ECCV 2026 Papers

Flow matching (FM) models generate data by learning a ve-locity field that transforms samples from a simple latent distribution,typically an isotropic Gaussian, to the data distribution. However, thegeometric mismatch between a unimodal, full-dimensional Gaussian anda multimodal, approximately low-dimensional data distribution leads toa complex velocity field that is costly to learn at training time andto integrate at inference time. In this paper, we propose ParsimoniousFlow Matching (PFM), which adopts a mixture of Gaussians (MoG)as the latent distribution whose geometry better aligns with the data.We identify key design choices that enable efficient FM, including anoptimal-transport data-latent coupling, MoG estimation via k-means,and eigenvalue regularization of the per-mode covariances. Theoretically,when both data and latent distributions are MoGs, we show that gra-dient descent for FM with affine velocity fields converges faster whenthe corresponding modes of the latent and data are more similar andthe per-mode covariances are well-conditioned. Further, for data thatfollows a separated MoG, we prove that replacing the isotropic Gaus-sian latent with a MoG in FM training accelerates gradient descentconvergence and lowers the initial loss. On CIFAR-10 and ImageNet(32 × 32), PFM achieves the same generation quality as the baselinein fewer training iterations, and produces higher-quality samples for thesame number of function evaluations at inference time. Code is available:https://github.com/tianjiaoding/pfm.

Conversational talking face generation has recently attractedincreasing attention, aiming to synthesize interactive talking videos wherecharacters speak, listen, and respond dynamically to each other. Thistask presents three core challenges: 1) Flexibility: enabling multi-rounddialogues with an arbitrary number of participants; 2) Naturalness: main-taining coherent motion and appropriate non-verbal feedback through-out the interaction; and 3) Efficiency: achieving real-time generation andlow computation overhead for long-term continuous online conversation.Despite recent advances, existing methods still fall short in balancingall three requirements. To bridge this gap, we introduce InterTalk, anovel and efficient framework designed for highly interactive conver-sational talking face generation. Built upon a motion-based architec-ture, InterTalk supports real-time conversation synthesis. Our methodachieves strong flexibility by explicitly modeling multi-round conversa-tional dynamics among each participant, eliminating constraints on theirnumbers. To enhance interactivity, we incorporate motion feedback frommultiple participants and introduce an iterative generation strategy formore natural behaviors. Besides, we disentangle motion into several fa-cial components, enabling targeted refinements for natural response suchas precise lip-sync and realistic eye-blinking. Finally, we construct a newmulti-person conversational dataset and enrich it with 3D face-baseddata augmentation. Extensive experiments demonstrate that InterTalkachieves superior interaction quality while maintaining real-time perfor-mance at 30 FPS.

Vision-Language-Action (VLA) models enable robots to per-form manipulation tasks directly from natural language instructions andare increasingly viewed as a foundation for generalist robotic policies.However, their reliability under Out-Of-Distribution (OOD) instructionsremains underexplored. In this paper, we reveal a critical failure modein which VLA policies continue executing visually plausible actions evenwhen the language instruction contradicts the scene. We refer to this phe-nomenon as linguistic blindness, where VLA policies prioritize visualpriors over instruction semantics during action generation. To systemati-cally analyze this issue, we introduce ICBench, a diagnostic benchmarkconstructed from the LIBERO dataset that probes language–action cou-pling by injecting controlled OOD instruction contradictions while keep-ing the visual environment unchanged. Evaluations on three represen-tative VLA architectures, including π0 , π0.5 , and OpenVLA-OFT, showthat these models frequently succeed at tasks despite logically impossi-ble instructions, revealing a strong visual bias in action generation. Tomitigate this issue, we propose Instruction-Guided Attention Re-calibration (IGAR), a train-free inference-time mechanism that rebal-ances attention distributions to restore the influence of language instruc-tions. IGAR operates without retraining or architectural modificationand can be directly applied to existing VLA models. Experiments across30 LIBERO tasks demonstrate that IGAR substantially reduces erro-neous execution under OOD contradictory instructions while preservingbaseline task performance. We additionally validate the approach on areal Franka robotic arm, where IGAR effectively prevents manipulationtriggered by inconsistent instructions.

3D Gaussian Splatting has revolutionized novel view synthe-sis with real-time rendering and photorealistic quality. However, standard3DGS methods fail in aerial scenes due to sparse anisotropic UAV captures.Two critical issues emerge: incomplete triangulated points in texturelessregions that destabilizes Gaussian initialization and view-aligned degen-eracies where Gaussians overfit limited observations rather than respecttrue geometry. We present StratoSplat, a framework that exploits the in-herent layered structure of aerial scenes for robust sparse-view 3DGS. Ourkey insight is that buildings, terrain, and roads naturally stratify alonggravity, providing strong geometric priors to regularize this problem. Weaddress point sparsity through layered memory guided initialization thatenforces multi-view consistency. To prevent optimization degeneracies, weintroduce neural multi-plane Gaussians where virtual primitives anchor tolearned parallel planes for geometric regularization. Extensive experimentsdemonstrate state-of-the-art performance, surpassing the best baseline byover 3dB. Code available at https://github.com/keloee/StratoSplat.

Animal Re-Identification (Animal ReID) is crucial for eco-logical monitoring in the wild, yet remains challenging due to the fine-grained and diverse visual features of animals, including intricate tex-tures, distinctive patterns, and various pose or form representations.While current Animal ReID methods mostly rely on holistic global fea-tures, they frequently overfit to the surrounding environment rather thancapturing localised, identity-relevant cues. Existing pattern-based meth-ods primarily focus on global generalised signals that are sensitive tobackground interventions and often fail to highlight identity evidencefor individual discrimination. Additionally, these methods are mainlyconstrained to the visual space only and lack pattern-based semanticinteractions from the multimodal space. To address these challenges,we introduce a Region-Aware Multimodal Interleaving (RAMI) frame-work, which formulates Animal ReID as a visual-textual semantic in-teraction over informative, localised regions. Specifically, RAMI projectssegmentation-derived region tokens into a shared token space and inter-leaves them with semantic tokens within a Transformer-based module tofacilitate dense cross-modal interactions. Given the lack of region-leveldata, we design a simple pipeline to produce biologically representativeregions from segmented animals. To the best of our knowledge, we are thefirst to formulate multimodal interleaving tailored for fine-grained dis-criminative tasks in Vision-Language Models (VLMs). Experiments showthat RAMI outperforms state-of-the-art methods across both benchmarkand in-the-wild datasets, with an average mAP gain of up to 18.7%. Ourcode is available at https://github.com/ML-4-SocialGood/RAMI.git.

Reconstructing articulated 3D objects from a single imagerequires jointly inferring object geometry, part structure, and motionparameters from limited visual evidence. A key difficulty lies in theentanglement between motion cues and object structure, which makesdirect articulation regression unstable. Existing methods address thischallenge through multi-view supervision, retrieval-based assembly, orauxiliary video generation, often sacrificing scalability or efficiency. Wepresent MonoArt, a unified framework grounded in progressive struc-tural reasoning. Rather than predicting articulation directly from im-age features, MonoArt progressively transforms visual observations intocanonical geometry, structured part representations, and motion-awareembeddings within a single architecture. This structured reasoning pro-cess enables stable and interpretable articulation inference without ex-ternal motion templates or multi-stage pipelines. Extensive experimentson PartNet-Mobility demonstrate that MonoArt achieves state-of-the-artperformance in both reconstruction accuracy and inference speed. Theframework further generalizes to robotic manipulation and articulatedscene reconstruction.

Reaction diagram parsing (RxnDP) is critical for extractingchemical synthesis information from literature. Although recent Vision-Language Models (VLMs) have emerged as a promising paradigm toautomate this complex visual reasoning task, their application is funda-mentally bottlenecked by the inability to align visual chemical entitieswith pre-trained knowledge, alongside the inherent discrepancy betweentoken-level training and reaction-level evaluation. To address these dualchallenges, this work enhances VLM-based RxnDP from two comple-mentary perspectives: prompting representation and learning paradigms.First, we propose Identifier as Visual Prompting (IdtVP), whichleverages naturally occurring molecule identifiers (e.g., bold numeralslike 1a) to activate the chemical knowledge acquired during VLM pre-training. IdtVP enables powerful zero-shot and out-of-distribution ca-pabilities, outperforming existing prompting strategies. Second, to fur-ther optimize performance within fine-tuning paradigms, we introduceRe3 -DAPO, a reinforcement learning algorithm that leverages verifi-able rewards to directly optimize reaction-level metrics, thereby achiev-ing consistent gains over standard supervised fine-tuning. Additionally,we release the ScannedRxn benchmark, comprising scanned historicalreaction diagrams with real-world artifacts, to rigorously assess model ro-bustness and out-of-distribution ability. Our contributions advance theaccuracy and generalization of VLM-based reaction diagram parsing.

Evaluation of autonomous vehicle (AV) planners in safety-critical closed-loop simulation is essential for real-world deployment. How-ever, generating controllable safety-critical scenarios remains challeng-ing. Existing approaches use soft guidance that provides only probabilis-tic preferences and cannot guarantee the satisfaction of geometric andseverity constraints associated with specific collision types. We introduceCollision-Constrained Flow Matching (CCFM), a novel framework thatguarantees precise collision control through hard physical constraints.CCFM consists of three key components: (i) a heuristic collision selectorthat optimally identifies an adversarial agent and collision type via com-posite scoring; (ii) structured hard constraints that explicitly define fourcollision types (rear-end, side, cut-in, head-on) through contact point,heading, and severity requirements; and (iii) a collision-constrained flowmatching sampler that enforces the constraints via Gauss-Newton man-ifold projection. CCFM achieves collision rate up to 46.4% on nuScenesand 83.1% on nuPlan, significantly outperforming baselines while pre-serving realistic driving behavior. By enabling controllable collision char-acteristics in safety-critical scenario generation, CCFM provides a reli-able foundation for AV safety evaluation and sim-to-real crash data gen-eration. The code and implementation details are available at https://github.com/KELISBU/CCFM.

Feed-forward 3D reconstruction provides an efficient paradigmfor scene modeling from image sequences. Scaling these models to largemonocular scenarios are constrained by excessive GPU memory foot-print, degraded local geometry, and long-term trajectory drift. Exist-ing chunk-based optimization strategies provide limited geometric con-straints and fail to maintain global consistency over extended trajecto-ries. We present a unified framework for stable and scalable feed-forward3D reconstruction from long monocular sequences. Our approach buildson coarse-to-fine trajectory alignment augmented by lightweight geomet-ric prior injection. Distilling monocular geometric cues into the feed-forward backbone via LoRA adaptation improves depth accuracy on finestructures while preserving inference efficiency. We introduce a hybrid-weight sparse ray-field optimization that leverages high-frequency geo-metric features to guide local point-cloud refinement and enforce con-sistent inter-frame ray constraints. Unlike prior chunk-based methods,this establishes strong cross-frame geometric coupling while maintainingscalability. Finally, an efficient trajectory stitching strategy with jointray-error optimization explicitly reduces accumulated drift. Extensiveexperiments show that our approach achieves competitive trajectory ac-curacy compared with representative SLAM systems, while maintainingglobally consistent 3D reconstruction in large-scale scenarios.

Pansharpening aims to fuse a high-resolution panchromatic(PAN) image and a low-resolution multispectral (LRMS) image to pro-duce a high-resolution multispectral (HRMS) image. Recent deep modelshave achieved strong performance, yet they typically rely on large-scalepretraining and often generalize poorly to unseen real-world image pairs.Prior zero-shot approaches improve real-scene generalization but requireper-image optimization, hindering weight reuse, and the above methodsare usually limited to a fixed scale. To address this issue, we proposeG-ZAP, a generalizable zero-shot framework for arbitrary-scale pansharp-ening, designed to handle cross-resolution, cross-scene, and cross-sensorgeneralization. G-ZAP adopts a feature-based implicit neural representa-tion (INR) fusion network as the backbone and introduces a multi-scale,semi-supervised training scheme to enable robust generalization. Extensiveexperiments on multiple real-world datasets show that G-ZAP achievesstate-of-the-art results under PAN-scale fusion in both visual qualityand quantitative metrics. Notably, G-ZAP supports weight reuse acrossimage pairs while maintaining competitiveness with per-pair retraining,demonstrating strong potential for efficient real-world deployment.

3D Gaussian Splatting (3DGS) has revolutionized fast novelview synthesis, yet its opacity-based formulation makes surface extrac-tion fundamentally difficult. Unlike implicit methods built on SignedDistance Fields or occupancy, 3DGS lacks a global geometric field, forc-ing existing approaches to resort to heuristics such as TSDF fusion ofblended depth maps. Inspired by the Objects as Volumes framework [38],we derive a principled occupancy field for Gaussian Splatting and showhow it can be used to extract highly accurate watertight meshes of com-plex scenes. Our key contribution is to introduce a learnable orientednormal at each Gaussian element and to define an adapted attenuationformulation, which leads to closed-form expressions for both the normaland occupancy fields at arbitrary locations in space. We further intro-duce a novel consistency loss and a dedicated densification strategy toenforce Gaussians to wrap the entire surface by closing geometric holes,ensuring a complete shell of oriented primitives. We modify the differen-tiable rasterizer to output depth as an isosurface of our continuous model,and introduce Primal Adaptive Meshing for Region-of-Interest mesh-ing at arbitrary resolution. We additionally expose fundamental biasesin standard surface evaluation protocols and propose two more rigor-ous alternatives. Overall, our method Gaussian Wrapping sets a newstate-of-the-art on DTU and Tanks and Temples, producing complete,watertight meshes at a fraction of the size of concurrent work—recoveringthin structures such as the notoriously elusive bicycle spokes. Our projectpage is available here.

Reconstructing and tracking dynamic 3D scenes is a funda-mental challenge in computer vision. Existing methods typically decou-ple geometry from motion: static multi-view reconstruction systems as-sume a rigid world, whereas dynamic tracking frameworks rely on explicitego-motion estimation or separate object motion models. In this work,we propose Flow4R, a unified framework that treats relative scene flowas the central representation linking 3D structure, camera ego-motion,and dynamic object motion. Given a two-view input, Flow4R employs ashared Vision Transformer to predict a compact, pixel-aligned propertyset comprising 3D point positions, scene flow, pose weights, and con-fidence maps. This flow-centric formulation allows local geometry andbidirectional motion to be jointly inferred in a single feedforward pass,eliminating the need for explicit pose regression heads or complex bundleadjustment. By training jointly on static and dynamic datasets, Flow4Rachieves state-of-the-art performance on 4D reconstruction and trackingbenchmarks, demonstrating the power of the flow-centric formulation forspatiotemporal scene understanding.

Mixture-of-Experts (MoE) architectures have emerged as a powerful paradigm for scaling diffusion models in visual generation. Recent advancements have focused on adaptively allocating computational resources across diverse tokens to improve efficiency and performance. However, we identify a routing assignment problem in existing diffusion MoE frameworks: the router fails to accurately allocate more computational resources to salient tokens. Our analysis attributes this failure to the router’s reliance on noise-corrupted latent features throughout the denoising process. Such stochastic noise obscures the critical structural and textural information, thereby preventing the router from effectively distinguishing salient tokens. To address this, we propose SharpMoE, a post-training framework with a saliency-harnessing accurate routing mechanism, which utilizes clean latent features as a noise-free guidance signal for routing. By bypassing the noise-distorted inputs, SharpMoE provides the router with clear saliency guidance, enabling the identification of salient tokens even in high-noise stages. Furthermore, we introduce a trajectory routing loss to constrain the compute allocation throughout the multi-step denoising trajectory, ensuring precise resource allocation along the generation rollout. Extensive experiments demonstrate that SharpMoE serves as a versatile, plug-and-play solution that further enhances the pretrained, converged MoE models, achieving stateof-the-art performance in visual generation.

Automatically retrieving videos from large camera-trap data-sets remains challenging. Text-to-Video retrieval (TVR) methods basedon large video-language models (VLMs) have potential to retrieve eventsof interest by describing them with simple text queries. However, cur-rent methods often lack spatiotemporal understanding and do not gen-eralize well to ecological data. In this work, we introduce Prompting-MammAlps, the first camera-trap TVR benchmark, and propose a fine-grained and interpretable TVR method. Specifically, we trained a vi-sion transformer to perform spatiotemporal action localization, and con-vert its output to structured text, describing each video. Independently,ethology-inspired queries are processed by a Large-Language Model (LLM)based coding agent to parse the structured text per video and retrievevideos accordingly. We harnessed the LLM to use functions from a cus-tom parsing library to minimize the risk of LLM hallucinations and toimprove method interpretability. This retrieval approach applied on thePrompting-MammAlps benchmark achieved a set-based F1-score of 34%on a test set of 135 ecologically-relevant queries and 775 candidate videos.In comparison the best zero-shot VLM achieved a F1-score of 18%, whilealso lacking interpretability.Project page: eceo-epfl.github.io/prompting-mammalps/

Understanding instructional videos requires recognizing fine-grained actions and modeling their temporal relations, which remainschallenging for current Video Foundation Models (VFMs). This diffi-culty stems from noisy web supervision and a pervasive “static bias”,where models rely on objects rather than motion cues. To address this,we propose InstrAct, a pretraining framework for instructional videos’action-centric representations. We first introduce a data-driven strategy,Project page: https://zyyangzy.github.io/InstrAct/which filters noisy captions and generates action-centric hard negativesto disentangle actions from objects during contrastive learning. At thevisual feature level, an Action Perceiver extracts motion-relevant tokensfrom redundant video encodings. Beyond contrastive learning, we intro-duce two auxiliary objectives: Dynamic Time Warping alignment (DTW-Align) for modeling sequential temporal structure, and Masked ActionModeling (MAM) for strengthening cross-modal grounding. Finally, weintroduce the InstrAct Bench to evaluate action-centric understanding,where our method consistently outperforms state-of-the-art VFMs onsemantic reasoning, procedural logic, and fine-grained retrieval tasks.

Models for image representation learning are typically de-signed for either recognition or generation. Various forms of contrastivelearning help models learn to convert images to embeddings that areuseful for classification, detection, and segmentation. On the other hand,models can be trained to reconstruct images with pixel-wise, perceptual,and adversarial losses in order to learn a latent space that is compati-ble with image generation. We seek to unify these two directions with afirst-of-its-kind model that learns representations which are simultane-ously well-suited for recognition and generation. We train our model asa hyper-network for implicit neural representation, which learns to mapimages to model weights for fast, accurate reconstruction. We furtherintegrate our INR hyper-network with knowledge distillation to improveits generalization and performance. Beyond the novel training design, themodel also learns an unprecedented compressed embedding space withoutstanding performance for various visual tasks. The complete modelcompetes with state-of-the-art results for image representation learning,enables downstream generative capabilities, and produces high-qualitytiny embeddings.

Intermediate representations are key to bridging the modal-ity gap between generalizable manipulation policies and large-scale pre-trained vision-language models (VLMs). Among these, trajectory-basedrepresentations compactly represent motion-relevant cues, yet most exist-ing approaches predict trajectories in 2D image space, resulting in intrin-sic 3D ambiguity. Moreover, using 2D trajectories with depth still leavesthe free-space waypoints ambiguous, limiting reliable 3D reasoning. Toaddress this, we propose predicting 3D consistent waypoints (3DWay)from multi-view images. By reformulating 3D waypoints prediction asgenerating multi-view consistent 2D waypoints followed by geometric tri-angulation, we enable explicit 3D motion specification while preservingthe strong priors of pretrained VLMs. The predicted waypoints can guideexisting VLA models for better generalization or be directly executed onsimple tasks. Extensive experiments show that 3DWay substantially im-proves 3D spatial grounding and vision-language reasoning, demonstrat-ing strong potential for generalizable robot manipulation. Codes will bereleased at https://github.com/ziqin-h/3DWay.

Medical image segmentation relies on the ability of encoder-decoderarchitectures to translate rich feature representations into accurate pixel-level pre-dictions under challenging conditions such as low contrast, structural ambigu-ity, and scale variability. While recent advances in large-scale pretraining andtransformer-based encoders have substantially improved feature extraction, seg-mentation accuracy remains constrained by decoder design, particularly in termsof cross-scale alignment, contextual integration, and boundary preservation. Inthis work, we revisit medical image segmentation from a decoder-centric per-spective and propose a context-aware gated decoder that systematically regulatesfeature fusion and contextual aggregation throughout the decoding process. Theproposed decoder integrates lightweight multi-scale channel recalibration, gatedskip fusion with spatial competition and a global context aggregation mechanismthat injects encoder-wide information into intermediate decoding stages. This de-sign enables effective translation of strong pretrained encoder representations intospatially consistent predictions. Extensive experiments across 11 medical imagesegmentation benchmarks validate the effectiveness and demonstrate that the pro-posed approach consistently outperforms strong baselines while remaining com-putationally practical. Code: https://github.com/saadwazir/MedCAGD

Existing Vision Foundation Model (VFM)-based iterativestereo pipelines under-exploit three information pathways: multi-scalebackbone features are collapsed into single-level correlations, geometricpriors remain untapped at initialization, and context propagates only lo-cally. These gaps widen under degraded photometric cues, making under-water scenes a stringent generalization test. To address this, we proposeLinStereo, built upon Depth Anything V3, whose core is a Position-Aware Linear Attention (PALA) module that replaces local recurrencewith global aggregation at linear cost, propagating reliable estimatesfrom well-matched regions into degraded areas while preserving dispar-ity structure. PALA is made effective by two enabling components: Hi-erarchical Semantic Cost Volumes (HSCV), which supply scale-alignedcorrelations from the VFM feature hierarchy, and a Depth Prior Ini-tialization (DPI) that converts monocular depth into a metrically cali-brated warm start. LinStereo achieves state-of-the-art-level accuracy onstandard benchmarks and strong cross-domain generalization, particu-larly on underwater scene where severe photometric degradation makesstereo matching particularly challenging, attaining the best overall ac-curacy with consistent gains (28% lower AbsRel on TartanAir-UW,26% on SQUID, a real-world underwater dataset). Code is availableat https://u7079256.github.io/LinStereo/.