CVPR 2026 论文列表

Diffusion models achieve state-of-the-art video generation quality, but their inference remains expensive due to the large number of sequential denoising steps. This has motivated a growing line of research on accelerating diffusion inference. Among training-free acceleration methods, caching reduces computation by reusing previously computed model outputs across timesteps. Existing caching methods rely on heuristic criteria to choose cache/reuse timesteps and require extensive tuning. We address this limitation with a principled sensitivity-aware caching framework. Specifically, we formalize the caching error through an analysis of the model output sensitivity to perturbations in the denoising inputs, i.e., the noisy latent and the timestep, and show that this sensitivity is a key predictor of caching error. Based on this analysis, we propose Sensitivity-Aware Caching (SenCache), a dynamic caching policy that adaptively selects caching timesteps on a per-sample basis. Our framework provides a theoretical basis for adaptive caching, explains why prior empirical heuristics can be partially effective, and extends them to a dynamic, sample-specific approach. Experiments on Wan 2.1, CogVideoX, and LTX-Video show that SenCache achieves better visual quality than existing caching methods under similar computational budgets.

The Visual Geometry Grounded Transformer (VGGT) marks a significant leap forward in 3D scene reconstruction, as it is the first model that directly infers all key 3D attributes (camera poses, depths, and dense geometry) jointly in one pass. However, this joint inference mechanism requires global attention layers that perform all-to-all attention computation on tokens from all views. For reconstruction of large scenes with long-sequence inputs, this causes a significant latency bottleneck. In this paper, we propose head-wise temporal merging (HTTM), a training-free 3D token merging method for accelerating VGGT. Existing merging techniques merge tokens uniformly across different attention heads, resulting in identical tokens in the layers' output, which hinders the model's representational ability. HTTM tackles this problem by merging tokens in multi-head granularity, which preserves the uniqueness of feature tokens after head concatenation. Additionally, this enables HTTM to leverage the spatial locality and temporal correspondence observed at the head level to achieve higher merging ratios with lower merging costs compared to existing methods. Thus, HTTM achieves up to 7x acceleration with negligible performance drops in a GPU-based inference.

The introduction of negative labels (NLs) has proven effective in enhancing Out-of-Distribution (OOD) detection. However, existing methods often lack an understanding of OOD images, making it difficult to construct an accurate negative space. Furthermore, the absence of negative labels semantically similar to ID labels constrains their capability in near-OOD detection. To address these issues, we propose shaping an Adaptive Negative Textual Space (ANTS) by leveraging the understanding and reasoning capabilities of multimodal large language models (MLLMs). Specifically, we cache images likely to be OOD samples from the historical test images and prompt the MLLM to describe these images, generating expressive negative sentences that precisely characterize the OOD distribution and enhance far-OOD detection. For the near-OOD setting, where OOD samples resemble the in-distribution (ID) subset, we cache the subset of ID classes that are visually similar to historical test images and then leverage MLLM reasoning to generate visually similar negative labels tailored to this subset, effectively reducing false negatives and improving near-OOD detection. To balance these two types of negative textual spaces, we design an adaptive weighted score that enables the method to handle different OOD task settings (near-OOD and far-OOD), making it highly adaptable in open environments. On the ImageNet benchmark, our ANTS significantly reduces the FPR95 by 3.1%, establishing a new state-of-the-art. Furthermore, our method is training-free and zero-shot, enabling high scalability. Codes are available at https://github.com/ZhuWenjie98/ANTS.

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 diversity, we propose a pose-guided enriched feature learning method that explicitly augments pose-diverse samples. Specifically, a Pose-Extraction Module (PEM) disentangles pose-relevant and pose-irrelevant feature components, where pose-relationship knowledge distillation method helps identify the correct pose and semantic consistency maintenance 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 method synthesizes pose-swapped features, thereby substantially enhances contrastive learning of ReID models. Extensive experiments on Market1501 and MSMT17 under the federated-by-camera setting demonstrate that the proposed method consistently outperforms federated ReID baselines and their conjunctions with the existing feature augmentation methods; thus achieving state-of-the-art federated ReID performance.

In this work, we propose DiT360, a DiT-based framework that performs hybrid training on perspective and panoramic data for panoramic image generation. We attribute the main challenges in preserving geometric fidelity and photorealism to the scarcity of large-scale, high-quality real-world panoramic data, in contrast to prior methods that emphasize model design. Basically, DiT360 has several key modules for inter-domain transformation and intra-domain augmentation, applied at both the pre-VAE image level and the post-VAE token level. At the image level, we incorporate cross-domain knowledge through perspective image guidance and panoramic refinement, which enhance perceptual quality while regularizing diversity and photorealism. At the token level, hybrid supervision is applied across multiple modules, which include circular padding for boundary continuity, yaw loss for rotational robustness, and cube loss for distortion awareness. Extensive experiments on text-to-panorama, inpainting, and outpainting tasks demonstrate that our method achieves better boundary consistency and image fidelity across eleven quantitative metrics. Our code is available at https://github.com/Insta360-Research-Team/DiT360.

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 \color ddtD ecoupled \color ddtD iffusion \color ddtT ransformer(\color ddtDDT ), 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 256x256, Our DDT-XL/2 achieves a new state-of-the-art performance of 1.31 FID (nearly 4xfaster training convergence compared to previous diffusion transformers). For ImageNet 512x512, 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.

Creating interactive digital environments for gaming, robotics, and simulation relies on articulated 3D objects whose functionality emerges from their part geometry and kinematic structure. However, existing approaches remain fundamentally limited: optimization-based reconstruction methods require slow, per-object joint fitting and typically handle only simple, single-joint objects, while retrieval-based methods assemble parts from a fixed library, leading to repetitive geometry and poor generalization. To address these challenges, we introduce ArtLLM, a novel framework for generating high-quality articulated assets directly from complete 3D meshes. At its core is a 3D multimodal large language model trained on a large-scale articulation dataset curated from both existing articulation datasets and procedurally generated objects. Unlike prior work, ArtLLM autoregressively predicts a variable number of parts and joints, inferring their kinematic structure in a unified manner from the object's point cloud. This articulation-aware layout then conditions a 3D generative model to synthesize high-fidelity part geometries. Experiments on the PartNet-Mobility dataset show that ArtLLM significantly outperforms state-of-the-art methods in both part layout accuracy and joint prediction, while generalizing robustly to real-world objects. Finally, we demonstrate its utility in constructing digital twins, highlighting its potential for scalable robot learning.

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

Reconstructing 3D human-object interaction (HOI) from monocular images is highly challenging especially when human and object are mutually occluded. Existing methods primarily rely on single-view inputs, which fundamentally limit their ability to recover occluded regions and accurately estimate contact areas. To address these challenges, we for the first time, consider to introduce novel-view feature priors to enhance monocular 3D HOI reconstruction. We first design a cross-view generator that learns to infer novel-view image features from a single-view input, enriching spatial geometry at the feature level without requiring extra inputs during inference. Guided by both real and generated view features, a spatial cross-view feature fusion module adaptively aggregates complementary cues to enhance the initial reconstruction of human and object meshes. Built upon this reconstruction, we sample 3D vertex features from both views and introduce a bidirectional cross-view Transformer to integrate multi-view vertex representations for accurate contact estimation. Finally, the predicted contact maps are leveraged to refine human-object meshes, yielding geometrically consistent and physically plausible reconstructions.Experiments on BEHAVE and InterCap show that our proposed CrossHOI surpasses state-of-the-art methods in both reconstruction accuracy and contact prediction, especially under severe occlusions.

Face Anti-Spoofing (FAS) typically depends on a single visual modality when defending against presentation attacks such as print attacks, screen replays, and 3D masks, resulting in limited generalization across devices, environments, and attack types. Meanwhile, Multimodal Large Language Models (MLLMs) have recently achieved breakthroughs in image-text understanding and semantic reasoning, suggesting that integrating visual and linguistic co-inference into FAS can substantially improve both robustness and interpretability. However, the lack of a high-quality vision-language multimodal dataset has been a critical bottleneck. To address this, we introduce FaceCoT (Face Chain-of-Thought), the first large-scale Visual Question Answering (VQA) dataset tailored for FAS. FaceCoT covers 14 spoofing attack types and enriches model learning with high-quality CoT VQA annotations. Meanwhile, we develop a caption model refined via reinforcement learning to expand the dataset and enhance annotation quality. Furthermore, we introduce a CoT-Enhanced Progressive Learning (CEPL) strategy to better leverage the CoT data and boost model performance on FAS tasks. Extensive experiments demonstrate that models trained with FaceCoT and CEPL outperform state-of-the-art methods on multiple benchmark datasets.

Humanoid robots have achieved significant progress in motion generation and control, exhibiting movements that appear increasingly natural and human-like. Inspired by the Turing Test, we propose the Motion Turing Test, a framework that evaluates whether human observers can discriminate between humanoid robot and human poses using only kinematic information. To facilitate this evaluation, we present the Human-Humanoid Motion (HHMotion) dataset, which consists of 1,000 motion sequences spanning 15 action categories, performed by 11 humanoid models and 10 human subjects. All motion sequences are converted into SMPL-X representations to eliminate the influence of visual appearance. We recruited 30 annotators to rate the human-likeness of each pose on a 0-5 scale, resulting in over 500 hours of annotation. Analysis of the collected data reveals that humanoid motions still exhibit noticeable deviations from human movements, particularly in dynamic actions such as jumping, boxing, and running. Building on HHMotion, we formulate a human-likeness evaluation task that aims to automatically predict human-likeness scores from motion data. Despite recent progress in multimodal large language models, we find that they remain inadequate for assessing motion human-likeness. To address this, we propose a simple baseline model and demonstrate that it outperforms several contemporary LLM-based methods. The dataset, code, and benchmark release publicly at http://www.lidarhumanmotion.net/mtt/

Generative video models achieve high visual fidelity but often violate basic physical principles, limiting reliability in real-world settings. Prior attempts to inject physics rely on conditioning: frame-level signals are domain-specific and short-horizon, while global text prompts are coarse and noisy, missing fine-grained dynamics. We present PhysVid, a physics-aware local conditioning scheme that operates over temporally contiguous chunks of frames. Each chunk is annotated with physics-grounded descriptions of states, interactions, and constraints, which are fused with the global prompt via chunk-aware cross-attention during training. At inference, we introduce negative physics prompts (descriptions of locally relevant law violations) to steer generation away from implausible trajectories. On VideoPhy, PhysVid improves physical commonsense scores by ~ 33% over baseline video generators, and by up to ~ 8% on VideoPhy2. These results show that local, physics-aware guidance substantially increases physical plausibility in generative video and marks a step toward physics-grounded video models.

Inference-time optimization of diffusion latents enables powerful control but often degrades the statistical structure of true Gaussian noise, causing artifacts and reward hacking. To address this, we propose a Gaussianity regularizer that aligns a sample's local statistics with a typical Gaussian realization, rather than relying on pointwise likelihood. We formalize this by computing the KL divergence between the sample distribution and the Gaussian prior. To make the divergence computation tractable from a single sample, we lift each candidate latent into an empirical distribution induced by its statistics and model it as a pairwise Markov Random Field. This yields a Bethe-Kikuchi-style regularizer with 1D marginal, 2D spatial, and multi-scale terms. Our results show improved latent optimization stability and generation quality over prior approaches.

Estimating the 3D pose of unseen objects from a single image remains a fundamental yet challenging problem in computer vision, especially under a CAD model-free setting.Pioneering attempts address this issue by matching templates generated through Novel View Synthesis (NVS), which essentially aims to learn the geometric transformation from a reference to a target view. While promising, these methods can only approximate this transformation under pixel-level supervision, as the starting orientation remains undefined. In the absence of explicit geometric constraints to verify the correctness of the predicted transformation, existing methods often synthesize novel views with geometry-distorted structures or severely blurred local textures, leading to unreliable template matching and suboptimal pose estimation results. To this end, we propose OrienPose, a novel object pose estimation framework via orientation-aware NVS from a single image. Specifically, we introduce the Orientation-Aware Guidance, which explicitly injects object orientation cues into the reference latent embedding to enhance orientation awareness during viewpoint transformation. We also introduce an orientation consistency loss that supervises viewpoint transformation at the geometric level, establishing sufficient supervision for explicit and geometry-consistent transformation guidance beyond pixel-level similarity. This loss justifies estimating the reference orientation rather than using its ground-truth pose, thereby ensuring the alignment of coordinate domains between the injected and supervised priors. Extensive experiments demonstrate that OrienPose achieves state-of-the-art performance in single-view unseen object pose estimation and impressive robustness to image degradations. Our code is available at https://github.com/pubyLu/OrienPose.

Visual query localization (VQL) aims to predict a spatial-temporal response of the most recent occurrence from a sequence given a query. Currently, most research focuses on visual query localization from 2D videos, while its counterpart in 3D space has received little attention. In this paper, we make the first attempt to visual query localization in the 3D world by introducing a novel benchmark, dubbed 3DVQL. Specifically, 3DVQL contains 2,002 sequences with around 170,000 frames and 6.4K response track segments from 38 object categories. Each sequence in 3DVQL is provided with multiple modalities including point clouds (PC), RGB and depth images to support flexible research. To ensure high-quality annotation, each sequence is manually annotated with multiple rounds of verification and refinement. To our best knowledge, 3DVQL is the first benchmark towards 3D multimodal visual query localization. To facilitate comparison for subsequent research, we implement a series of representative 3D multimodal VQL baselines using PC and RGB. The experimental results show that existing methods exhibit significant performance variations across different fusion modules. To encourage future research, we propose a lift and attention fusion algorithm named LaF, which significantly outperforms existing baseline models. Our benchmark and model will be publicly released at our webpage https://github.com/wuhengliangliang/3DVQL.

Existing single object trackers typically treat temporal modeling superficially by passing limited inter-frame information, such as propagated tokens or template updates, without intrinsic temporal supervision learning. To address this limitation, we propose TGTrack, a new unified tracking framework that incorporates a temporally generative supervision task to guide the model in learning temporal dynamics. The core of TGTrack is a temporally generative learning paradigm equipped with a transformer-based generative decoder, which consists of a gated fusion module and an autoregressive prediction mechanism. This joint design enables the model to infer future scenarios from preceding information, thereby improving its ability to model both visual appearance and temporal dynamics. Furthermore, we introduce a time token embedding to explicitly encode the temporal position of each frame. Experiments on 11 benchmarks spanning five modalities show that TGTrack achieves state-of-the-art performance in robust unified tracking. For instance, TGTrack-B384 achieves an AUC of 75.3% on LaSOT. Code is available at https://github.com/wtg1/TGTrack.

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.

Open-vocabulary semantic segmentation (OVSS) aims to segment arbitrary category regions in images using open-vocabulary prompts, necessitating that existing methods possess pixel-level vision-language alignment capability. Typically, this capability involves computing the cosine similarity, ie, logits, between visual and linguistic features, and minimizing the distribution discrepancy between the logits and the ground truth (GT) to generate optimal logits that are subsequently used to construct segmentation maps, yet it depends on time-consuming iterative training or model-specific attention modulation. In this work, we propose a more direct approach that eschews the logits-optimization process by directly deriving an analytic solution for the segmentation map. We posit a key hypothesis: the distribution discrepancy encodes semantic information; specifically, this discrepancy exhibits consistency across patches belonging to the same category but inconsistency across different categories. Based on this hypothesis, we directly utilize the analytic solution of this distribution discrepancy as the semantic maps. In other words, we reformulate the optimization of the distribution discrepancy as deriving its analytic solution, thereby eliminating time-consuming iterative training, freeing us from model-specific attention modulation, and achieving state-of-the-art performance on eight benchmark datasets.

The rapid development of large vision-language model (VLM) has greatly promoted the research of GUI agent. However, GUI agents still face significant challenges in handling long-horizon tasks. First, single-agent models struggle to balance high-level capabilities and low-level execution capability, facing prevalent issues of responsibility coupling and capability conflicts. Second, agents lack awareness of the task state, leading to progress loss in long-horizon tasks. To address these challenges, we propose a staged execution-feedback reinforcement learning algorithm. Unlike training a unified policy model, we focus on training high-level scheduling models. Specifically, we propose and train two agents: a Coordinator, responsible for the strategic planning and task decomposition; and a State Tracker, responsible for context compression and information management to maintain the task's state and coherence. Based on this, we built the Coordinator-Executor-State Tracker (CES) multi-agent framework, which can be integrated with any low-level Executor model, assisting the Executor in solving long-horizon tasks through task scheduling and state management. Experiments on long-horizon task benchmarks demonstrate that CES significantly enhances the system's planning and state management capabilities. Furthermore, analysis confirms that our trained high-level scheduling module is a generalizable, plug-and-play module that significantly enhances the long-horizon capabilities of various Executors.