ECCV 2026 Papers

Multimodal dataset distillation (MDD) aims to synthesize acompact image-text dataset on which a visual-language model can betrained to achieve performance comparable to that trained on the orig-inal, full-scale dataset. While previous methods have achieved promis-ing results by using trajectory matching, they fail to consider the gapbetween single-semantic texts and multi-semantic images. Additionally,they usually use contrastive losses to pull together paired samples andpenalize non-paired samples, ignoring the textual information from un-paired samples that are semantically related to the corresponding images.To reduce the semantic gap, we introduce a multi-text fusion module thatstrengthens the cross-modal interaction between the image and text. Inaddition, we leverage uncertainty to adaptively guide the contribution ofnon-paired samples in the synthetic dataset, thereby improving the ef-fective utilization of information from these samples. We then provide aninformation theoretic analysis of the semantic gap and the limitations ofcontrastive supervision. Experiments on Flickr-30K and MS-COCO showthat our method consistently outperforms previous state-of-the-art MDDmethods, achieving significant improvements in retrieval performance(e.g., +6.5% in IR@10 and +4.6% in TR@10 in a 500-pair setting).Code is available at https://github.com/yiqiqiandefanhua/P2DE.

Frontier Multimodal Large Language Models (MLLMs) ex-hibit remarkable capabilities in Visual-Language Comprehension (VLC)tasks. However, they are often deployed as zero-shot solution to newtasks in a black-box manner. Validating and understanding the behav-ior of these models become important for application to new task. Wepropose an Explicit Logic Channel, in parallel with the black-box modelchannel, to perform explicit logical reasoning for model validation, selec-tion and enhancement. The frontier MLLM, encapsulating latent vision-language knowledge, can be considered as an Implicit Logic Channel. Theproposed Explicit Logic Channel, mimicking human logical reasoning, in-corporates a LLM, a VFM, and logical reasoning with probabilistic infer-ence for factual, counterfactual, and relational reasoning over the explicitvisual evidence. A Consistency Rate (CR) is proposed for cross-channelvalidation and model selection, even without ground-truth annotations.Additionally, cross-channel integration further improves performance inzero-shot tasks over MLLMs, grounded with explicit visual evidence toenhance trustworthiness. Comprehensive experiments conducted for tworepresentative VLC tasks, i.e., MC-VQA and HC-REC, on three chal-lenging benchmarks, with 11 recent open-source MLLMs from 4 frontierfamilies. Our systematic evaluations demonstrate the effectiveness of pro-posed ELC and CR for model validation, selection and improvement onMLLMs with enhanced explainability and trustworthiness.

We introduce Elastic Looped Transformers (ELT), a highlyparameter-efficient class of visual generative models based on a recurrenttransformer architecture. While conventional generative models rely ondeep stacks of unique transformer layers, our approach employs iterative,weight-shared transformer blocks to drastically reduce parameter countswhile maintaining high synthesis quality. To effectively train these modelsfor image and video generation, we propose the idea of Intra-Loop SelfDistillation (ILSD), where student configurations (intermediate loops)are distilled from the teacher configuration (maximum training loops)to ensure consistency across the model’s depth in a single training step.Our framework yields a family of elastic models from a single trainingrun, enabling Any-Time inference capability with dynamic trade-offs be-tween computational cost and generation quality, with the same parame-ter count. ELT significantly shifts the efficiency frontier for visual synthe-sis. With 4 × reduction in parameter count under iso-inference-computesettings, ELT achieves a competitive FID of 2.0 on class-conditional Im-ageNet 256 × 256 and FVD of 72.8 on class-conditional UCF-101.Vanilla Looped Transformers Elastic Looped TransformersXk 0 Xk 0ILSDTrainingXk intXk max Xk int Xk maxSolution Space Solution Space✗ Xk int away from solution space. ✓ Xk int moves towards solution space.✗ Fixed Depth Inference ✓ Any-Time Inference.Fig. 2: Latent Trajectories of Standard vs. Elastic Looped Transformers.Xk int & Xk max represent output of intermediate (Lint ) & final loops (Lmax ) respectivelyfor kth generation sampling step. Unlike standard recurrent models (left) where onlythe final iteration Xk max reaches the solution space, our ILSD training (right) guidesintermediate states Xk int also toward the target space. This transformation shifts themodel from a fixed-depth architecture to an Any-Time inference framework, supportingflexible computational budgets through early exits within a sampling step.

Scaling video diffusion transformers is fundamentally bottle-necked by two compounding costs: the expensive quadratic complexityof attention per step, and the iterative sampling steps. In this work, wepropose EFlow, a efficient few-step training framework, that tacklesthese bottlenecks simultaneously. To reduce sampling steps, we build ona solution-flow objective that learns a function mapping a noised state attime t to time s. Making this formulation computationally feasible andhigh-quality at video scale, however, demands two complementary innova-tions. First, we propose Gated Local–Global Attention, a token-droppablehybrid block which is efficient, expressive, and remains highly stable underaggressive random token-dropping, substantially reducing per-step com-pute. Second, we develop an efficient few-step training recipe. We proposePath-Drop Guided training to replace the expensive guidance target witha computationally cheap, weak path. Furthermore, we augment this witha Mean-Velocity Additivity regularizer to ensure high fidelity at extremelylow step counts. Together, our EFlow enables a practical from-scratchtraining pipeline, achieving up to 2.5× higher training throughput overstandard solution-flow, and 45.3× lower inference latency than standarditerative models with competitive performance on Kinetics and large-scaletext-to-video datasets.

Single-Source Domain Generalization (SDG) in medical multi-organ image segmentation confronts severe challenges of distributionshifts caused by varying imaging protocols. However, we identify a crit-ical yet overlooked phenomenon in current methodologies: CatastrophicForgetting in SDG (CF-SDG). Specifically, existing generalization strate-gies often lack explicit protection for source domain features, leading toperformance degradation on the source domain while improving general-ization capabilities. To address this, we propose the Collaborative CausalLearning Network (CCL-Net), a unified framework that simultaneouslyachieves robust generalization and source knowledge preservation. Froma causal perspective, we construct a dual-path causal intervention mech-anism with source domain knowledge constraints. First, we design anOrthogonal Structure Disentanglement (OSD) module to learn struc-tural mediator representations, establishing a pure structural pathwayby blocking the intrusion of non-causal information via orthogonaliza-tion strategies. Second, we introduce an Anatomy-Guided Local CausalIntervention (AG-LCI) module to physically sever the spurious corre-lations between anatomical semantics and local appearances throughmask-guided counterfactual generation. Furthermore, to mitigate CF-SDG, a novel Collaborative Learning Constraint mechanism is designedto anchor source domain memory via dynamic local-global consistencyregularization. Extensive experiments on multi-modality and multi-organdatasets demonstrate that CCL-Net not only achieves state-of-the-artgeneralization performance on unseen domains but also maintains a sig-nificantly lower forgetting rate on the source domain.

Existing methods for Visual Document Question Answering(DocVQA) often predict answers without explicit intermediate analysisor produce ungrounded intermediate steps. The few approaches whichemploy grounding produce rectangular bounding boxes which fail to ac-curately represent complex structures such as pie-chart wedges, flowchartnodes, and curved text. We present DoCoG, a mask-based multi-typegrounded step-wise Chain-of-Thought architecture for Document QA.The model supports multi-type grounding, i.e., it localizes both textualcontent and fine-grained graphical elements such as legends, pie slices,and flowchart nodes. It generates step-wise intermediate outputs and fi-nal answers grounded through precise segmentation masks, enabling con-sistent evidence alignment. DoCoG integrates a Vision-Language Modelwith a promptable segmenter via a novel Grounding Interaction Module(GIM). During decoding, special [GND] tokens trigger mask prediction,linking each step and the final answer to supporting evidence. GIM re-fines [GND] tokens during decoding to improve spatial consistency andalignment. We introduce DoCoG-QA, a large-scale diverse dataset com-prising 325K documents and 1.5M step-wise grounded QA pairs. Wefurther release DoCoG-PQA with 20K preference pairs to improve align-ment using Direct Preference Optimization (DPO), and DoCoG-Bench,a human-verified benchmark. Experiments show that DoCoG achievesstate-of-the-art performance across multiple benchmarks, improving an-swer accuracy and grounding precision.

Dataset distillation (DD) compresses a large training set intoa small synthetic set, reducing storage and training cost, and has shownstrong results on general benchmarks. Decoupled DD further improvesefficiency by splitting the pipeline into pretraining, sample distillation,and soft-label generation. However, existing decoupled methods largelyrely on coarse class-label supervision and optimize samples within eachclass in a nearly identical manner. On fine-grained datasets, this oftenyields distilled samples that (i) retain large intra-class variation withsubtle inter-class differences and (ii) become overly similar within thesame class, limiting localized discriminative cues and hurting recognition.To solve the above-mentioned problems, we propose FD2 , a dedicatedframework for Fine-grained Dataset Distillation. FD2 localizes discrimi-native regions and constructs fine-grained representations for distillation.During pretraining, counterfactual attention learning aggregates discrim-inative representations to update class prototypes. During distillation,a fine-grained characteristic constraint aligns each sample with its classprototype while repelling others, and a similarity constraint diversifies at-tention across same-class samples. Experiments on multiple fine-grainedand general datasets show that FD2 integrates seamlessly with decou-pled DD and improves performance in most settings, indicating strongtransferability. Code is available at https://github.com/Guang000/FD2.

Current visual generation models are capable of producinghigh-quality content, yet they lack a coherent perception of the spatialstructure. Existing generative novel view synthesis methods typically in-troduce explicit geometry priors, which enforce spatial consistency butinherently restrict generalization in large view changes. In contrast, re-cent interactive generative methods favor implicit scene modeling, offer-ing greater flexibility at the cost of precise camera control and geome-try consistency. In this paper, we propose MetaView, a diffusion-basedmonocular novel view synthesis framework that enables rendering underlarge view changes from a single image. Our key insight is to combineimplicit geometry modeling with minimal yet essential explicit 3D cues:we incorporate implicit geometry priors from a feed-forward geometryperception network to regularize structure without imposing restrictivereconstruction pipelines, while leveraging metric depth to anchor the gen-eration to a metric scale. This design allows MetaView to achieve bothgeometry consistency and precise controllability. Extensive experimentsdemonstrate that, under challenging monocular large viewpoint changes,MetaView significantly outperforms existing methods and exhibits supe-rior generalization. Our code will be made publicly available.

Domain adaptive object detection (DAOD) is essential fordeploying reliable detectors in real-world environments with inevitabledomain shifts. While recent approaches rely on image-to-image transla-tion, they often suffer from instability inherent in pixel-level domaintransformation. To address these limitations, we propose DA-F2F, anovel framework that directly aligns domains within the feature space.DA-F2F introduces a style-aware feature modulation (SFM) module thatextracts style statistics from the target domain to dynamically modulatesource representations. Furthermore, to improve adversarial learning, wepropose a soft-weighted proposal alignment (SPA) module that stablyseparates foreground and background regions in the target domain. Ex-tensive experiments demonstrate that DA-F2F significantly outperformsexisting state-of-the-art methods across diverse adaptation scenarios,achieving 60.5% mAP on clear-to-adverse weather (Foggy Cityscapes),46.9% mAP on small-to-large-scale datasets (BDD100K), and 69.9%mAP on synthetic-to-real adaptation (SIM10K). Code is available athttps://github.com/ohhotaek/DA-F2F.

Knowledge Distillation (KD) is widely used to obtain com-pact models for efficient inference in resource-constrained environments.Yet the computational overhead of the distillation process itself is oftenoverlooked, raising the question of whether a better student model canbe obtained with less data and less compute via data pruning. However,existing data pruning methods are not designed for KD: some introducesubstantial overhead (e.g., obtaining training dynamics through retrain-ing), while others rely on heuristic selection rules that fail to capture whatKD actually requires, often resulting in suboptimal subsets. To addressthese issues, we propose IF-Beta, an efficient data pruning frameworkthat combines influence function and a learnable sampling policy. Em-pirically, we first demonstrate that influence functions can serve as aneffective and efficient estimator of sample impact in KD settings, whereonly a pretrained teacher is available. Building on this, our samplingpolicy is specifically parameterized by a Beta distribution, whose highlyflexible two-parameter family allows the policy to adapt to diverse prun-ing regimes rather than being tied to fixed heuristic forms. Next, weformulate KD pruning as optimizing this policy through a bilevel objec-tive, where the inner loop operates in the teacher’s feature space with aKD-aligned objective, enabling fast proxy training, while the outer loopupdates the policy parameters to maximize the distillation performance.This design ensures IF-Beta is both computationally efficient and inher-ently aligned with the goals of KD. Extensive experiments on CIFAR-10/100 and ImageNet show that IF-Beta consistently outperforms otherbaselines across a wide range of pruning ratios. Remarkably, IF-Betadoes enable students trained on less data and less compute to surpassthe performance of students distilled on the full dataset.

Precise human mesh recovery (HMR) from multi-view im-ages remains challenging: end-to-end methods produce entangled errorshard to localize, while _x001C_tting-based methods rely on sparse keypointsthat provide limited surface constraints. We observe that the true bot-tleneck lies in the quality of intermediate representations, and that densepixel-to-surface correspondences can be e_x001B_ectively generated by repur-posing pre-trained di_x001B_usion models with rich visual priors. We proposeDi_x001B_Proxy, a Stable-Di_x001B_usion-based framework trained on large-scalesynthetic data with pixel-perfect annotations. A multi-conditional proxygenerator predicts dense correspondences from multi-view images, pro-viding uniform surface constraints that enable precise _x001C_tting. Hand re-_x001C_nement feeds enlarged hand crops alongside full-body images for _x001C_ne-grained detail, while test-time scaling exploits di_x001B_usion stochasticity toestimate per-pixel uncertainty. Trained only on synthetic data, Di_x001B_Proxyachieves state-of-the-art results on _x001C_ve diverse real-world benchmarks.Project page: https://wrk226.github.io/DiffProxy.html

Recent advances in video generation have revealed an un-expected capability: diffusion-based video models can perform reason-ing through Chain-of-Frames (CoF), suggesting that reasoning unfoldssequentially across video frames. In this work, we revisit this assump-tion and uncover a different mechanism. Through systematic analysis,we show that video reasoning primarily develops along the diffusion de-noising steps instead, where early steps exhibit multiple hypotheses thatgradually converge. We term this mechanism Chain-of-Steps (CoS),which is validated through qualitative analysis and probing tests. Ourinvestigation reveals several intriguing emergent behaviors: 1) Modelsdemonstrate a form of persistent working memory that supports tasksrequiring consistent reference, such as object permanence. 2) They canself-correct intermediate mistakes during generation instead of com-mitting to incorrect trajectories. 3) Early steps function differently fromlater steps, exemplified by a perception before action phenomenon.Moreover, analysis of Diffusion Transformer layers shows that middlelayers conduct key reasoning procedures. Motivated by these insights,we propose Training-Free Ensemble (TFE), a simple strategy thatintegrates reasoning paths by merging latents from identical models withdifferent random seeds at inference time. This approach encourages theexploration of diverse reasoning trajectories and improves reasoning per-formance. Together, our findings provide the first systematic dissectionof the mechanisms underlying video reasoning and offer practical insightsfor developing more capable video reasoning models.

Infrared–visible video fusion aims to integrate complemen-tary thermal and visual information to produce stable and informativefused videos for real-world applications. While recent advances haveachieved remarkable progress in image-level fusion, extending these meth-ods to dynamic video scenarios remains challenging. Existing video fu-sion approaches typically adopt a ‘temporal-first’ processing strategy,where temporal modeling is performed within each modality before cross-modal fusion. However, such designs may become vulnerable to modality-specific degradations, including low illumination in visible videos andnoise or flicker in infrared sequences, which often lead to unreliablemotion estimation and temporally inconsistent fusion results. To ad-dress this limitation, we advocate a ‘fusion-first’ strategy and proposeDRT-VF, a unified degradation-robust and temporally consistent in-frared–visible video fusion method built upon a one-step diffusion ar-chitecture with a two-stage progressive training strategy. In Stage I, aModality-Collaborative Attention module integrates complementary in-frared and visible cues to establish robust intra-frame fusion representa-tions. In Stage II, temporal consistency is modeled through a hierarchicaldesign consisting of a Global Temporal Consistency module and a LocalMotion Compensation module, which jointly enhance global temporalconsistency and refine motion-sensitive regions. Extensive experimentsdemonstrate that DRT-VF improves temporal stability while preservingfine structural details.

Robust out-of-distribution (OOD) detection is crucial fordeploying deep learning systems in high-stakes settings such as medi-cal imaging and autonomous driving. Post-hoc feature-based detectorsare attractive because they operate on pre-trained models, yet they of-ten treat all feature dimensions as equally informative. We show thatOOD-discriminative information in deep feature representations is of-ten concentrated in subsets of features and is frequently axis-aligned. Toexploit this structure, we introduce a distribution-aware feature selec-tion strategy that ranks feature dimensions according to the discrepancybetween in-distribution (ID) and OOD feature distributions, using theWasserstein-1 distance as a principled metric.To avoid the need for curated OOD validation data, we construct proxy-OOD data based on cross-domain mixup and evaluate adversarial per-turbations as an alternative. Our method is a lightweight, broadly ap-plicable add-on to existing feature-based OOD detectors that requiresno retraining or changes to scoring functions. Evaluations across pub-lic benchmarks spanning medical and natural image domains show con-sistent performance improvements, while the reduced feature set lowerscomputational complexity and enables substantially faster inference. Ourcode is available at https://github.com/remic-othr/mfs-ood.

Recently, 3D Gaussian Splatting (3DGS) greatly acceleratedmesh extraction from posed images due to its explicit representation andfast software rasterization. While the addition of geometric losses andother priors has improved the accuracy of extracted surfaces, mesh ex-traction remains difficult in scenes with abundant view-dependent effects.To resolve the resulting ambiguities, prior works rely on multi-view tech-niques, iterative mesh extraction, or large pre-trained models, sacrificingthe inherent efficiency of 3DGS. In this work, we present a simple and ef-ficient alternative by introducing a self-supervised confidence frameworkto 3DGS: within this framework, learnable confidence values dynam-ically balance photometric and geometric supervision. Extending ourconfidence-driven formulation, we introduce losses which penalize per-primitive color and normal variance and demonstrate their benefits tosurface extraction. Finally, we complement the above with an improvedappearance model, by decoupling the individual terms of the D-SSIMloss. Our final approach delivers state-of-the-art results for unboundedmeshes while remaining highly efficient.

Color transfer aims to align the color distribution of a sourceimage with that of a reference image while preserving structural andsemantic consistency. However, existing methods often suffer from inac-curate global mapping, semantic misalignment, and visual artifacts. Toaddress these issues, we propose ColorFM, an optimization-to-learningframework. ColorFM connects online optimization to offline inference byreformulating color transfer as the transport of pixel distributions alongvelocity fields via Flow Matching. Specifically, we introduce ColorFM-O, an instance-specific optimization scheme that fits the velocity fieldthrough hierarchical color coupling guided by semantic priors. By nu-merically integrating the induced flow trajectories, ColorFM-O producesprecise and semantically consistent color transfer results, while generat-ing high-quality paired data as pseudo-supervision. Building upon this,we design ColorFM-L, an efficient feed-forward model trained on thegenerated pairs. Through implicit state modeling, ColorFM-L extractsdeep semantic features to predict flow parameters for bidirectional lin-earized transport, ensuring accurate color transfer. Extensive experi-ments demonstrate that ColorFM-L outperforms state-of-the-art meth-ods in visual quality, structural fidelity, and semantic consistency, suc-cessfully combining the accuracy of optimization with the speed of feed-forward inference.

Spatial reasoning is a critical capability for intelligent robots,yet current vision-language models (VLMs) still fall short of human-levelperformance in video-based spatial reasoning. This gap mainly stemsfrom two challenges: a semantic-geometric misalignment that preventsconsistent 3D understanding, and the absence of persistent memory toretain 3D representation and understanding across frames. To addressthese limitations, we present VLM2 , a Vision-Language Model with per-sistent Memory for spatial reasoning with a view-consistent, 3D-awarerepresentation purely from 2D videos. Specifically, we incorporate a dual-memory module consisting of a working memory that operates as a slid-ing window to focus on immediate context, and an episodic memory thatconsolidates and stores critical information across frames. This designenables bounded and efficient spatial reasoning under a fixed computa-tional cost. Extensive experiments on multiple benchmarks show thatVLM2 achieves state-of-the-art performance among video-based models,significantly advancing the frontier of visual-spatial intelligence.

Retrieval-Augmented Text-to-Motion (RAG-T2M) modelshave demonstrated superior performance over conventional T2M ap-proaches, particularly in handling uncommon and complex textual de-scriptions by leveraging external motion knowledge. Despite these gains,existing RAG-T2M models remain limited by two closely related factors:coarse-grained text-motion retrieval that overlooks the hierarchical struc-ture of human motion, and underexplored mechanisms for effectivelyfusing retrieved information into the generative process. In this work,we present ReMoMask, a structure-aware RAG framework for text-to-motion generation that addresses these limitations. To improve retrieval,we propose Hierarchical Bidirectional Momentum (HBM) contrastivelearning, which employs dual objectives to jointly align global motionsemantics and fine-grained part-level features with text. To bridge thegap between structured retrieval and generation, we introduce TopologyStructured Masking (TSM), a training strategy that adaptively masksmotion tokens based on semantic relevance, forcing the model to learnrobust part-level grounding. Furthermore, we design Semantic Spatial-Temporal Attention (SSTA), a topology-aware fusion module that inte-grates retrieved knowledge via an asymmetric attention mechanism. Ex-tensive experiments on HumanML3D, KIT-ML, and SnapMoGen demon-strate that ReMoMask consistently outperforms prior methods on bothtext-motion retrieval and text-to-motion generation benchmarks.

Open World Classification (OWC) with Large MultimodalModels (LMMs) is an emerging and promising task that moves beyondthe closed-world setting by enabling direct answer generation in responseto open-ended queries. However, we observe that the OWC performanceof LMMs significantly degrades on fine-grained tasks. Empirically, wefind that their performance is sensitive to contextual information in thequery. To address this limitation, we propose Adaptive Memory ContextInjection (AMCI), a training-free and model-agnostic framework that dy-namically restores contextual information during OWC inference. AMCImaintains a non-parametric memory that aligns visual indices with theircorresponding attribute-augmented descriptions. By employing an adap-tive similarity threshold to preserve memory diversity and utilizing arefinement LLM to transform accumulated context into a structuredprompt, AMCI transforms the initial zero-shot description into a cali-brated fine-grained prediction without parameter updates. Furthermore,we construct a specialized low-resource fine-grained benchmark to sim-ulate extreme data scarcity and introduce a new metric that penalizesoverly generic predictions, tailored for fine-grained OWC evaluation. Ex-tensive experiments demonstrate that AMCI consistently improves per-formance by a substantial margin while remaining computationally effi-cient and compatible with various LMMs.