaluation, we further construct RealCOD-25K, a comprehensive real-world dataset containing paired camera poses and object-level 3D trajectories across diverse indoor and outdoor scenes, addressing a key data gap in unified motion control. Extensive experiments and user studies show that SymphoMotion significantly outperforms existing methods in visual fidelity, camera controllability, and object-motion accuracy, establishing a new benchmark for unified motion control in video generation.