ECCV 2026 Papers

Achieving reliable single-shot high dynamic range (HDR)imaging under extreme illumination conditions remains a long-standingchallenge, yet no comprehensive benchmark exist for evaluating HDRperception in multi-sensor robotic systems. To fill this gap, we intro-duce a large-scale dataset collected via a custom robotic vision platformand an iPhone 13 Pro: 121 real-world scenes spanning modest and ultra-high dynamic range conditions, alongside 20 synthetic video sequencesfrom the CARLA simulator. As a reference pipeline for this dataset,we propose Depth-guided Multi-view Exposure Bracketing (DMEB), asingle-shot HDR method that distributes drastically different exposuresacross multi-view low-bit-depth cameras and fuses them via depth-guidedconfidence-aware fusion. Evaluations on our dataset show that DMEBestablishes a strong reference point and highlight the promise of thissensor configuration for robust HDR perception in diverse multi-cameraand depth sensor system.

Food cutting is a representative non-rigid manipulation taskinvolving deformation, contact forces, and material separation, yet itRobot Simulation remains largely unexplored in VLA research. Existing robot manipula-tion datasets are primarily built around rigid objects and struggle tojointly capture the spatial precision, deformation, topology changes, andCut the banana force interactions required for cutting. To systematically study theseat the center. challenges, we introduce CulinaryCut, a benchmark that integrates anniskill MPM-based deformable simulator with a robot environment to generatedataset rather than provided as policy inputs, enabling evaluation of ex-isting VLAs while exposing where physical grounding is needed. Usingthree representative VLA baselines, we identify two core limitations innon-rigid cutting. First, a geometry gap: VLAs struggle to map ratio-and direction-based instructions onto object geometry, especially whensequential cuts change the object’s topology. Second, a physics gap: be-cause policies output motion without a notion of material stiffness orcontact resistance, a geometrically accurate path may still fail to severthe object. We show that training VLAs on physics-grounded trajecto-ries, without modifying the policy architecture or adding force inputs,improves cutting success and sim-to-real transfer. Together, CulinaryCutprovides a foundation for evaluating spatial reasoning and physical plau-sibility in deformable manipulation.

Recent 3D Gaussian Splatting (3DGS) GANs for humanheads synthesize and render photorealistic 3D models in real-time andoffer a vast variety in identity and appearance. However, controlling spe-cific semantic attributes such as hair color or glasses remains challenging,as edits in the entangled latent space often induce unintended changesin identity or appearance. Although there are several methods that aimto disentangle the latent space post training by estimating directionsthat only modify certain features, these methods cannot guarantee com-plete disentanglement and often require pre-trained classifiers. In ourapproach, we propose a new generator architecture that synthesizes com-ponents, such as hair, skin, glasses, and torso, completely independently.This allows for changing the latent vector for one region while keepingthe remaining parts fixed. Further, we achieve this separation using onlysparse information such as the hair or skin color, eliminating the re-quirement of segmentation masks or geometric priors, often seen in priorwork. To ensure matching shape and lighting conditions during editing,we allow minimal shared information via context tokens between the in-dependent generators. These tokens even allow us to control the shapeand light, without any prior annotation. Compared to existing works onGAN-based generation and editing, our method shows better disentan-glement, more precise editing control, and competitive visual quality.

Recent multi-view 3D human pose estimation methods de-compose the problem into per-image 2D pose detection and 3D lift-ing stages, which can naturally scale to multi-frame input and achievestrong results by leveraging temporal context. However, this 2D-3D lift-ing paradigm introduces two levels of geometric inconsistency: (i) projec-tion inconsistency, i.e., the detected 2D keypoints may disagree with the2D projections of the inferred 3D pose, and (ii) cross-view inconsistency,i.e., 2D poses across views are mutually inconsistent because off-the-shelfdetectors operate independently per camera, so there may exist no sin-gle 3D pose whose projections match all detections. We observe that 2Ddetection errors closely follow a Gaussian distribution, meaning that de-tector outputs can be naturally viewed as noisy samples from a diffusionforward process. Based on this insight, we propose CoDePose, a coupled2D-3D denoising diffusion framework that jointly performs multi-viewdenoising of the detected 2D poses and estimates the underlying 3Dpose. As training proceeds, the generated joint distribution progressivelyapproaches the real data distribution, leading to increased projection andcross-view consistency and more reliable 3D estimation. Unlike existingdiffusion-based methods that condition on fixed 2D detections, CoDe-Pose refines the 2D observations during the reverse process, preventing2D detection errors and cross-view inconsistencies from being reinforcedthroughout the whole 3D pose generation. Experimental results on Hu-man3.6M show that CoDePose achieves 12.1 mm MPJPE, a 19.9% im-provement over the previous state-of-the-art method. Cross-dataset eval-uation on HumanEva and MPI-INF-3DHP further demonstrates stronggeneralization.

Document comprehension is a challenging yet impactful taskfor Multimodal Large Language Models, especially as these systems seegrowing adoption in real-world, human-centric applications. However,this adoption is limited for low-resource languages such as Bangla dueto the scarcity of high-quality annotated data. To address this gap, weintroduce BaFCo, a benchmark dataset for Bangla form comprehensionwith a focus on Document Layout Analysis (DLA) and Key InformationExtraction (KIE). BaFCo curates 200 multi-page complex Bangladeshigovernment forms, sourced from across diverse sectors including agricul-ture, education, banking, and land management. To accurately capturethe structural and contextual complexity of these forms, we define a fine-grained annotation schema comprising 26 types of form entities, alongwith a separate coarse form entity set consisting of 5 types. We evaluatethe latest MLLMs from the ChatGPT, Gemini, Claude, Qwen, and Kimiseries using zero-shot and chain-of-thought prompts under both low andhigh reasoning setups. Our results reveal limitations in current MLLMs’ability in comprehending Bangla forms, particularly in accurately local-izing highly granular form entities. Our dataset and code is available at:https://huggingface.co/datasets/Mausul/bafco

Enabling reliable long-horizon robotic manipulation is a cru-cial step toward open-world embodied intelligence. However, VLM-basedplanners treat each step as an isolated observation-to-action mapping,forcing them to reinfer scene geometry from raw pixels at every decisionstep while remaining unaware of how prior actions have reshaped the en-vironment. Despite strong short-horizon performance, these systems lackthe spatio-temporal reasoning required for persistent geometric anchoringand memory of action-triggered state transitions. Without persistentstate tracking, perceptual errors accumulate across the execution horizon,temporarily occluded objects are catastrophically forgotten, and com-pounding failures lead to precondition violations that cascade throughsubsequent steps. In contrast, humans maintain a persistent mental modelthat continuously tracks spatial relations and action consequences acrossinteractions rather than reconstructing them at each instant. Inspiredby this human capacity for causal spatio-temporal reasoning with persis-tent memory, we propose RoboStream, a training-free framework thatachieves geometric anchoring through Spatio-Temporal Fusion Tokens(STF-Tokens), which bind visual evidence to 3D geometric attributesfor persistent object grounding, and maintains causal continuity via aCausal Spatio-Temporal Graph (CSTG) that records action-triggeredstate transitions across steps. This design enables the planner to tracecausal chains and preserve object permanence under occlusion withoutadditional training or fine-tuning. RoboStream achieves a 90.5% successrate on long-horizon RLBench tasks and a 44.4% success rate on challeng-ing real-world block-building tasks, where both SoFar and VoxPoser score11.1%, demonstrating that spatio-temporal reasoning and causal memoryare critical missing components for reliable long-horizon manipulation.

In person re-identi_x001C_cation, neighbor-based methods haveachieved signi_x001C_cant success by interacting with neighbor samples to ob-tain more robust representations. However, existing methods rely onlyon a_x001E_nity relations, causing their success to depend heavily on the re-liability of selected neighbors. We _x001C_nd that a_x001E_nity-only interaction of-ten fails in challenging scenarios due to the inevitable presence of noisyneighbors. To enable e_x001B_ective interactions under noisy neighborhoods, werevisit neighbor-based methods under distinct reliability conditions andpropose a novel Adaptive Neighbor Feature Interaction (ANFI)method. The core idea of ANFI is to account for negative e_x001B_ects fromnoisy neighbors, allowing samples to remain distinguishable from falsepositive neighbors. Unlike existing methods, ANFI models not only a_x001E_n-ity relations but also discrepancy relations, and employs sample-wiseadaptive weighting for these two types of relations. Given that capturingnegative e_x001B_ects from noisy neighbors di_x001B_ers signi_x001C_cantly from traditionalrelation learning, we derive discrepancy relations from a new neighbor-hood similarity, which provides more information than pairwise sim-ilarity. In addition, we propose Noisy Relation Supervision (NRS)to train ANFI, gradually injecting robustness to noisy relations into themodel. Extensive experiments conducted under standard, cross-modal,and cross-domain settings, including comparisons with neighbor-basedmethods and re-ranking methods, demonstrate the superiority of ourmethod across various neighbor distributions.

Image editing has become an important capability of modernvisual generative models. Most existing approaches generate the targetimage directly given an input image and an editing instruction. Whileeffective for simple appearance modifications, this paradigm often failswhen edits require multi-step inference over spatial, temporal, causal, orcommonsense relationships, which entails reasoning about how the sceneshould evolve through a series of plausible intermediate states from thesource image to the desired result. Notably, modern video generationmodels naturally capture such transformation processes: by modelinghow scenes evolve over time, they encode strong temporal coherence andimplicit physical priors. However, directly applying video models to imageediting remains challenging due to the mismatch between editing instruc-tions and video generation objectives. To bridge this gap, we proposeVGEdit, a reinforcement learning framework that empowers video gener-ation models for image editing. Our method consists of two components:(1) Instruction transfer : an MLLM converts the editing instruction into avideo description, reframing the edit as a plausible transformation process.(2) Reinforcement learning: we adapt the video model with reward-basedoptimization, treating intermediate frames as implicit reasoning chainsand the last frame as the target edited image. This formulation eliminatesthe need for ground-truth editing videos and enables direct optimizationfrom rewards that capture instruction faithfulness, contextual consistency,and user preference alignment. Experiments show that VGEdit signifi-cantly outperforms prior methods on reasoning-informed image editingbenchmarks, demonstrating that connecting image editing and videogeneration in a process-centric manner provides a scalable path towardmore physically grounded and reasoning-capable visual editing.

Camera control has been extensively studied in conditional video generation; however, stable altering camera trajectories while faithfully preserving video content remains a challenging task. The mainstream approach to achieving camera control is warping a 3D representation according to the target trajectory. However, such methods fail to fully leverage the 3D priors of video diffusion models (VDMs) and often fall into the Inpainting Trap, resulting in subject inconsistency and degraded generation quality. To address this problem, we propose DepthDirector, a video re-rendering framework with stable camera controllability. Our key insight is to decompose the conditional injection of video content and camera trajectories by leveraging the depth video from explicit 3D representation as camera-control guidance. Specifically, we design a View-Content Dual-Stream Condition mechanism, enabling VDMs to comprehend camera movements and leverage their 3D understanding capabilities. Additionally, we construct a large-scale multi-camera synchronized dataset named MultiCamWarp Dataset using Unreal Engine 5. Extensive experiments show that DepthDirector outperforms existing methods. Our code and dataset will be publicly available on our project website: https://eleanor6725.github.io/DepthDirector/.

Embodied Visual Tracking (EVT) requires an agent to con-tinuously follow a language-specified target from egocentric observations.Most existing methods adopt single-agent imitation learning, which de-pends on expert demonstrations and can struggle to generalize to sce-narios beyond the training distribution. Inspired by competition-drivencapability evolution in nature, we propose CoMaTrack, a competitivemulti-agent reinforcement learning (RL) framework that trains agentsin a dynamic adversarial setting with competitive subtasks, yieldingstronger adaptive planning and interference-resilient strategies. We fur-ther propose CoMaTrack-Bench, a Habitat-based benchmark for language-conditioned competitive EVT, featuring dynamic dueling scenarios be-tween a tracker and adaptive opponents across diverse environments andinstructions, enabling standardized robustness evaluation under activeadversarial interactions. Experiments show that CoMaTrack achievesstate-of-the-art results on both standard benchmarks and CoMaTrack-Bench. Notably, a 3B VLM trained with our framework surpasses pre-vious single-agent imitation learning methods based on 7B models onthe challenging EVT-Bench, achieving 92.1% in STT, 74.2% in DT, and57.5% in AT.

Pixel-space diffusion has recently re-emerged as a strong al-ternative to latent diffusion, enabling high-quality generation withoutpretrained autoencoders. However, standard pixel-space diffusion mod-els receive relatively weak semantic supervision and are not explicitlydesigned to capture high-level visual structure. Recent representation-alignment methods (e.g., REPA) suggest that pretrained visual featurescan substantially improve diffusion training, and visual co-denoising hasemerged as a promising direction for incorporating such features intothe generative process. However, existing co-denoising approaches of-ten entangle multiple design choices, making it unclear which are trulyessential. We therefore present V-Co, a systematic study of visual co-denoising in a unified JiT-based framework. This controlled setting allowsus to isolate the ingredients that make visual co-denoising effective. Ourstudy reveals two main ingredients. First, co-denoising benefits from pre-serving feature-specific computation while enabling flexible cross-streaminteraction, which leads to a fully dual-stream architecture together witha structurally defined unconditional prediction for classifier-free guid-ance. Second, it requires both stronger semantic supervision and propercross-stream calibration, which we realize through a perceptual-driftinghybrid loss and RMS-based feature rescaling. Together, these findingsyield a simple recipe for visual co-denoising. Experiments on ImageNet-256 show that, at comparable model sizes, V-Co outperforms the un-derlying pixel-space diffusion baseline and strong prior pixel-diffusionmethods while using fewer training epochs, offering practical guidancefor future representation-aligned generative models.

Cross-domain Few-shot Segmentation (CD-FSS) aims to tra-nsfer knowledge learned from source domain to distinct target domains,segmenting unseen target classes with only a few annotated samples.Although existing methods have made significant progress, they stillrely on training or fine-tuning processes, which incur high computationalcosts and risk overfitting. We observe that when powerful and general-purpose vision foundation models are incorporated into these methods,their performance shows only marginal improvement or even degradesdue to overfitting. To address this, we eliminate trainable parameters andpropose a training-free framework to avoid both training overhead andoverfitting. Built upon the self-supervised vision encoder DINOv3, ourframework addresses cross-domain challenges through three core mod-ules. First, the Semantic-aware Feature Re-fusion (SAFR) module iden-tifies and re-fuses features that emphasize semantic patterns, generatingrepresentations with enhanced semantic discriminability. Additionally,the Adaptive Support Enhancement (ASE) module narrows semanticgaps between support and query through robust query information ag-gregation. Finally, the Hybrid Prototype Matching (HPM) module inte-grates matching results from diverse prototypes to adapt to varying se-mantic complexity across domains. Extensive experiments on four targetdomain datasets demonstrate that our method achieves state-of-the-artperformance in CD-FSS without any training. Our code is available athttps://github.com/Sparkling-Water/RSRM.

Recent advances in sparse voxel representations have significantly improved the quality of 3D content generation, enabling highresolution modeling with fine-grained geometry. However, existing frameworks suffer from severe computational inefficiencies due to the quadratic complexity of attention mechanisms in their two-stage diffusion pipelines. In this work, we propose Ultra3D, an efficient 3D generation framework that significantly accelerates sparse voxel modeling without compromising quality. Our method leverages the compact VecSet representation to efficiently generate a coarse object layout in the first stage, reducing token count and accelerating voxel coordinate prediction. To refine pervoxel latent features in the second stage, we introduce Part Attention, a geometry-aware localized attention mechanism that restricts attention computation within semantically consistent part regions. This design preserves structural continuity while avoiding unnecessary global attention, achieving up to a 6.7× speed-up in latent generation. To support this mechanism, we construct a scalable part annotation pipeline that converts raw meshes into part-labeled sparse voxels. Extensive experiments demonstrate that Ultra3D supports high-resolution 3D generation at 1024 resolution and achieves state-of-the-art performance in both visual fidelity and user preference.

Trained policies for real-world robotics rely on computer vi-sion components, typically in the form of pre-trained visual encoders.These encoders are an essential component and it has been shown thattheir power does not emerge from training on robotics downstream lossesalone. Pre-training with auxiliary losses in the form of computer-visionpre-text tasks is a defining factor and heavily conditions agent perfor-mance in robotics tasks. In this unprecedented large-scale study, we ran966 navigation episodes of static point goal navigation in a real-worldbuilding for 24km and asked which components really matter for thecomputer vision aspects of robotics: we evaluate state-of-the art visualencoders in realistic conditions. We explore the usefulness of heteroge-neous multi-teacher distillation leading to encoders with multiple dif-ferent and complementary skills. We investigate how much informationfrom these encoders is necessary for robotics by bottlenecking them in aprincipled and “spatially useful” way and we show that this leads to theemergence of interpretable features linked to affordances. We also arguethat training policies on RGB data alone does not lead to an optimalusage of visual features and show this by finetuning policies pre-trainedon privileged information. All in all, we paint a more complete picture ofwhat aspects of computer vision are relevant for real-world navigation.

Vision-Language Models (VLMs) exhibit strong visual rea-soning capabilities, yet they still struggle with 3D understanding. Inparticular, VLMs often fail to infer a text-consistent goal 6D pose of atarget object in a 3D scene. However, we find that with some inference-time techniques and iterative reasoning, VLMs can achieve dramatic per-formance gains. Concretely, given a 3D scene represented by an RGB-Dimage (or a compositional scene of 3D meshes) and a text instructionspecifying a desired state change, we repeat the following loop: observethe current scene; evaluate whether it is faithful to the instruction; pro-pose a pose update for the target object; apply the update; and ren-der the updated scene. Through this closed-loop interaction, the VLMeffectively acts as an agent. We further introduce three inference-timetechniques that are essential to this closed-loop process: (i) multi-viewreasoning with supporting view selection, (ii) object-centered coordinatesystem visualization, and (iii) single-axis rotation prediction. Withoutany additional fine-tuning or new modules, our approach surpasses priormethods at predicting the text-guided goal 6D pose of the target object.It works consistently across both closed-source and open-source VLMs.Moreover, when combining our 6D pose prediction with simple robot mo-tion planning, it enables more successful robot manipulation than recentVision-Language-Action models (VLAs). Finally, we conduct an ablationstudy to demonstrate the necessity of each proposed technique.

LiDAR-based Human Mesh Reconstruction (HMR) is crucialfor understanding human activities in real-world environments. How-ever, the inherent sparsity of LiDAR point clouds often causes severeloss of anatomical details, limiting reconstruction accuracy. Althoughrecent methods have made promising progress, most still follow a decou-pled pipeline that first estimates skeletal joints and then reconstructsthe mesh, leading to error propagation under sparse observations. Toaddress this issue and fully exploit both the topological guidance of theskeleton on the mesh and the spatial constraints of the surface mesh onthe skeleton, we propose SynHMR, a Synergistic Joint-Mesh Modelingframework for robust LiDAR-based HMR. Specifically, we first initializeboth the skeletal joints and the mesh vertices as queries equipped withpositional information for co-optimization. We then design a Geometry-Aware Transformer that leverages point cloud features to apply the at-tention mechanism between joint and vertex queries, enabling their it-erative refinement through mutual information exchange. Furthermore,we introduce a Noise-Augmented Learning strategy that injects per-turbed joint queries during training. This reduces the impact of noisewhile encouraging bidirectional association between joints and mesh. Ex-tensive experiments on LiDARHuman26M, SLOPER4D, and Human-M3 demonstrate that SynHMR achieves state-of-the-art performanceamong LiDAR-based HMR approaches. The project page is availableat https://github.com/ShiRui1208/SynHMR.

Contrastive vision–language models (VLMs) such as CLIP,EVACLIP, ImageBind, and SigLIP exhibit impressive zero-shot general-ization, yet their performance on specific data distributions can furtherbe enhanced by improving the synchronization of the vision and languagerepresentations. For this purpose, we present SyncVL, a bi-directionalVL synchronization framework that enables fully unsupervised adapta-tion of VLMs. SyncVL jointly refines visual and textual representationsthrough two lightweight transformer-based encoder-decoder synchroniz-ers (Vision-to-Text and Text-to-Vision) that progressively project VLrepresentation into a shared latent subspace without requiring labeleddata. To ensure robust synchronization, we form groups using multipleimage-views and text templates. Both synchronizers are trained using areward function motivated by group relative policy optimization whilekeeping the original VLMs frozen. Additionally, a mutual-distillationmechanism is proposed to iteratively reinforce each synchronizer. Weextensively evaluate SyncVL on 24 benchmark datasets spanning unsu-pervised VL adaptation, out-of-distribution matching, unsupervised clus-tering, cross-modal retrieval, object detection, and segmentation. Acrossall tasks, SyncVL consistently improves over state-of-the-art baselines,establishing a new standard for unsupervised adaptation of VLMs.

Autonomous agents require representations that are gen-eral, causal, and physically structured to operate in real-time stream-ing environments. However, current vision foundation models remainfragmented, specializing narrowly in image semantic perception, offlinetemporal modeling, or spatial geometry. This paper introduces Om-niStream, a unified streaming visual backbone that effectively per-ceives, reconstructs, and acts from diverse visual inputs. By incorporat-ing causal spatiotemporal attention and 3D rotary positional embeddings(3D-RoPE), our model supports efficient, frame-by-frame online process-ing of video streams via a persistent KV-cache. We pre-train OmniStreamusing a synergistic multi-task framework coupling static and temporalrepresentation learning, streaming geometric reconstruction, and vision-language alignment on 29 datasets. Extensive evaluations show that, evenwith a strictly frozen backbone, OmniStream achieves consistently com-petitive performance with specialized experts across image and videoprobing, streaming geometric reconstruction, complex video and spatialreasoning, and robotic manipulation. Rather than pursuing benchmark-specific dominance, our work demonstrates the viability of training a sin-gle, versatile vision backbone that generalizes across semantic, spatial,and temporal reasoning, i.e., a meaningful step towards general-purposevisual understanding for interactive and embodied agents.

Vision-Language-Action (VLA) models excel in static manipulation but struggle in dynamic environments with moving targets. This performance gap primarily stems from a scarcity of dynamic manipulation datasets and the reliance of mainstream VLAs on single-frame observations, restricting their spatiotemporal reasoning capabilities. To address this, we introduce DOMINO, a large-scale dataset and benchmark for generalizable dynamic manipulation, featuring 35 tasks with hierarchical complexities, over 110K expert trajectories, and a multidimensional evaluation suite. Through comprehensive experiments, we systematically evaluate existing VLAs on dynamic tasks, explore effective training strategies for dynamic awareness, and validate the generalizability of dynamic data. Furthermore, we propose PUMA, a dynamicsaware VLA architecture. By integrating scene-centric historical optical flow and specialized world queries to implicitly forecast object-centric future states, PUMA couples history-aware perception with short-horizon prediction. Results demonstrate that PUMA achieves state-of-the-art performance, yielding a 6.3% absolute improvement in success rate over baselines. Moreover, we show that training on dynamic data fosters robust spatiotemporal representations that transfer to static tasks.