ECCV 2026 Papers

Reinforcement learning (RL) has become a standard tech-nique for post-training diffusion-based image synthesis models, as itenables learning from reward signals to explicitly improve desirable as-pects such as image quality and prompt alignment. In this paper, wepropose an online RL variant that reduces the variance in the modelupdates by sampling paired trajectories and pulling the flow velocityin the direction of the more favorable image. Unlike existing methodsthat treat each sampling step as a separate policy action, we considerthe entire sampling process as a single action. We experiment with bothhigh-quality vision language models and off-the-shelf quality metrics forrewards, and evaluate the outputs using a broad set of metrics. Ourmethod converges faster and yields higher output quality and promptalignment than previous approaches.

Bird’s-eye-view (BEV) perception is a core component of camera-based 3D understanding in autonomous driving, where view transformation (VT) maps multi-camera image features into a uni(cid:28)ed BEV representation. Sampling-based view transformation (Sampling-VT) is attractive because it supports dense and continuous BEV aggregation for high-resolution and long-range perception. Its deployment bottleneck, however, is systems-level: standard tensorized implementations of Sampling-VT(cid:22)which we refer to as Tensorized Sampling-VT(cid:22)explicitly materialize large height-dependent intermediate tensors, causing memory and latency costs that scale poorly with vertical resolution and the number of cameras. We revisit Tensorized Sampling-VT from an operator-execution perspective and show that it follows a gather(cid:21)reduction pattern: each BEV query independently accumulates contributions across cameras and height bins. Unlike splatting-based VT, which requires index sorting and prevents fully thread-local reduction from voxel construction to BEV output, the gather(cid:21)reduction structure of Sampling-VT enables thread-local accumulation with on-the-(cid:29)y recomputation, eliminating the need to materialize heightand camera-dependent intermediates. Based on this insight, we propose FlashBEV, a fully fused and IOaware execution strategy that is mathematically equivalent to Tensorized Sampling-VT (same operator output) while substantially reducing global memory tra(cid:30)c and kernel-launch overhead. Experiments show that FlashBEV achieves more than an order of magnitude lower peak GPU memory and signi(cid:28)cant inference-latency speedups, with memory usage effectively independent of the number of height bins, reducing the operator’s peak memory to O(BCXY ) (output only). This unlocks higher BEV range/resolution and vertical discretization within (cid:28)xed deployment budgets on memory-constrained devices, where tensorized execution would otherwise be infeasible. Our contribution is therefore an execution redesign(cid:22)same math, di(cid:27)erent execution(cid:22)that removes a key scalability barrier for deployment-ready Sampling-VT. Code is available at https://github.com/yokosyun/FlashBEV. • •

Understanding human hand motion requires reasoning thatgoes beyond coarse action semantics to capture detailed finger articula-tion and interaction dynamics. However, existing Video-MLLMs primar-ily capture global action intent and often fail to represent fine-grainedfinger-level movements and contact semantics. In this work, we proposeFine-grained Finger-level Hand Motion Captioning (FingerCap), whichaims to generate textual descriptions that capture detailed finger-levelsemantics of hand actions. To support this task, we curate FingerCap-40K, a large-scale corpus of 40K paired hand-motion videos and captionsspanning two complementary sources: concise instruction-style finger mo-tions and diverse, naturalistic hand–object interactions. Furthermore,we design HandJudge, an LLM-based rubric that measures finger-level correctness and motion completeness for effective evaluation. Toestablish a strong yet lightweight baseline under the FingerCap task,we introduce FiGOP (Finger Group-of-Pictures), a module tailoredfor Video-MLLM settings with sparse RGB sampling. FiGOP augmentskeyframes with intermediate 2D hand keypoints, enabling complemen-tary motion cues without increasing the density of RGB frames. Exper-iments on FingerCap-40K show that strong open- and closed-sourceVideo-MLLMs still struggle with finger-level reasoning, while our FiGOP-enhanced model yields consistent gains under both HandJudge and hu-man studies. The dataset and benchmark are available at § FingerCap.

The rapid advancement of diffusion models has exacerbatedprivacy risks by enabling realistic manipulations of personal facial im-ages. Existing proactive defenses primarily target model-specific internalmodules under restrictive white-box assumptions, rendering their pro-tective efficacy highly vulnerable to unseen architectures and manipula-tion paradigms. To bridge this gap, we propose FaceArmor, a universalfacial image protection framework against both unknown generative ar-chitectures and real-world environmental transformations. Unlike priormodel-dependent approaches, FaceArmor focuses on the intrinsic fea-ture composition of facial images. It leverages an ensemble of surrogateextractors to decompose representations into a structural tier targetinglatent reconstruction and fine textures, and a semantic tier obscuringprompt alignment and biometric identities. These complementary di-mensions are universally exploited by mainstream diffusion-based ma-nipulation methods. FaceArmor then employs an attention-based softweighting scheme to generate effective adversarial perturbations acrossthese heterogeneous features. By extracting intrinsic attention heatmapsdirectly from the surrogate models, this strategy smoothly decouplescompeting gradients and resolves spatial optimization conflicts. Exten-sive experiments demonstrate that FaceArmor not only preserves visualimperceptibility but also achieves exceptional zero-knowledge transfer-ability, robustly thwarting diverse malicious manipulations across state-of-the-art architectures.

Fabric moiré is a sampling-induced aliasing artifact caused by the interaction between fine textile patterns and camera sensor grids, producing structured interference that severely degrades image quality. Unlike screen-induced moiré, which stems from strictly periodic display lattices, fabric moiré is intrinsically more challenging due to the broadband and semi-periodic nature of textile weaves. The heavy spectral overlap between intrinsic texture and aliasing components renders fabric demoiréing substantially more ill-posed. Consequently, existing models trained on screen moiré datasets generalize poorly to these complex textile patterns. Despite its practical importance, fabric image demoiréing remains underexplored and lacks standardized benchmarks. We present the first comprehensive benchmark for fabric image demoiréing. To address the difficulty of acquiring pixel-aligned real-world pairs, we develop a physically motivated synthesis framework and construct a large-scale dataset comprising 16,050 paired multi-resolution fabric images with controllable aliasing severity. Furthermore, we customize a baseline model, which establishes promising performance on the proposed benchmark dataset with strong generalization ability. Our benchmark provides a standardized platform for advancing research in fabric image demoiréing.

Modern vision generators transport a base distribution to data through time-indexed measures, implemented as deterministic flows (ODEs) or stochastic diffusions (SDEs). Despite strong empirical performance, standard flow-matching objectives do not directly control the information geometry of the trajectory, allowing low-entropy bottlenecks that can transiently deplete semantic modes. We propose Entropy-Controlled Flow Matching (ECFM): a constrained variational principle over continuityequation paths enforcing a global entropy-rate budget d H(µ ) ≥ −λ. dt t ECFM is a convex optimization in Wasserstein space with a KKT/Pontryagin system, and admits a stochastic-control representation equivalent to a Schrödinger bridge with an explicit entropy multiplier. In the pure transport regime, ECFM recovers entropic OT geodesics and Γ-converges to classical OT as λ → 0. We further obtain certificate-style mode-coverage and density-floor guarantees with Lipschitz stability, and construct nearoptimal collapse counterexamples for unconstrained flow matching.

Modeling the cognitive and experiential factors of humannavigation is central to deepening our understanding of human–environmentinteraction and to enabling safe social navigation and effective assis-tive wayfinding. Most existing methods focus on forecasting motionsin fully observed scenes and often neglect human factors that capturehow people feel and respond to space. To address this gap, we proposeEgoCogNav, a multimodal egocentric navigation framework that jointlyforecasts perceived path uncertainty, trajectories and head motion fromegocentric video, gaze, and motion history. To facilitate research in thefield, we introduce the Cognition-aware Egocentric Navigation (CEN)dataset consisting of 6 hours real-world egocentric recordings capturingdiverse navigation behaviors in real-world scenarios. Experiments showthat EgoCogNav learns the perceived uncertainty that strongly correlateswith human-like behaviors such as scanning, hesitation, and backtrack-ing while improving trajectory and head-motion forecasting on held-outnavigation recordings. Project page: https://calvinzqiu.github.io/egocognav-project/

We consider the problem of regenerating 3D objects from2D images and initial 3D shapes. Most 3D generators operate in a one-shot fashion, converting text or images to a 3D object with limited con-trollability. We introduce instead MeshReGen, a 3D regenerator that isconditioned on an initial 3D shape. This conceptually simple formula-tion allows us to support numerous useful tasks, including 3D enhance-ment, reconstruction, and editing. MeshReGen uses a new condition-ing mechanism based on VecSet, which allows the regenerator to up-date or improve the input geometry with consistent fine-grained details.MeshReGen learns a widely applicable regeneration prior from o!-the-shelf 3D datasets via self-supervised pretext tasks and augmentations,without additional annotations. We evaluate both the geometric consis-tency and fine-grained quality of MeshReGen, achieving state-of-the-artperformance in controllable 3D generation across several tasks.Query pointsConditional ImagesCond.(a) Architecture Latents !! PE❌…3D Coarse Cond.DINO V2 OutputDiT blocks⟳ ×#Zero-InitVAE Proj. VAEEncoder C DecoderCross attnSelf attnMLP… … …RefinedLatents…Random(b) Applications LatentsCoarse Mesh Enhancement Partial Point Cloud Reconstruction Masked Mesh EditingFig. 2: Overview. MeshReGen takes both 2D image and initial 3D geometry as input,enabling explicit control over global geometry (e.g., pose, coarse shape) while improvingfine-grained details. The 3D condition is encoded as VecSet latents (z L ) that compactlyrepresent the global geometry. After taking the sum with the positional embeddings,these conditioning and random latents are di!used by a DiT into enhanced latents andthen decoded into a complete, high-quality 3D shape.

To serve as a scalable data source for embodied AI, worldmodels should act as true simulators that infer interaction dynamicsstrictly from user actions, rather than mere conditional video genera-tors relying on privileged future object states. In this context, egocentricHand Object Interaction (HOI) world models are critical for predictingphysically grounded first-person rollouts. However, building such modelsis profoundly challenging due to rapid head motions, severe occlusions,and high-DoF hand articulations that abruptly alter contact topolo-gies. Consequently, existing approaches often circumvent these physicschallenges by resorting to conditional video generation with access toknown future object trajectories. We introduce EgoHOI, an egocentricHOI world model that breaks away from this shortcut to simulate pho-torealistic, contact-consistent interactions from action signals alone. Toensure physical accuracy without future-state inputs, EgoHOI distillsgeometric and kinematic priors from 3D estimates into physics-informedembeddings. These embeddings regularize the egocentric rollouts towardphysically valid dynamics. Experiments on the HOT3D dataset demon-strate consistent gains over strong baselines, and ablations validate theeffectiveness of our physics-informed design.

We propose DOME-HDR, a dual-output multi-exposure HDRreconstruction framework that jointly produces a perceptually balancedSDR image and a consistent HDR image via gain map inverse tone map-ping. Given three bracketed LDR inputs, DOME-HDR first synthesizesa base SDR using a LoRA-adapted latent diffusion model. A dual cross-attention fusion module injects complementary structural and color cuesfrom the under- and over-exposed images while anchoring on the mid ex-posure for stability. The synthesized SDR then guides HPGM, our HDRPrior-guided Gain Map network, to predict a spatially varying gain mapfor reliable dynamic-range expansion. We evaluate on Kalantari, Tel, andChallenge123 using both full-reference and no-reference metrics, whereDOME-HDR achieves state-of-the-art HDR reconstruction quality; abla-tions further confirm the effectiveness of dual cross-attention and SDR-guided gain map estimation.

Taxonomic classification of ecological families, genera, andspecies underpins biodiversity monitoring and conservation. Existing com-puter vision methods typically address fine-grained recognition and long-tailed learning in isolation. However, additional challenges such as spa-tiotemporal domain shift, hierarchical taxonomic structure, and previ-ously unseen taxa often co-occur in real-world deployment, leading tobrittle performance under open-world conditions. We propose TaxoNet,an embedding learning framework with a theoretically grounded dual-margin objective that reshapes class decision boundaries under classimbalance to improve fine-grained discrimination while strengtheningrare-class representation geometry. We evaluate TaxoNet in open-worldsettings that capture co-occurring recognition challenges. Leveraging di-verse plant datasets, including Google Auto-Arborist (urban tree im-agery), iNaturalist (Plantae observations across heterogeneous ecosys-tems), and NAFlora-Mini (herbarium collections), we demonstrate thatTaxoNet consistently outperforms strong baselines, including multimodalfoundation models.

Forecasting future scene evolutions is crucial for autonomousdriving safety. Most existing driving world models focus on appearancegeneration and fail to model geometric and semantic evolutions. In thispaper, we propose a native 3D world model, Doe-2, to forecast the com-prehensive scene evolutions in a unified 3D latent representation space.This representation contains both appearance, geometry, and semanticinformation and can be decoded into multi-view RGB, depth, semantics,and 3D occupancy. For efficiency, we adopt an autoregressive transformerarchitecture to generate all latents for the next scene in one pass and thenuse a lightweight diffusion head to refine them for better fidelity. Operat-ing on the compact 3D latent space, our model supports efficient futurerollouts conditioned on different actions. Extensive experiments on thenuScenes dataset show the ability of Doe-2 to efficiently and comprehen-sively forecast future driving scenes. Despite being trained from scratchwith only 700 scenes, Doe-2 already shows high-quality scene generationand good generalization with action control ability.

The deployment of Industrial Anomaly Detection (IAD) inreal-world manufacturing frequently encounters a challenging cold-startbottleneck, in which limited normal samples fail to represent the fullnormal distribution and only a few anomalies are available. Under sucha regime, existing methods struggle to form compact normal boundariesand fail to effectively exploit supervised signals from rare defects. To ad-dress this challenge, we propose Anomaly-Rectified Cold-start AD (Ar-cAD), a plug-and-play calibration framework for reconstruction-basedIAD baselines. ArcAD follows a push–pull learning paradigm to constructa compact and discriminative normal boundary under data scarcity. Onthe one hand, ArcAD projects limited normal samples onto a hypersphereand pulls them into multiple compact clusters to maximize coverage ofthe normal manifold. On the other hand, it synthesizes pseudo-anomalieson the hypersphere and leverages real anomalies to push the boundaryinward and sharpen anomaly discrimination. Extensive experiments onMVTec-AD, VisA, Real-IAD, and MANTA demonstrate that ArcADsignificantly outperforms state-of-the-art supervised and unsupervisedmethods in both single-class and multi-class settings under cold-startconditions. Code is available at: https://github.com/LGC-AD/ArcAD.

Despite remarkable progress in text-guided image editing, gen-erative models frequently fail to preserve visual object consistency, definedas the preservation of a subject’s key attributes throughout the editingprocess. We address this limitation through three contributions. First, weintroduce ABO-Edit, a dataset specifically designed to study object con-sistency, comprising over 12,000 triplets of source images, editing prompts,and high-quality target images rendered from artist-designed 3D assets,with multi-view coverage and human-verified quality control. Second, weuncover an overlooked property of image-editing rectified flow models: theconditioning embedding space, not directly supervised during training,encodes a prediction of the final generated image even at high noise levels.Third, exploiting this finding, we propose FlowMirror, a parameter-freeauxiliary loss that supervises this conditioning embedding space. Withoutarchitectural changes, our method improves generation quality acrossseveral metrics over baselines. Page: francescotaioli.github.io/DiTailed

Multimodal Large Language Models (MLLMs) have beengrowing the capability for scienti_x001C_c writing and collaboration. For exam-ple, OpenAI Prism is a free workspace for scienti_x001C_c writing and collab-oration. One important feature in Prism is turning scienti_x001C_c diagramsdirectly into LATEX TikZ code. In this paper, we build a benchmark,Diagram-MMU, a multi-modal benchmark designed to assess MLLMs'ability for scienti_x001C_c diagram parsing and understanding. Diagram-MMUfeatures 3.7k curated diagrams and 18.3k human-validated questionsacross six domains. It evaluates MLLMs on three tasks common in vibewriting workspaces: diagram-to-code parsing, diagram-to-code editing,and diagram question answering, alongside agentic settings per task.The evaluation of 12 MLLMs reveals that diagram-to-code tasks aremore challenging than diagram question answering: models can reasonwell over diagrams but struggle to parse and edit them, underscoringthe need for methods to enhance MLLMs' capability in diagram-to-codegeneration. Under agentic settings, most models improve parsing andediting performance but degrade on question answering, while Claude-4.6 Opus consistently improves across all three tasks. Our benchmark ispublicly available at https://huggingface.co/datasets/AIGrounding/Diagram-MMU.

Generalizable animal Re-Identification (ReID) aims to recognize individual animals across species with diverse morphologies and ecological contexts. Unlike person ReID, where different domains share similar body structures, animal species often exhibit drastically different anatomical structures and visual patterns, making it difficult to establish shared visual correspondences. As a result, representations learned across species tend to form fragmented embedding spaces, which severely limits cross-species generalization. To address this challenge, we propose Semantic Consistency Learning (SCL), a framework designed to learn representations that remain stable across appearance variations while preserving semantic structures shared across species. SCL consists of two complementary components. Foreground–Background Decoupled Spectral Normalization (FDSNorm) stabilizes feature statistics by suppressing environment-induced style variations in a region-aware manner, while Cross-species Neighborhood Modeling (CNM) captures transferable relational structures across species through dynamic feature neighborhoods. Extensive experiments on 11 public animal ReID datasets demonstrate that SCL consistently outperforms state-of-the-art methods under multiple cross-species evaluation protocols and generalizes effectively to previously unseen species and ecological domains. Code is available at https://github.com/Kemalau/ECCV-26-SCL.

Generative video models have significantly advanced the photorealistic synthesis of adverse weather for autonomous driving; however, they consistently demand massive datasets to learn rare weather scenarios. While 3D-aware editing methods alleviate these data constraints by augmenting existing video footage, they are fundamentally bottlenecked by costly per-scene optimization and suffer from inherent geometric and illumination entanglement. In this work, we introduce AutoWeather4D, a feed-forward 3D-aware weather editing framework designed to explicitly decouple geometry and illumination. At the core of our approach is a G-buffer Dual-pass Editing mechanism. The Geometry Pass leverages explicit structural foundations to enable surface-anchored physical interactions, while the Light Pass analytically resolves light transport, accumulating the contributions of local illuminants into the global illumination to enable dynamic 3D local relighting. Extensive experiments demonstrate that AutoWeather4D achieves comparable photorealism and structural consistency to generative baselines while enabling finegrained parametric physical control, serving as a practical data engine for autonomous driving.

Recent advances in motion-aware large language models haveshown remarkable promise for jointly learning motion understandingand generation knowledge. However, these models typically treat un-derstanding and generation separately, limiting the mutual benefits thatcould arise from interactive feedback between tasks. In this work, wereveal that motion assessment and refinement tasks can act as crucialbridges to enable knowledge flow from motion understanding to genera-tion. Specifically, we propose Interleaved Reasoning for Motion Genera-tion (IRMoGen), a novel paradigm that tightly couples motion genera-tion with assessment and refinement through iterative text-motion dia-logue. To realize this, we introduce IRG-MotionLLM, the first model thatseamlessly interleaves motion generation, assessment, and refinement toimprove the alignment between generated motion and goal text. IRG-MotionLLM is developed progressively with a novel three-stage train-ing scheme, initializing and subsequently enhancing native IRMoGencapabilities. To facilitate this development, we construct an automateddata engine to synthesize interleaved reasoning annotations from existingtext-motion datasets. Extensive experiments demonstrate the propertiesbrought by IRMoGen training, and the advanced cross-benchmark andcross-evaluator performance of IRG-MotionLLM. Code and models areavailable at https://github.com/HumanMLLM/IRG-MotionLLM.

While 3D Vision-Language Models (3D VLMs) have demon-strated remarkable spatial reasoning capabilities, they suffer from mas-sive visual token counts that create severe computational bottlenecksduring inference. Existing token pruning methods primarily rely on di-versity-based selection, discarding similar tokens to maximize dispersion.However, in 3D environments, this approach frequently drops represen-tative prototype tokens in favor of outliers, breaking the multi-view con-sistencies and geometric structures essential for spatial reasoning. In thispaper, we propose a paradigm shift for 3D VLM token pruning: frommaximizing diversity to preserving visual evidence coverage. We intro-duce CoverPrune, a training-free framework that formulates inference-time token pruning as an Optimal Transport (OT) problem. To over-come the intractable combinatorial subset selection inherent in this for-mulation, we design the Feature-Spatial-Temporal (FST) transport costand target capacity, along with an efficient Spatial-Guided Greedy Selec-tion (SGS) algorithm to approximate the OT objective. Furthermore, wepropose CoverPrune-Lite, an accelerated variant utilizing spatially struc-tured local matching for minimal overhead. Extensive experiments acrossmultiple 3D visual-spatial reasoning benchmarks demonstrate that ourmethods achieve state-of-the-art token efficiency, maintaining robust rea-soning performance even under highly aggressive pruning budgets. Visitour project website at https://github.com/Brucess/CoverPrune.