ECCV 2026 论文列表

Although existing video editing methods are generally feasi-ble, they often require many costly iterations and still struggle to deliverhigh-quality yet satisfying editing results. We attribute this limitationto the prevalent data-to-data paradigm, which is less compatible withmodern generative models than noise-to-data generation. To address thisgap, we revisit video editing from a noise-to-data perspective and pro-pose Streaming-Generation-based Video Editing (StreamEdit), whichpreserves few-step sampling while seamlessly injecting source-video con-ditions. Built on pre-trained streaming generation models, StreamEditintroduces dual-branch fast sampling with a self-attention bridge andcross-attention grounding/boosting to satisfy both sampling and con-ditioning requirements. We further propose source-oriented guidance toimprove target-generation quality, and a visual prompting strategy to en-hance editing flexibility and practicality. The method is effective, robust,and generalizable across different models. Extensive experiments on di-verse video editing tasks show that StreamEdit consistently outperformsexisting approaches, even in few-step settings with minimal time cost.Code and results are available at: dsl-lab.github.io/StreamEdit/.

Benefiting from the strong ability to capture data distributions, diffusion models have become powerful tools for solving image inverse problems. The key is to controllably steer the sampling trajectory toward the measurements while respecting the diffusion prior. In this work, we introduce Stochastic Optimal Control Sampling (SOCS), which models the denoising process as a dynamical system and injects control signals via SOC. Previous SOC-based approach addresses inverse problems by optimizing over the entire trajectory, which is computationally expensive. In contrast, we derive a closed-form control update and apply it at each sampling step, pulling the measurement-consistent clean prediction back onto the denoising flow. In SOCS, we can readily modulate the control strength to align with the diffusion model’s native capabilities and thereby enhance perceptual quality. Our method is compatible with a variety of linear stochastic differential equation backbones. Extensive experiments across a broad spectrum of image inverse tasks demonstrate that SOCS achieves accurate measurement-aligned reconstructions with improved visual fidelity and stronger quantitative performance. Code is available at https://github.com/zjqwq01/SOCS-DIP.

As Text-to-Video (T2V) models progress towards higher vi-sual realism, the artifacts and errors they produce are highly nuanced,fine-grained, and spatio-temporally localized. Vision Language Models(VLMs) are actively being adopted as automatic evaluators for videogeneration, driven by the promise of their perception and reasoning abili-ties. Yet, it remains unclear whether they can detect, localize, and explainfine-grained errors in modern high-fidelity video generations. We intro-duce Spotlight, a novel benchmark to rigorously assess whether cur-rent VLMs can precisely localize and explain nuanced video-generationerrors. Spotlight comprises 600 videos generated by state-of-the-artT2V models (Veo3, Seedance, LTX-2), annotated with over 1,600 fine-grained error localizations and explanations spanning physics, semanticsand anatomy. Our experiments reveal that current VLMs lag behindhumans significantly, with humans outperforming our best baselines bynearly 2× on the task. Our analysis highlights key areas of improvementfor utilizing VLMs as automated video evaluators, including the need formore robust perception and hallucination mitigation. Data and code isavailable at https://spotlight-video.github.io.

Agentic multimodal large language models (MLLMs) (e.g.,OpenAI o3 [41] and Gemini Agentic Vision [10]) achieve remarkablereasoning capabilities through the iterative invocation of visual tools.However, the cascaded perception, reasoning, and tool-calling loops intro-duce significant sequential overhead. This overhead, termed agentic depth,incurs prohibitive latency and seriously limits system-level concurrency.To this end, we propose SpecEyes, an agentic-level speculative accelera-tion framework that breaks this sequential bottleneck. Our key insight isthat a lightweight MLLM can plan a tool-free execution path to directlyanswer many queries, bypassing the expensive tool-use loop. To regulatethis speculative planning, we introduce a cognitive gating mechanismbased on answer separability, which quantifies the model’s confidence inself-verification without requiring oracle labels. Furthermore, we designa heterogeneous parallel funnel that exploits the small model’s statelessconcurrency to mask the large model’s stateful serial execution, therebymaximizing system throughput. Extensive experiments on V* Bench,HR-Bench, and POPE demonstrate that SpecEyes achieves 1.1 − 3.35×speedup over the baseline while preserving or even improving accuracy,thereby boosting serving throughput under concurrent workloads.

Vision–language models (VLMs) have recently shown promise for assisting clinical reasoning in dermatological diagnosis. However, their trustworthiness and clinical utility remain limited by three key challenges: heterogeneous datasets with inconsistent diagnostic labels and concept annotations, the lack of grounded diagnostic rationales for reliable reasoning supervision, and limited scalability when transferring knowledge from small, densely annotated datasets to large collections with sparse labels. To address these challenges, we propose Skin-R1, a dermatology-oriented VLM that integrates textbook-grounded clinical reasoning supervision with reinforcement learning (RL) to improve the accuracy and robustness of diagnostic prediction. First, we construct a textbook-based reasoning generator that synthesizes hierarchy-aware and differential-diagnosis (DDx) diagnostic trajectories derived from authoritative dermatology knowledge. Second, these trajectories are used for supervised fine-tuning (SFT), establishing a clinically grounded reasoning foundation for the model. Finally, we introduce an RL training framework that incorporates the hierarchical structure of dermatological diseases into the reward design, enabling the model to generalize grounded diagnostic reasoning to large-scale datasets with sparse annotations. Extensive experiments across multiple dermatology benchmarks demonstrate that Skin-R1 consistently improves diagnostic accuracy and robustness compared to state-of-the-art Med-VLM baselines. Ablation studies further highlight the critical role of grounded reasoning supervision introduced during the SFT stage. Code and model weight are available at https://github.com/l593191569/Skin-R1.

Inferring physical properties such as mass, stiffness, and elas-ticity from a single image is essential for simulation and embodied AI, yetmost existing approaches rely on multi-view reconstruction or physics-based supervision. We introduce SiPhy, a unified framework for single-image physical property reasoning that aligns 3D-aware visual cues, depthwith language-based material knowledge. From one RGB image, SiPhysamples pseudo-voxel points, extracts CLIP features, and grounds themto material candidates proposed by an VLM. A part-based contrastiveaggregator enforces region consistency, while a heaviness-aware refine-ment improves thickness and volume estimation for dense objects.Across ABO-500, MVImgNet-100, and PhysXNet-100, SiPhy achievesstate-of-the-art single-image performance, surpassing multi-view recon-struction methods by improving mass MnRE by up to 93% (vs. PUGS),reducing density MAE by 35.5% (vs. NeRF2Physics), and loweringYoung’s modulus error by 23.5%. We further validate SiPhy on realhand–object interaction datasets, demonstrating its potential as a dataannotation engine for physical understanding from single-view imagery.Codes are available at https://github.com/DominoAI-Lab/SiPhy-ECCV-2026.

Embodied Question Answering (EQA) requires an agent toexplore unseen environments, gather visual evidence, and answer natural-language questions. Efficient exploration requires both global planningand task-aware prioritization of informative regions. However, existingfrontier-based methods typically rely on greedy, step-wise decisions, whichoften cause redundant motion and frequent Vision-Language Model (VLM)queries. We propose SAFE-EQA, a semantic-aware framework that im-proves exploration efficiency through globally coherent path planning andreduced VLM interaction. SAFE-EQA incrementally builds a landmarkgraph to capture scene topology and a query-conditioned semantic mapto estimate task relevance of different regions. To support comprehen-sive, question-oriented exploration, we formulate planning as a Semantic-Aware Traveling Salesman Problem (TSP), which optimizes the visita-tion order over discovered landmarks. We further introduce an adaptivereplanning mechanism that selectively interrupts local navigation andre-optimizes trajectories when higher-priority frontiers emerge. Exten-sive experiments on OpenEQA, EXPRESS-Bench and HM-EQA showthat SAFE-EQA matches or outperforms state-of-the-art EQA methodswhile reducing VLM token consumption by 27.8–45.6%.

Pretrained vision models are widely used as backbones fordownstream tasks, yet adapting them to multi-task learning (MTL) re-mains challenging. Naive full fine-tuning incurs substantial overhead andoften exacerbates task interference, while existing parameter-efficientfine-tuning (PEFT) methods are largely designed for single-task adap-tation. We propose Residual Upcycling with Task Ladder (RUTaL), aparameter-efficient multi-task vision adaptation framework. To accom-modate the diverse and heterogeneous representation demands in MTL,We proposed Task-General Residual Upcycling (TGRU) to transforma pre-trained vision Transformer backbone into a Mixture-of-Expertsarchitecture through low-rank residual weight reparameterization, en-abling efficient and task-scalable capacity expansion. Built upon this up-cycled representation paradigm, we further introduce Task-Specific Lad-der Adaptation (TSLA), which extracts task-relevant features from theshared upcycled representations in a decoupled manner to accommodatethe unique requirements of each task. Experiments on multi-task densescene understanding benchmarks show that RUTaL achieves state-of-the-art performance while demonstrating superior computational efficiency.

In real-world deployment under strict real-time constraints,weather and imaging variations induce significant distribution shifts,severely degrading detectors. Single-Domain Generalized Object Detec-tion aims to mitigate this issue, yet existing methods rarely investi-gate—at the level of problem formulation—the generalization capabil-ity of real-time detectors under such constrained inference budgets. Tothis end, we introduce Real-Time Single-Domain Generalized ObjectDetection (RT-SDGOD), which focuses on how real-time detectors canachieve cross-domain generalization under zero extra inference overheadby relying solely on training-time representation learning. We observethat, under domain shift, DETR-based real-time detectors mainly de-grade through increased missed detections, rooted in limited and un-stable object-level discriminative evidence. Based on this, we proposeRT-SDGDet, a multi-evidence collaborative modeling framework for RT-SDGOD. The core idea is to enable multiple queries of the same ob-ject to collaboratively cover more sufficient discriminative evidence whilemaintaining the stability of such evidence modeling across views. Specifi-cally, we use one-to-many (O2M) supervision to construct stable object-specific query groups, and further design Discriminative Evidence Di-versity Learning (DEDL) and Dual-view Evidence Consistency Learning(DvECL) to expand object-level evidence coverage and improve evidencestability under appearance perturbations, respectively. Since all compo-nents are introduced only during training, our method incurs no ex-tra inference overhead. Extensive experiments show that the proposedmethod achieves better generalization performance than existing ap-proaches across multiple unseen target domains.

Recovering 3D human motion from sparse wearable sensorsis challenging due to the inherent noise, occlusion, and instability ofreal-world tracking signals. Existing approaches are restricted to fixedsensor configurations and, when built on diffusion models, are computa-tionally prohibitive for real-time deployment. We propose OmniPoser, auniversal framework that recovers full-body SMPL motion from a flex-ible combination of heterogeneous inputs, such as 3D keypoints, sparseIMU rotations, head-mounted 6-DoF poses, or any subset thereof, withina single model. OmniPoser introduces a cross-modal masking mechanismpaired with a Masked Conditioning Encoder, enabling one architectureto incorporate diverse sensor configurations through a masked design. Adecoupled dual-stream generative backbone generates complete motionvia Conditional Flow Matching, requiring only a single ODE integrationstep at inference. Across three settings and seven benchmarks, includingin-the-wild Nymeria and Ego-Exo4D dataset, OmniPoser performs onpar with state-of-the-art methods while running at over 700 FPS. Thecode will be open-sourced.

Out-of-Distribution (OOD) detection is essential for ensur-ing the robustness and reliability of object detection systems deployed insafety-critical applications. While prior research has mainly focused onuni-modal detectors or vision-language model (VLM) based classifiers,the potential of VLM-based object detectors in OOD scenarios remainsunderexplored. In this work, we take the first step toward building OODobject detection methods upon VLMs. We identify two key challenges:(i) learning discriminative features to separate in-distribution (ID) fromOOD instances, and (ii) designing scoring functions consistent with VLMprobabilistic outputs. Hence, we introduce Negative Label Guided Atten-tion and Scoring (NegAS). To address (i), we propose a negative labelguided attention module (NegA), where LLM-generated, visually-similarbut semantically-different negative labels are used to guide attention to-ward potential OOD background regions. To address (ii), we introducea novel sigmoid-based OOD scoring function (NegS) that leverages bothID and negative labels, producing strong responses for ID instances andsuppressed responses for OOD ones. Extensive experiments demonstratethat our approach improves OOD detection performance by a large mar-gin while maintaining competitive ID accuracy, e.g., reducing the FPR95by 11.4% on the COCO dataset and 25.5% on the OpenImages datasetcompared to the baseline model. While initially designed for dense VLMdetectors like YOLO-World, we successfully adapt NegAS to GroundingDINO, a query-based VLM transformer, achieving significant improve-ments, demonstrating the generalizability of our framework. The code isavailable at https://github.com/yjzzyj/NegAS.

Tracking multi-person 3D human meshes from in-the-wildvideos is a highly challenging problem due to complex interactions, fre-quent occlusions, and severe truncation inherent in unconstrained envi-ronments. While recent approaches have improved robustness againstthese issues, they largely overlook the critical challenge prevalent inreal-world footage: frequent shot changes. These abrupt transitions incamera viewpoints often cause existing methods to lose track of humanidentities and fail in reconstructing temporally coherent trajectories. Al-though several recent works have explored 3D human mesh tracking un-der shot changes, they are still limited to single-person scenarios, mak-ing them inadequate for real-world videos where multiple people interactand appear simultaneously. To address this limitation, we propose Multi-THuMBS (Multi-person Tracking of 3D Human Meshes Beyond VideoShots) that leverages a state-of-the-art 3D scene prior to reconstructthe two boundary frames in a single shared 3D space. Human meshesare then registered within the shared 3D space, maintaining per-personidentity and motion consistency across shot changes. Extensive experi-ments demonstrate that our approach yields significant improvements in3D human mesh recovery, camera pose estimation, and identity track-ing, thereby ensuring high-fidelity motion reconstruction with consistentidentity preservation across shots compared to previous state-of-the-artmethods.

Many real-world applications of spatial intelligence, such asrobotic control, autonomous driving, and automated assembly, requirespatial reasoning across multiple sequential steps. However, the extent towhich current Multimodal Large Language Models (MLLMs) possess thiscapability remains largely unexplored. Inspired by LEGO construction, arecreational activity that critically relies on multi-step spatial reasoning,we introduce LEGO-Puzzles: a benchmark designed to systematicallyevaluate the spatial reasoning capabilities of MLLMs from basic spatialunderstanding to multi-step planning. LEGO-Puzzles contains two tasksets. The Elementary set covers 11 visual question-answering (VQA)tasks with 1, 100 carefully curated samples to test elementary spatialreasoning skills that are cruical for LEGO assembly. The Planning setdirectly requires the model to generate a step-by-step plan for assemblinga target LEGO structure, where the tasks are organized into subsets withdifferent planning horizons ranging up to 8. Our evaluation of 29 state-of-the-art MLLMs shows that even the strongest models struggle withelementary reasoning tasks in LEGO construction, falling at least 20%behind human performance. The planning accuracy also quickly dropsto 0% as the number of planning steps increases, whereas our humanparticipants solve all the tasks perfectly. Switching the output formatfrom multiple choice to image generation degrades model performanceeven further, leading to zero accuracy even for planning 3 steps. Over-all, LEGO-Puzzles reveals critical limitations in current MLLMs’ spatialreasoning capabilities and highlights the need for substantial advances.

Analyzing fine-grained skill activities (e.g., sports, surgery) requires not only recognizing visual patterns but also performing stepby-step visual reasoning that leads to the final judgment. While recent advances in action quality assessment have achieved remarkable progress in evaluating performance, existing models remain black boxes, where they lack the ability to explicitly reveal the reasoning processes underlying their judgments. To address this limitation, we propose Latent Visual Diffusion Reasoning (LVDR), a novel framework that integrates keypoint-guided Monte Carlo Tree Search (MCTS) to model and visualize the latent visual reasoning process. LVDR not only produces more accurate skill assessments but also uncovers the critical visual reasoning sequences that contribute to the final evaluation. Extensive experiments across four datasets spanning diverse sports and surgical domains demonstrate that LVDR achieves competitive quantitative performance while providing interpretable visual reasoning trajectories leading to the final predictions. Source codes and models can be found through the following link: https://github.com/XiruiTeng/LVDR_Official.git.

Existing Stochastic 3D Human Motion Prediction models are fundamentally constrained by hard-coding the skeleton kinematics, severely limiting generalization, preventing cross-dataset training, and requiring complex data retargeting. We introduce EquiFusion, the first kinematics-agnostic model to solve this bottleneck, implementing a latent diffusion model with a permutation equivariant architecture. EquiFusion treats the kinematics’ connectivity as an explicit input parameter, ensuring its internal computations are inherently agnostic to joint ordering and graph structure. This novel design enables truly cross-dataset generalization to unseen kinematics and unlocks novel zero-shot directions, such as motion prediction from partial or occluded observations and targeted limb generation. EquiFusion achieves state-of-the-art results on major benchmarks, being up to 75% more compact than previous kinematics-specific methods, while achieving faster training and inference. EquiFusion thus establishes a new, flexible standard for robust human motion prediction.

Knowledge-Based Visual Question Answering (KB-VQA) aimsto evaluate whether Visual Language Models (VLMs) can retrieve, ground,and reason over external structured knowledge beyond visual evidence.In practice, answer accuracy is widely adopted as the primary evalu-ation metric, implicitly treating correctness as a proxy for knowledge-grounded reasoning. However, for existing KB-VQA benchmarks, thisproxy relies on critical assumptions that are often overlooked and ren-dered unreliable by benchmark issues: annotated answer must be deriv-able from the associated knowledge base, question must be well-posedwith su!cient constraints, and visual setting must meaningfully requiregrounded disambiguation. In this work, we show that these assumptionsare systematically violated in existing KB-VQA benchmarks. Our au-dit reveals substantial instances with missing or contradicted answersand underspecified questions that render accuracy a misleading metric.Furthermore, we find that existing datasets rely on visually trivial, single-entity scenes that bypass the need for sophisticated visual-to-knowledgemapping. We demonstrate that even with controlled architectures, theseflaws lead to distorted model rankings and overestimations of reasoningcapabilities. To address this, we introduce (1) a principled audit-and-repair protocol that restores answer derivability and question clarity,and (2) a controlled multi-entity augmentation protocol that introducesvisual ambiguity to challenge initial retrieval and grounded reasoning.Re-evaluation under corrected and augmented settings yields markedlydi"erent performance trends. Our findings call for rethinking evaluationprotocols and designing more interaction-aware KB-VQA benchmarksthat prioritize verifiable reasoning over simple matching. 1

The physical world is not merely visual; it is governed byrigorous structural and procedural constraints. Yet, the evaluation ofvision-language models (VLMs) remains heavily skewed toward percep-tual realism, prioritizing the generation of visually plausible 3D lay-outs, shapes, and appearances. Current benchmarks rarely test whethermodels grasp the step-by-step processes and physical dependencies re-quired to actually build these artifacts—a capability essential for au-tomating design-to-construction pipelines. To address this, we introduceDreamHouse, a novel benchmark for physical generative reasoning:the capacity to synthesize artifacts that concurrently satisfy geomet-ric, structural, constructability, and code-compliance constraints. Weground this benchmark in residential timber-frame construction, a do-main with fully codified engineering standards and objectively verifiablecorrectness. We curate over 26,000 structures spanning 13 architecturalstyles—each verified to construction-document standards (LOD 350)—and develop a deterministic 10-test structural validation framework. Un-like static benchmarks that assess only final outputs, DreamHouse sup-ports iterative agentic interaction. Models observe intermediate buildstates, generate construction actions, and receive structured environ-mental feedback, enabling a fine-grained evaluation of planning, struc-tural reasoning, and self-correction. Extensive experiments with state-of-the-art VLMs reveal substantial capability gaps that are largely invis-ible on existing leaderboards. These findings establish physical validityas a critical evaluation axis orthogonal to visual realism, highlightingphysical generative reasoning as a distinct and underdeveloped fron-tier in multimodal intelligence. The benchmark is available at https://luluyuyuyang.github.io/dreamhouse/.

We present HairWeaver, a diffusion-based pipeline that ani-mates a single human image with realistic and expressive hair dynamics.While existing methods successfully control body pose, they lack specificcontrol over hair, and as a result, fail to capture the intricate hair mo-tions, resulting in stiff and unrealistic animations. HairWeaver overcomesthis limitation using two specialized modules: a Motion-Context-LoRAto integrate motion conditions and a Style-Alignment-LoRA to pre-serve the subject’s photoreal appearance across different data domains.These lightweight components are designed to guide a video diffusionbackbone while maintaining its core generative capabilities. By trainingon a specialized dataset of dynamic human motion generated from aCG simulator, HairWeaver affords fine control over hair motion and ulti-mately learns to produce highly realistic hair that responds naturally tomovement. Comprehensive evaluations demonstrate that our approachsets a new state of the art, producing lifelike human hair animationswith dynamic details.

End-to-end planning has emerged as a dominant paradigmfor autonomous driving, where recent models often adopt a scoring-selection framework to choose trajectories from a large set of candidates,with diffusion-based decoding showing strong promise. However, directlyselecting from the entire candidate space remains difficult to optimize,and Gaussian perturbations used in diffusion often introduce unrealis-tic trajectories that complicate the denoising process. In addition, fortraining these models, reinforcement learning (RL) has shown promise,but existing end-to-end RL approaches typically rely on a single coupledreward without structured signals, limiting optimization effectiveness.To address these challenges, we propose HAD, an end-to-end planningframework with a Hierarchical Diffusion Policy that decomposes planninginto a coarse-to-fine process. To improve trajectory generation, we intro-duce Structure-Preserved Trajectory Expansion, which produces realisticcandidates while maintaining kinematic structure. For policy learning, wedevelop Metric-Decoupled Policy Optimization (MDPO) to enable struc-tured RL optimization across multiple driving objectives. Extensive ex-periments show that HAD achieves new state-of-the-art performance onboth NAVSIM and HUGSIM, outperforming prior arts by a huge margin:+2.3 EPDMS on NAVSIM and +4.9 Route Completion on HUGSIM.