ECCV 2026 Papers

Latent diffusion models (LDMs) achieve state-of-the-art im-age synthesis, yet their reconstruction-style denoising objective providesonly indirect semantic supervision: high-level semantics emerge slowly,requiring longer training and limiting sample quality. Recent works injectsemantics from Vision Foundation Models (VFMs) either externally viarepresentation alignment or internally by jointly modeling only a nar-row slice of VFM features inside the diffusion process, under-utilizingthe rich, nonlinear, multi-layer spatial semantics available. We intro-duce REGLUE (Representation Entanglement with Global-Local UnifiedEncoding), a unified latent diffusion framework that jointly models (i)VAE image latents, (ii) compact local (patch-level) VFM semantics, and(iii) a global (image-level) [CLS] token within a single SiT backbone.A lightweight convolutional semantic compressor nonlinearly aggregatesmulti-layer VFM features into a low-dimensional, spatially structuredrepresentation, which is entangled with the VAE latents in the diffusionprocess. An external alignment loss further regularizes internal repre-sentations toward frozen VFM targets. On ImageNet 256×256, REGLUEconsistently improves FID and accelerates convergence over SiT-B/2 andSiT-XL/2 baselines, as well as over REPA, ReDi, and REG. Extensive exper-iments show that (a) spatial VFM semantics are crucial, (b) non-linearcompression is key to unlocking their full benefit, and (c) global tokensand external alignment act as complementary, lightweight enhancementswithin our global–local–latent joint modeling framework. The code isavailable at https://github.com/giorgospets/reglue.

We propose Ref-GeNVS, a training-free, reflection-awaremethod for generative novel view synthesis (NVS) in mirror scenes. Ex-isting multi-view diffusion models often fail to recognize the mirror in thescene and cannot exploit reflected content for scene generation. To fix thisissue without additional training, our key idea is to treat a mirror imageas two complementary views. From input images, we estimate the mirrorplane and reflect camera poses to form virtual views. Based on this vir-tual view setup, we propose a two-stage generation method consisting ofMirror-gated attention and Reflection injection, which enables reflection-consistent NVS by explicitly leveraging reflection relationships in a multi-view diffusion model. Ref-GeNVS inherits the strong generalizability ofthe multi-view diffusion backbone, while it does not require finetuning.On synthetic and real scenes including mirrors, Ref-GeNVS outperformsrecent generative NVS methods by generating reflection-consistent andcontextually coherent novel views, revealing scene structure visible onlythrough mirrors. Project page: https://kim-geonu.github.io/Ref-GeNVS/

Current multimodal re_x001D_ection mechanisms for long video un-derstanding predominantly rely on closed-loop self-re_x001D_ection within in-ternal parameters. Lacking objective external evidence, models are fre-quently trapped in blind con_x001C_dence and often fail to correct errors.Furthermore, applying reinforcement learning to multi-stage re_x001D_ectionpipelines introduces severe policy coupling, which is exacerbated by acritical scarcity of dedicated training data. To address these limitations,this work proposes Re_x001D_ect-R1, the _x001C_rst Evidence-Driven self-correctionframework for long video understanding. The framework constructs athree-stage pipeline consisting of intuition, veri_x001C_cation, and arbitration.By dynamically retrieving objective visual evidence to verify initial in-tuitions and autonomously executing multiple temporal searches to re-solve con_x001D_icts, it completely breaks the hallucination loop. To over-come policy coupling, we design a stage-decoupled reinforcement learn-ing algorithm named SD-GRPO that independently computes advantagefunctions across di_x001B_erent reasoning stages. Concurrently, we constructa dataset of 120K samples to bridge the training data gap. Extensiveexperiments on benchmarks such as VideoMME and LongVideoBenchdemonstrate that Re_x001D_ect-R1 achieves state-of-the-art performance. Ourmethod signi_x001C_cantly improves the genuine recti_x001C_cation rate and enablesauthentic self-correction strictly grounded in objective evidence.

Digital cameras digitize scene light into linear raw representations,which the image signal processor (ISP) converts into display-ready outputs. Whileraw data preserves full sensor information—valuable for editing and vision tasks—formats such as Digital Negative (DNG) require large storage, making them im-practical in constrained scenarios. In contrast, JPEG is a widely supported format,offering high compression efficiency and broad compatibility, but it is not well-suited for raw storage. This paper presents Raw-JPEG Adapter, a lightweight,learnable, and invertible pre-processing pipeline that adapts raw images for stan-dard JPEG compression. Our method applies spatial and optional frequency-domain transforms, with compact parameters stored in the JPEG comment field,enabling accurate raw reconstruction. Experiments across multiple datasets showthat our method achieves higher fidelity than direct JPEG storage, supports othercodecs, and provides a favorable trade-off between compression ratio and recon-struction accuracy.

Photographs of people taken by professional photographerstypically present the person in beautiful lighting, with an interestingpose, and flattering quality. This is unlike common photos people takeof themselves in uncontrolled conditions. In this paper, we explore howto canonicalize a person’s "in-the-wild" photograph into a controllable,high-fidelity avatar—reposed in a simple environment with standard-ized minimal clothing. A key challenge is preserving the person’s uniquewhole-body identity, facial features, and body shape while stripping awaythe complex occlusions of their original garments. While a large paireddataset of the same person in varied clothing and poses would simplifythis, such data does not exist. To that end, we propose two key insights:1) Our method transforms the input photo into a canonical full-body UVspace, which we couple with a novel reposing methodology to model oc-clusions and synthesize novel views. Operating in UV space allows us todecouple pose from appearance and leverage massive unpaired datasets.2) We personalize the output photo via multi-image finetuning to ensurerobust identity preservation under extreme pose changes. Our approachyields high-quality, reposed portraits that achieve strong quantitativeperformance on real-world imagery, providing an ideal, clean biometriccanvas that significantly improves the fidelity of downstream applicationslike Virtual Try-On (VTO).

Existing world model-based planners for visual navigationtypically follow a verification-centric paradigm, decoupling goal intentfrom trajectory synthesis. This approach suffers from candidate depen-dence, heavy computational overhead, and inconsistencies between sam-pled actions and predicted visuals. To address these issues, we proposeSWAM (Spatial-perceiving World Action Model), a task-centric jointobservation-action generation framework. Given start and goal RGB ob-servations, SWAM performs single-pass inference to simultaneously gen-erate intermediate RGB-D sequences and corresponding action trajecto-ries, promoting goal-consistent trajectory generation and improved spa-tial feasibility. SWAM leverages depth pseudo-labels during training tointernalize spatial priors, but it requires only monocular RGB input atinference time by using depth estimation model. We further introducea visual-guided action refinement module and a trajectory-scale regular-ization loss to enforce fine-grained alignment between motion and visualcues while stabilizing predictions across varying distances. Extensive ex-periments show that SWAM significantly outperforms state-of-the-arttwo-stage planners in success rate, trajectory accuracy, and inference ef-ficiency, while demonstrating robust zero-shot generalization to unseenenvironments.

Recent text-to-image (T2I) models generate semantically co-herent images from textual prompts, yet evaluating how well they alignwith individual user preferences remains an open challenge. Conventionalevaluation methods, general reward functions or similarity-based metrics,fail to capture the diversity and complexity of personal visual tastes. Wepresent PIGReward, a personalized reward modeling framework that dy-namically generates user-conditioned evaluation dimensions and assessesimages through CoT reasoning. To address the scarcity of user data, PI-GReward adopts a self-bootstrapping strategy that reasons over limitedreference data to construct rich user contexts, enabling personalizationwithout user-speci_x001C_c training. Beyond evaluation, PIGReward providespersonalized feedback that drives user-speci_x001C_c prompt optimization, im-proving alignment between generated images and individual intent. Ex-tensive experiments demonstrate that PIGReward outperforms existingmethods in both accuracy and interpretability, establishing a scalable,reasoning-based foundation for personalized T2I evaluation and opti-mization. Project page: https://jeongeunnn-e/PIGReward/

Despite the remarkable progress of Large Multimodal Models(LMMs), deploying autonomous agents to navigate web Graphical UserInterfaces (GUIs) remains a significant challenge. Most existing agentsare “blind” when encountering unfamiliar websites, as they rely heav-ily on patterns memorized during in-domain training, which inevitablyfails in the open-world web. To bridge this generalization gap, we pro-pose that agents should mimic human behavior: leveraging external ex-pertise to navigate unknown environments. In this paper, we introduceWebOne, a novel benchmark designed to evaluate an agent’s ability to mas-ter unseen websites by referencing Multimodal Tutorials—heterogeneousknowledge sources derived from instructional videos, historical trajec-tories, and human demonstrations. WebOne comprises 1,342 real-worldtasks and 970 high-quality tutorials across 60+ websites. Building uponthis, we propose WebLearner, a reinforcement learning-based frameworkthat utilizes a hierarchical referencing strategy to synthesize informa-tion from these multimodal tutorials. Experimental results demonstratethat WebLearner significantly outperforms current state-of-the-art open-source models, achieving a 56.9% success rate on held-out websites andshowing competitive performance against recent proprietary models.

The training-free integration of expert models via modelmerging has exposed significant security risks, enabling free-riders tocombine specialized models without authorization. Recent works pro-pose parameter-level defenses that employ linear parameter transforma-tions to neutralize this threat. In this paper, we systematically analyzesuch defenses and reveal that their protected task vectors are inherentlysmall in magnitude. Consequently, the protected weights remain over-whelmingly dominated by the pretrained model. Based on this observa-tion, we designate the pretrained model as a static reference anchor andpropose the Anchor-Guided Attack (AGA) to circumvent existing safe-guards. Specifically, AGA aligns the protected model with this anchorto recover the transformation matrix analytically. Extensive evaluationsvalidate that AGA consistently bypasses both individual and compositedefenses under realistic defense-agnostic scenarios. Furthermore, we pro-vide Anchor-Repulsive Fine-tuning (ARF), a defense method to mitigatethe anchor dominance leveraged by AGA. Empirical results confirm thatARF effectively defeats the proposed attack. Our code is available athttps://github.com/krumpguo/secure-merge-attack.

Recent advances in video diffusion models have enabled either long single-view generation through temporal autoregression, or short multi-view synthesis through bidirectional attention. However, generating long, multi-view consistent videos of dynamic scenes remains unsolved. In this work, we present MV-Forcing, a framework that composes temporal and view-wise autoregression within a single diffusion model by introducing a 4D geometric bridge between sequentially generated views. Our key insight is that an autoregressive 3D reconstruction model naturally interfaces between autoregressively generated views. Given a completed source view, we reconstruct its 3D structure and render a geometric prior of the next target viewpoint, which the diffusion model refines into a high-quality video. To extend generation beyond the teacher’s fixed temporal window, we introduce a joint denoising regime where both view slots are initialized from noise during training, enabling temporally unbounded generation. We distill the model via Distribution Matching Distillation with Spatio-Temporal Self-Forcing, closing the train-inference exposure bias gap for both temporal and view-sequential autoregression. Extensive experiments on both synthetic and real-world data demonstrate that MV-Forcing produces geometrically consistent multi-view videos of dynamic scenes at arbitrary lengths and viewpoint counts using a single few-step student model.

Recent advancements in scaling dataset distillation rely heav-ily on decoupled information extraction pipelines, comprising Squeeze,Recover, and Relabel stages. Despite their scalability to large-scaledatasets, these methods suffer from prohibitive computational overheadand poor cross-architecture generalization. In this paper, we reveal theroot cause of these bottlenecks: the implicit dual-compression process,from data to model and back to images, inherently induces severe infor-mation loss. Crucially, we empirically and theoretically demonstrate thatthis loss creates a distribution shift that fundamentally compromises thewidely adopted Relabel strategy, transforming the pre-trained modelinto an unreliable labeler that yields sub-optimal labels. To overcomethese critical flaws, we propose CIM, a novel, metric-driven frameworkthat abandons the flawed dual-compression paradigm. Instead, CIM ex-plicitly quantifies and minimizes the information gap between the originaland synthetic datasets. By directly aligning the data distributions, ourapproach ensures high-fidelity information condensation and inherentlysatisfies the prerequisites for effective relabeling. Extensive experimentsdemonstrate that CIM establishes a new state-of-the-art. Notably, itdistills ImageNet-1K at an IPC =10 in merely 80 minutes on a singleRTX-4090 GPU, achieving an unprecedented 48.7% Top-1 accuracy onResNet-18 and significantly outperforming previous SOTA approaches,such as NRR-DD and DELT, by 2.6% and 2.9%, respectively. Our codeis available at https://github.com/LINs-lab/CIM.

Nighttime color constancy still remains a challenging prob-lem in computational photography due to low-light noise and complexillumination conditions. We present RL-AWB, a novel framework com-bining statistical methods with deep reinforcement learning for night-time white balance. Our method begins with a statistical algorithmtailored for nighttime scenes, integrating salient gray pixel detectionwith novel illuminant estimation. Building on this foundation, we de-velop the first deep reinforcement learning approach for color constancythat leverages the statistical algorithm as its core, mimicking profes-sional AWB tuning experts by dynamically determining image-specificparameters at inference time, without requiring ground-truth illumi-nants or reference images. To further facilitate cross-sensor evaluation,we introduce the first multi-sensor nighttime dataset. Experiment re-sults demonstrate that our method achieves strong generalization capa-bility across low-light and well-illuminated images. Project page: https://ntuneillee.github.io/research/rl-awb/

Foundation models have transformed vision and languageprocessing by providing rich, reusable representations that transfer acrossdiverse tasks. Sheet music, as a visual encoding of musical language,lacks such a strong domain-specific backbone. We introduce MuSViT(Music Score Vision Transformer): the first foundation vision modelfor sheet music representation—a ViT encoder pre-trained via MaskedAutoencoders on 9.7 million pages from the International Music ScoreLibrary Project (IMSLP). To handle the complexity of real-world scores,we adopt a two-stage curriculum: a synthetic warm-up on typeset scoresfollowed by large-scale training on the full IMSLP corpus. We evaluateMuSViT on four downstream tasks—full-page and staff-level music scorerecognition, music symbol detection, and score difficulty classification—under two scenarios: linear probing (frozen encoder) and fine-tuning.Under linear probing, MuSViT consistently outperforms modern visionencoders, revealing that general-purpose representations, regardless ofscale, fall systematically short on the structured symbolic propertiesMuSViT PaliGemma 2 Qwen3-VL MuSViT SoTAMuSViT-Light Kosmos-2.5 DINOv3-7B MuSViT-LightFull-Page Music Full-Page MusicScore Recognition Score RecognitionScore Difficulty Staff-Level Music Score Difficulty Staff-Level MusicClassification Score Recognition Classification Score RecognitionMusic Symbol Music SymbolDetection DetectionFig. 2: MuSViT performance across four downstream tasks. Left: Linear probing(frozen encoder)—MuSViT (solid) consistently outperforms general-purpose vision en-coders (dashed), demonstrating superior representation quality. Right: Fine-tuning—MuSViT generally outperforms state-of-the-art methods (SoTA). Axes represent nor-malized performance on each task (higher is better); see Section 3 for detailed results.of musical notation. Under fine-tuning, MuSViT generally improvesupon task-specific state-of-the-art methods. An additional embedding-transcription consistency analysis reveals that MuSViT encodes sym-bolic musical structure directly in its representation space—unlike otherencoders, whose embeddings do not correlate with music notation con-tent. These results establish MuSViT as a foundation backbone for sheetmusic understanding.

Understanding field imagery such as detecting plants anddistinguishing individual crop and weed instances, is a central challengein precision agriculture. Despite progress in vision–language tasks likecaptioning and visual question answering, Visual Grounding (VG), local-ising language-referred objects, remains unexplored in agriculture. A keyreason is the lack of suitable benchmark datasets for evaluating groundingmodels in field conditions, where many plants look highly similar, appearat multiple scales, and the referred target may be absent from the image.To address these limitations, we introduce gRef-CW, the first dataset de-signed for generalised visual grounding in agriculture, including negativeexpressions. Benchmarking current state-of-the-art grounding models ongRef-CW reveals a substantial domain gap, highlighting their inabilityto ground instances of crops and weeds. Motivated by these findings, weintroduce Weed-VG, a modular framework that incorporates multi-labelhierarchical relevance scoring and interpolation-driven regression. Weed-VG advances instance-level visual grounding and provides a clear baselinefor developing VG methods in precision agriculture. Code and data areavailable at https://github.com/MHaghighat98/WeedVG-gRefCW.

Medical image enhancement is clinically important, but ex-isting methods often rely on large-scale datasets to learn complex image-to-image mappings. Such reliance leads to high training and storage costsand hinders practical deployment.Dataset distillation (DD) provides a promising solution by synthesizingcompact datasets that preserve the training behavior of the original data.However, existing DD methods mainly focus on high-level tasks, wheremultiple samples share a common semantic label and the distilled datacan compress shared representations. In contrast, low-level medical im-age enhancement involves dense image-to-image mappings, where eachsample corresponds to a unique pixel-level target. This fundamental dif-ference makes low-level DD significantly more challenging and inherently

We study action-conditioned world modeling as a scalableway to learn transferable dynamics priors for robot learning. By pre-training a model to predict how actions drive visual scene evolution,the resulting world model captures reusable interaction dynamics be-yond appearance-level video generation. Concretely, we pretrain a multi-view interactive base diffusion world model, A2World, on large-scalerobot manipulation data with real action annotations. We validate thelearned dynamics priors from two complementary perspectives. First,we adapt A2World into a task- or scene-specialized real-world simula-tor, A2World-sim, whose long-horizon rollouts support simulator-basedpolicy evaluation and scalable what-if analysis by replacing real-robotrollouts with world model rollouts. Second, starting from the same pre-trained weights, we adapt A2World into a video-action joint predictionmodel, A2World-policy, that predicts actions under visual and instruc-tion conditioning. Experiments across simulation benchmarks and real-robot settings demonstrate that action-conditioned world model pre-training yields transferable dynamics priors that benefit both simulator-centric and policy-centric robot learning.

This paper concerns the problem of unsupervised temporalaction segmentation for long, untrimmed videos. Recent successful ap-proaches follow a joint representation learning and clustering paradigm,where optimal transport (OT) is adopted to produce pseudo labels forlearning frame representations. These approaches alternate between es-timating pseudo labels using OT and optimizing the parameters withgradient descent during training, where OT is used for obtaining the fi-nal temporal action segmentation. A major limitation of these works isthat they learn a deterministic embedding for frame representations. Theiterative procedure between learning deterministic embeddings based onpseudo labels and estimating pseudo labels from the learned embeddingcan thus get quickly stuck in a local optimum. As an alternative, we thuspropose to learn a probabilistic embedding for frame representations. Theembeddings are modeled by Gaussian distributions and we sample fromthe distributions before estimating the pseudo labels. We evaluate ourapproach on several challenging temporal action segmentation datasetsand achieve results comparable to, and in some cases, better than thestate of the art. Compared to baselines with deterministic embeddings,our approach improves MoF up to 20.7% and F1-score up to 19.0%. Ourcode is available at https://github.com/derkbreeze/PEOT.

Generating realistic 3D hand-object interactions (HOI) is afundamental challenge in computer vision and robotics, requiring bothtemporal coherence and high-fidelity physical plausibility. Existing meth-ods remain limited in their ability to learn expressive motion repre-sentations for generation and perform temporal reasoning. In this pa-per, we present HO-Flow, a framework for synthesizing realistic hand-object motion sequences from texts and canonical 3D objects. HO-Flowfirst employs an interaction-aware variational autoencoder to encode se-quences of hand and object motions into a unified latent manifold byincorporating hand and object kinematics, enabling the representationto capture rich interaction dynamics. It then leverages a masked flowmatching model that combines auto-regressive temporal reasoning withcontinuous latent generation, improving temporal coherence. To furtherenhance generalization, HO-Flow predicts object motions relative to theinitial frame, enabling effective pre-training on large-scale synthetic data.Experiments on the GRAB, OakInk, and DexYCB benchmarks demon-strate that HO-Flow achieves state-of-the-art performance in both phys-ical plausibility and motion diversity for interaction motion synthesis.

Chain-of-Thought fine-tuning enhances the reasoning capabilities of Large Vision-Language Models, but can paradoxically exacerbate hallucinations, revealing a reasoning-faithfulness trade-off. We identify a key cause of this phenomenon as the “Confidence Trap”: under visual ambiguity or degradation, models may produce low-entropy, high-confidence errors instead of expressing appropriate uncertainty. This failure mode challenges conventional mitigation strategies that rely on uncertainty cues or contrastive decoding signals, since confident hallucinations can remain highly stable under visual perturbations. To address this problem, we propose Hallucination-Entropy Regulated Optimization (HERO), an on-policy reinforcement learning framework for mitigating confident hallucinations. HERO introduces a dynamic entropy-aware objective that assigns larger optimization weights to low-entropy erroneous outputs, and further improves training efficiency through variance-gated sample selection that focuses updates on informative hard negatives. Experiments on POPE, THRONE, and AMBER show that HERO achieves strong overall hallucination mitigation, substantially improves truthfulness in critical low-entropy regions, and maintains competitive general reasoning capabilities. These results suggest that explicitly targeting confidence-misaligned errors is an effective direction for improving multimodal faithfulness in CoT-style LVLMs.