ECCV 2026 Papers

Out-of-distribution (OOD) detection is crucial for the safedeployment of deep neural models in applications such as autonomousdriving. With the emerging multimodal nature of modern applications,recent attention has shifted toward OOD detection in multimodal set-tings. However, current multimodal OOD detection methods fail to fullyexploit the synergy among modalities: they treat all modalities equally,disregarding their varying detection performance, and they are unableto capture the diverse uncertainty information encoded at the logit level.In this paper, we propose to exploit the dark knowledge within unimodalexperts as the key to revealing their synergy. To this end, we introducea self multimodal OOD distillation framework, which leverages logits asuncertainty-aware soft targets to train a holistic model that operates inthe joint embedding space of all modalities. Specifically, the proposedframework accounts for the negative effects of underperforming modal-ities and effectively fuses both the rich feature-level knowledge and thelogit-level knowledge of modalities. As a result, our method improvesthe performance of current state-of-the-art multimodal OOD detectionmethods, achieving gains of up to 30% across diverse OOD detectionbenchmarks, spanning two tasks and five multimodal OOD datasets.

Recent advances in diffusion models have enabled impressivevideo editing capabilities, yet production-grade Video Object Insertion(VOI) remains challenging due to inadequate 4D scene understandingand a lack of proper optical interactions, such as shadows and reflec-tions. To address these limitations, we present InsertAnywhere, a com-prehensive VOI framework that achieves geometrically grounded objectplacement and optics-aware video synthesis. Our approach first lever-ages a 4D-aware mask generation module that allows users to anchor anobject’s 3D pose in a single frame. The framework automatically prop-agates this placement across the video, accurately handling local scenedynamics and occlusions. To synthesize realistic physical lighting inter-actions, we introduce Optics-Aware Representation Alignment, a novelstrategy that utilizes an extended mask to guide feature extraction, en-abling optical effects to seamlessly extend beyond the inserted object’sboundary. Finally, to overcome the lack of training data for such phenom-ena, we construct and open-source ROSE++, a specialized quadrupletdataset tailored for the supervised learning of optical effects. Extensiveexperiments demonstrate that InsertAnywhere produces geometricallyplausible and photometrically realistic insertions in complex real-worldscenarios, significantly outperforming existing research and commercialgenerative tools.

While multimodal Large Language Models (MLLMs) excel at offline video understanding, an interesting question of how far they are from serving as a real-time procedural coach remains unknown. Such a role typically requires an MLLM to continuously monitor the execution, detect mistakes, and provide corrective guidance in a closed-loop interaction. In this paper, we construct GuideMe, the first multi-domain benchmark for streaming video that supports training and evaluation of MLLMs for closed-loop interactive task guidance. It comprises 2,458 videos spanning 223.7 hours across diverse domains (e.g., cooking, object manipulation, daily-life guidance, and fitness), with 47,775 interaction samples covering next-step instructions, completion feedback, error detection, and corrective guidance. To evaluate existing models on GuideMe, we design a three-component assessment framework to measure the capabilities of representative MLLMs, which consists of temporal-semantic bipartite matching for sequence-level alignment, behavioral classification for intervention timing, and LLM-as-a-Judge for content quality. Extensive experiments highlight a critical performance asymmetry: despite excelling at providing instructions, existing MLLMs consistently fail to identify execution errors and respond with corrective feedback. Code and data are released at (cid:140)homepage.

Music-to-dance generation aims to translate auditory sig-nals into expressive human motion. Despite promising progress, existingmethods remain underexplored in achieving high-quality generation un-der a strict efficiency budget, due to (i) generative strategies that sup-port high-fidelity few-step sampling, and (ii) per-step model architecturesthat are optimized for efficient yet refined long-horizon motion synthesis,thereby degrading downstream user experience in terms of responsive-ness and visual fidelity. Thus, we propose FlowerDance, which not onlygenerates refined motion with physical plausibility and artistic expres-siveness, but also achieves significant generation efficiency on inferencespeed and memory utilization. Specifically, FlowerDance combines Mean-Flow with Physical Consistency Constraints, which enables high-qualitymotion generation with only a few sampling steps. Moreover, Flower-Dance leverages a simple but efficient model architecture with BiMamba-based backbone and Channel-Level Cross-Modal Fusion, which generateslong-horizon dance with efficient non-autoregressive manner. Meanwhile,FlowerDance supports motion editing, enabling users to interactively re-fine dance sequences. Extensive experiments on AIST++ and FineDanceshow that FlowerDance achieves state-of-the-art results in both motionquality and generation efficiency. Code is available at https://sun-happy-ykx.github.io/FlowerDance/.

Rendering views using 3D scene representations such as Gaus-sian Splatting (3DGS), Neural Radiance Fields (NeRF), meshes, or evenpoint clouds produces artifacts when input views are sparse or targetviews lie far from the input. Recent work mitigates these artifacts usingdiffusion-based generative priors, but is specialized to individual repre-sentations and require custom architectures or extensive retraining. Wepresent FixAnything, a single model for fixing a wide range of render-ing artifacts. It does so by repurposing a pretrained video generativemodel, leveraging its implicit multi-view priors with only minimal mod-ification and lightweight finetuning. Our key insight is that even noisily-rendered sequences preserve camera motion and coarse scene structure,allowing cleanup to be formulated as video-to-video translation. To con-trol what scene structure should be preserved, we introduce a binarymask denoting the clean pixels, enabling the model to anchor its outputto high-quality inputs (e.g. training views) while refining the rest. To en-courage FixAnything to produce 3D-consistent renderings that supportdownstream reconstruction, we use camera pose accuracy (recovered viastructure-from-motion) as a reward signal for direct preference optimiza-tion (DPO). Across four distinct 3D representations, FixAnything con-sistently improves rendering quality with lightweight finetuning, demon-strating that a single generalist video prior can replace multiple specialistrefinement pipelines. The simplicity of the framework enables immediateadoption of stronger future video models without architectural redesign.

Accurate estimation of forest biomass, a major carbon sink, relies heavily on tree-level traits such as height and species. Unoccupied Aerial Vehicles (UAVs) capturing high-resolution imagery from a single RGB camera offer a cost-effective and scalable approach for mapping and measuring individual trees. We introduce BIRCH-Trees, the first benchmark for individual tree height and species estimation from tree-centered UAV images, spanning three datasets: temperate forests, tropical forests, and boreal plantations. We also present DINOvTree, a unified approach using a Vision Foundation Model (VFM) backbone with task-specific heads for simultaneous height and species prediction. Through extensive evaluations on BIRCH-Trees, we compare DINOvTree against commonly used vision methods, including VFMs, as well as biological allometric equations. We find that DINOvTree achieves top overall results with accurate height predictions and competitive classification accuracy while using only 54% to 58% of the parameters of the second-best approach.

Visual search in 3D environments requires embodied agentsto actively explore their surroundings and acquire task-relevant evidence.However, existing visual search and embodied AI benchmarks, includ-ing EQA, typically rely on static observations or constrained egocen-tric motion, and thus do not explicitly evaluate fine-grained viewpoint-dependent phenomena that arise under unrestricted 5-DoF viewpointcontrol, such as disambiguating object attributes observable only fromspecific angles. To address this limitation, we introduce E3VS-Bench,a benchmark for embodied 3D visual search where agents must controltheir viewpoints in 5-DoF to gather viewpoint-dependent evidence forquestion answering. E3VS-Bench consists of 99 high-fidelity 3D scenesreconstructed using 3D Gaussian Splatting and 2,014 question-drivenepisodes. 3D Gaussian Splatting enables photorealistic free-viewpointrendering that preserves fine-grained visual details (e.g., small text andsubtle attributes) often degraded in mesh-based simulators, thereby al-lowing the construction of questions that cannot be answered from asingle view and instead require active inspection across viewpoints in5-DoF. We evaluate multiple state-of-the-art VLMs and compare theirperformance with humans. Despite strong 2D reasoning ability, all mod-els exhibit a substantial gap from humans, highlighting limitations inactive perception and coherent viewpoint planning specifically under full5-DoF viewpoint changes. The benchmark, code, and dataset are publiclyavailable at https://k0uya.github.io/e3vs-proj/.

Adapting pre-trained text-to-image diffusion models, whether to learn new visual concepts or erase unwanted ones, is routinely evaluated on its intended effects alone. We argue this framing is incomplete. Through sparse autoencoder analysis and zero-shot classification, we demonstrate that adaptation systematically damages semantically unrelated concepts in ways that aggregate metrics structurally cannot surface: when damage is severe enough for FID and KID to respond, the model is already nearly unusable; when the model remains functional, FID and KID stay flat while specific classes silently suffer worst-case zero-shot accuracy drops of up to 18.9 points and concept-level distributions shift dramatically. This pattern appears at both ends of the adaptation spectrum (concept customization and concept unlearning), suggesting it is a systematic consequence of weight-level modification rather than an artifact of any particular method. To surface this hidden drift before deployment, we introduce DriftScope, a prompt-level diagnostic tool that takes any two model checkpoints and returns a ranked list of tokens whose visual concepts have shifted most between them. DriftScope optimizes a soft prompt to attribute drift at the token level without requiring access to real data or model internals. The result is an interpretable, concept-level audit that aggregate evaluation cannot provide.

The long-standing vision of general-purpose robots hinges on theirability to understand and act upon natural language instructions. Vision-Language-Action (VLA) models have made remarkable progress toward thisgoal, yet their generated actions can still misalign with the given instructions.In this paper, we investigate test-time verification as a means to shrinkthe “intention-action gap.” We first characterize the test-time scaling lawsfor embodied instruction following and demonstrate that jointly scaling thenumber of rephrased instructions and generated actions greatly increasestest-time sample diversity, often recovering correct actions more efficientlythan scaling each dimension independently. To capitalize on these scalinglaws, we present CoVer, a contrastive verifier for vision–language–actionalignment, and show that our architecture scales gracefully with additionalcomputational resources and data. We then introduce CoVer-VLA, a hierar-chical test-time verification pipeline using the trained verifier. At deployment,our framework precomputes a diverse set of rephrased instructions from aVision-Language-Model (VLM), repeatedly generates action candidates foreach instruction, and then uses the verifier to select the optimal high-levelprompt and low-level action chunks. Compared to scaling policy pre-training⋆ †Equal contribution. § Equal advising.on the same data, our verification approach yields 22% gains in-distributionand 13% out-of-distribution on the SIMPLER benchmark, with a further45% improvement in real-world experiments. On the PolaRiS benchmark,CoVer-VLA achieves 14% gains in task progress and 9% in success rate.

While Audio-Visual Language Models (AVLMs) have achievedremarkable progress over recent years, their reliability is bottleneckedby cross-modal hallucination. A particularly pervasive manifestation isvideo-driven audio hallucination: models routinely exploit visual short-cuts to hallucinate expected sounds, discarding true auditory evidence.To counteract this deeply ingrained visual dominance, we propose Audio-Contrastive Preference Optimization (ACPO). This dual-axis preferencelearning framework introduces an output-contrastive objective to penal-ize visual descriptions masquerading as audio facts, alongside an input-contrastive objective that swaps audio tracks to explicitly penalize gener-ation invariant to the true auditory signal. Extensive experiments demon-strate that ACPO establishes highly faithful audio grounding and mit-igates audio hallucination. Project page: https://vision.cs.utexas.edu/projects/acpo/

Pre-trained flow-based models excel at synthesizing complexscenes, yet lack a direct mechanism for disentangling and customizingtheir underlying concepts from one-shot real-world sources. To demystifythis process, we first introduce a novel differential probing technique toisolate and analyze the influence of individual concept tokens on the ve-locity field over time. This investigation yields a critical insight: the gen-erative process is not monolithic but unfolds in three distinct stages. Aninitial Blueprint Stage establishes low-frequency structure, followed bya pivotal Instantiation Stage where content concepts emerge with peakintensity and become naturally disentangled, creating an optimal windowfor manipulation. A final concept-insensitive Refinement Stage thensynthesizes fine-grained details. Guided by this discovery, we proposeConceptWeaver, a framework for one-shot concept disentanglement.ConceptWeaver learns concept-specific semantic offsets from a single ref-erence image using a stage-aware optimization strategy that aligns withthe three-stage framework. These learned offsets are then deployed dur-ing inference via our novel ConceptWeaver Guidance (CWG) mechanism,1 ∗Equal contribution; † Project Lead; ‡ Co-corresponding author.which strategically injects them at the appropriate generative stage. Ex-tensive experiments validate that ConceptWeaver enables high-fidelity,compositional synthesis and editing, demonstrating that understandingand leveraging the intrinsic, staged nature of flow models is key to un-locking precise, multi-granularity content manipulation. We will releaseour code at https://github.com/JasperChennn/ConceptWeaver.

When a real-world scene is captured by a smartphone cam-era and viewed on its screen, the displayed image often differs noticeablyfrom the original scene in color, brightness, and contrast. This gap per-sists despite substantial advances in both modern cameras and displays.A key reason is that most pipelines factor the high-dimensional capture-to-display process into two separately calibrated camera and displaystages, and then connect them through low-dimensional color transforms,leading to information bottlenecks and inevitable error accumulation. Toaddress this systemic challenge, we propose Color Pass-Through, anend-to-end learned framework that operates directly on captured images.Our key insight is to treat the camera and display as a coupled systemrather than calibrating them in isolation. Coupling the camera and dis-play yields two practical advantages: (1) it brings the entire real-worldscenes to the display via end-to-end optimization, and (2) it allows effi-cient one-step calibration for each distinct observer via complete capture-to-display path. We validate Color Pass-Through using both digi-tal and human observers. Compared with representative baselines, ourmethod achieves an average gain of +2.0 points on a 5-point user-studyand more than 2× improvement on quantitative metrics, demonstratingimproved reproduction of the perceived color of the original scene. Seeproject page: https://lyricccco.github.io/color-pass-through/

We introduce CameraAnything, the first unified frame-work for camera controlled video editing that enables joint control ofboth intrinsic and extrinsic camera parameters. Existing approaches ei-ther rely on expensive 3D reconstruction to achieve full camera function-ality or restrict editing to extrinsic parameter manipulation. Moreover,the coupled influence of intrinsic and extrinsic parameters on video ap-pearance makes disentangled modeling particularly challenging. To ad-dress this, we adopt per-pixel Plücker ray injection alongside resolution-aware 3D RoPE in self-attention, building both camera conditioning andspatial positional encoding on the target latent to jointly control cam-era position, focal length, and native resolution editing without croppingor outpainting. To overcome the scarcity of paired training data, wefurther develop a scalable synthetic pipeline that constructs diverse dy-namic scenes through structured multi-camera recording and generatessynchronized videos with varied camera configurations. With a tailoredorthogonal training strategy, CameraAnything enables expressive videoreshooting with arbitrary viewpoint control, focal length adjustment, res-olution adaptation, and multi-shot transitions within a single generationprocess, offering strong practical value for cinematic video editing andcross-platform content adaptation in video production.

Building high-fidelity digital twins of articulated objects fromvisual data remains a central challenge. Existing approaches depend onmulti-view captures of the object in discrete, static states, which severelyconstrains their real-world scalability. In this paper, we introduce Articu-lat3D, a novel framework that constructs such digital twins from casuallycaptured monocular videos by jointly enforcing explicit 3D geometric andmotion constraints. We first propose Motion Prior–Driven Initialization,which leverages 3D point tracks to exploit the low-dimensional structureof articulated motion. By modeling scene dynamics with a compact set ofmotion bases, we facilitate soft decomposition of the scene into multiplerigidly-moving groups. Building on this initialization, we introduce Ge-ometric and Motion Constraints Refinement, which enforces physicallyplausible articulation through learnable kinematic primitives parameter-ized by a joint axis, a pivot point, and per-frame motion scalars, yieldingreconstructions that are both geometrically accurate and temporally co-herent. Extensive experiments demonstrate that Articulat3D achievesstate-of-the-art performance on synthetic benchmarks and real-worldcasually captured monocular videos, significantly advancing the feasibilityof digital twin creation under uncontrolled real-world conditions. Here isour project page.

Learning-based stereo matching models struggle in under-water environments due to scarce in-domain data and the difficulty ofextracting discriminative correspondences from degraded imagery. In thiswork, we present AquaStereo, a perception-enhanced framework witha data simulation pipeline and a self-distillation strategy that jointly ad-dress data scarcity and feature degradation in underwater stereo match-ing. First, a depth-conditioned diffusion pipeline renders underwaterstereo pairs while preserving binocular geometry, with a lightweight left-right consistency module ensuring geometric alignment. Training on thissynthetic corpus effectively narrows the terrestrial–underwater gap andimproves zero-shot robustness. Second, a frozen binocular teacher trainedon clean terrestrial pairs guides a student exposed to rendered underwa-ter pairs with perturbations. A stage-weighted sequence loss is performedto align the student’s disparities with the teacher’s geometry, while aclean-branch supervision with shared pseudo targets prevents scale drift.To further enhance feature stability under turbidity and low texture,we introduce learnable perception frames, a perception-enhanced featureformulation that constructs robust matching descriptors by fusing tem-poral cues from two auxiliary views encoded by a video backbone withsemantic features extracted by a strong image encoder. Extensive experi-ments demonstrate that AquaStereo substantially improves robustnessand zero-shot generalization in challenging underwater scenarios.

Embodied visual navigation requires agents with real-time perception, logical reasoning and efficient decision-making capabilities. Existing collaborative modular frameworks struggle to balance accurate decision-making and efficient online execution, leading to error propagation and inference blocking. To tackle this limitation, we propose AMCoNav, an Asynchronous Multi-module Collaborative Framework that unites real-time lightweight backbone networks with on-demand zeroshot large model modules (ZLMM) for robust and real-time embodied navigation. The framework has four tightly coupled components: (1) a Real-Time Multimodal Decision Module (RMDM) for continuous decision making; (2) Zero-shot Large Model Modules (ZLMM) for on-demand asynchronous reasoning; (3) a latency-tolerant Multimodal Shared Context (MSC) for cross-module coordination; and (4) a Bayesian Probabilistic Decision Fusion Module (BPDFM) for robust score fusion and mode switching. Experiments on three representative embodied navigation benchmarks (HM3D-OVON, SG3D, and GOAT-Bench) demonstrate that AMCoNav consistently improves success rates while maintaining real-time execution. The most significant gain appears on HM3DOVON, where AMCoNav achieves an average 8.7% gain over prior stateof-the-art results, with 60.0%/60.0%/53.3% success rate (SR) on Val Seen/Val Seen Synonyms/Val Unseen. Ablation results show that asynchronous guidance from MSC and ZLMM and the BPDFM both contribute to performance improvement, and their combination yields the best overall results. These results show that asynchronous multi-module collaboration can improve navigation capability without sacrificing realtime responsiveness.

The ability to remotely measure an animal’s size and shape(i.e. its morphology) in its natural habitat is of key importance for ap-plications such as conservation, re-identification, and biomechanical un-derstanding. Monocular depth estimation (MDE) and 3D reconstructionapproaches can be used. This is useful to generate accurate deformablemesh models or to characterize biologically important parameters such assexual dimorphism or the size distribution of populations. Whilst depthestimation and 3D reconstruction have been extensively studied as coretopics in computer vision, ranging from early work on simple, rigid ob-jects, to more recent work on human and animal reconstruction, themajority of existing animal models are trained on video/image baseddatasets that lack metric scale and ground truth e.g. from camera trapimages or from public videos. In addition, they do not fully reflect thechallenges of accurately estimating the focal animal’s scale when it is dis-tant from a camera. To address this limitation, we present WildDepth,a multimodal dataset and benchmark suite for depth estimation, behav-ior detection, and 3D reconstruction from diverse categories of animalsranging from domestic to wild environments with synchronized RGB andLiDAR. We provide three focused benchmarks: (1) monocular depth es-timation with per-distance and temporal stability analysis, (2) behav-ior detection, (3) 3D reconstruction and densification. Our results showthat large-scale MDE models degrade significantly at long range, whileLiDAR-anchored fusion reduces metric error by up to 25% – 30% RMSEin mid-range scenarios. Our aim is to enable camera trap data to bemore accurately ‘lifted’ to 3D morphometrics, unlocking a new era ofscale-aware animal modelling.⋆ M. Aamir, N. Muramatsu, and S. Shin contributed equally and are listed alpha-betically by surname.

The ability to remotely measure an animal’s size and shape(i.e. its morphology) in its natural habitat is of key importance for ap-plications such as conservation, re-identification, and biomechanical un-derstanding. Monocular depth estimation (MDE) and 3D reconstructionapproaches can be used. This is useful to generate accurate deformablemesh models or to characterize biologically important parameters such assexual dimorphism or the size distribution of populations. Whilst depthestimation and 3D reconstruction have been extensively studied as coretopics in computer vision, ranging from early work on simple, rigid ob-jects, to more recent work on human and animal reconstruction, themajority of existing animal models are trained on video/image baseddatasets that lack metric scale and ground truth e.g. from camera trapimages or from public videos. In addition, they do not fully reflect thechallenges of accurately estimating the focal animal’s scale when it is dis-tant from a camera. To address this limitation, we present WildDepth,a multimodal dataset and benchmark suite for depth estimation, behav-ior detection, and 3D reconstruction from diverse categories of animalsranging from domestic to wild environments with synchronized RGB andLiDAR. We provide three focused benchmarks: (1) monocular depth es-timation with per-distance and temporal stability analysis, (2) behav-ior detection, (3) 3D reconstruction and densification. Our results showthat large-scale MDE models degrade significantly at long range, whileLiDAR-anchored fusion reduces metric error by up to 25% – 30% RMSEin mid-range scenarios. Our aim is to enable camera trap data to bemore accurately ‘lifted’ to 3D morphometrics, unlocking a new era ofscale-aware animal modelling.⋆ M. Aamir, N. Muramatsu, and S. Shin contributed equally and are listed alpha-betically by surname.

Foundation model pseudo-labeling—labeling data strictly viazero-shot inference—enables massive scale, but performance is under-mined by hallucinations that evade standard thresholds. To eliminatethese errors, we introduce the Turing-inspired Label Imitation Game(LIG), a framework that formalizes pseudo-label pruning as an adversar-ial interrogation. Rather than filtering labels via isolated thresholds, weuse the LIG to train a Turing Test Network (TTN), a task-agnostic“judge” that evaluates candidate pseudo-labels within a dataset-wide con-text. Experiments across four diverse datasets demonstrate the TTN’srobustness, consistently enhancing label accuracy for three state-of-the-art vision-language models without costly supervision or retraining. Cru-cially, we demonstrate that learned semantic-contextual logic is a robustalternative to spatial-geometric verification, enabling a unique zero-shottask transfer capability—a TTN trained strictly on image classificationdatasets can effectively prune complex object detection pseudo-labels.This pruning yields F1 -score gains of 28% for the worst-performing base-line categories and 44% with task-specific fine-tuning. Significantly, wealso observe Category Revival, where the TTN pruning “detoxifies” thetraining signal for downstream models and enables them to recover fromzero recall on transfer-vulnerable classes. The pre-trained TTN modelsand code are available at https://github.com/voxel51/ttn.