ECCV 2026 Papers

When videos extend from hours to days, directly process-ing them end-to-end becomes impractical for current Multi-modal LargeLanguage Models (MLLMs). This ultra-long setting necessitates a two-stage paradigm: query-agnostic memory construction followed by retrieval-based inference. Prior work invests in complex memory construction topre-model high-level relations in videos, despite not knowing the down-stream query at build time. We instead prioritize high-recall retrievabil-ity during memory building, and defer query-specific, high-level relationcomposition to inference time. To this end, we propose MERIT (Multi-key Episodic Retrieval with Inference-time Temporal expansion), a sim-ple yet effective agentic framework for ultra-long video understanding.First, we formulate an episodic multi-key representation that enablesprecise retrieval of fine-grained memories through a simple key-matchingmechanism. Second, we introduce a neighbor filtering mechanism to cap-ture broader semantic context without the massive computational over-head of global memory construction. This is achieved by expanding thetemporal scope exclusively around the retrieved segments at inferencetime. By leveraging simple key-matching with this on-demand tempo-ral expansion, MERIT achieves state-of-the-art performance across threelong-video benchmarks: EgoLifeQA, LVBench, and Video-MME(Long).

CLIP (Contrastive Language-Image Pre-training) has be-come a de facto paradigm for image-text alignment, but it struggleswith long-context descriptions (> 77 tokens) due to absolute positionalencoding and pretraining on short captions. In long contexts, sentencesare often reordered, summarized, or partially omitted. Although priorworks extend CLIP with longer positional encodings, they often sufferfrom degraded image-text alignment under such text perturbations. Weattribute this limitation to the Euclidean contrastive objective, whichenforces strict one-to-one matching and lacks explicit mechanisms formodeling hierarchical relationships between global context and its con-stituent elements. To address this issue, we propose HyFL-CLIP, a hy-perbolic fine-tuning framework that distills the well-established text-image alignment learned in Euclidean CLIP into hyperbolic space viacross-manifold similarity distillation, leveraging its geometry to capturehierarchical and entailment relations. Our method models hierarchicalsemantics by linking summarized token-wise features, long-context de-scriptions, constituent short textual components, and images, capturingpart–whole relationships via hyperbolic entailment with Einstein mid-point aggregation. Experiments on diverse benchmarks, including long-context cross-modal retrieval, cross-modal retrieval with caption pertur-bations, intra-modality retrieval, and short-text cross-modal retrieval,show that HyFL-CLIP achieves more robust long-context understand-ing. In particular, it yields up to 19.5% improvement in long-text cross-modal retrieval under textual perturbations over the best prior method.We also show HyFL-CLIP can be seamlessly integrated into other modelframeworks by applying it to Stable Diffusion XL (SDXL). The projectpage is available at https://janeyeon.github.io/hyflclip.

We present HSImul3R1 , a unified framework for simulation-ready 3D reconstruction of human-scene interactions (HSI) from casualcaptures, including sparse-view images and monocular videos. Existingmethods suffer from a perception-simulation gap: visually plausible re-constructions often violate physical constraints, leading to instability inphysics engines and failure in embodied AI applications. To bridge thisgap, we introduce a physically-grounded bi-directional optimiza-tion pipeline that treats the physics simulator as an active supervisorto jointly refine human dynamics and scene geometry. In the forwarddirection, we employ Scene-targeted Reinforcement Learning to optimizehuman motion under dual supervision of motion fidelity and contactstability. In the reverse direction, we propose Direct Simulation RewardOptimization, which leverages simulation feedback on gravitational sta-bility and interaction success to refine scene geometry. We further presentHSIBench, a new benchmark with diverse objects and interaction sce-narios. Extensive experiments demonstrate that HSImul3R produces thefirst stable, simulation-ready HSI reconstructions and can be directlydeployed to real-world humanoid robots.

Reconstructing strand-level 3D hair from a single-view image is highly challenging, especially when preserving consistent and realistic attributes in unseen regions. Existing methods rely on limited frontalview cues and small-scale/style-restricted synthetic data, often failing to produce satisfactory results in invisible regions. In this work, we propose a novel framework that leverages the strong 3D priors of video generation models to transform single-view hair reconstruction into a calibrated multi-view reconstruction task. To balance reconstruction quality and efficiency for the reformulated multi-view task, we further introduce a neural orientation extractor trained on sparse real-image annotations for better full-view orientation estimation. In addition, we design a two-stage strand-growing algorithm based on a hybrid implicit field to synthesize the 3D strand curves with fine-grained details at a relatively fast speed. Extensive experiments demonstrate that our method achieves state-ofthe-art performance on single-view 3D hair strand reconstruction on a diverse range of hair portraits in both visible and invisible regions.

When different objectives conflict with each other in multi-task learning, gradients begin to interfere and slow convergence, therebypotentially reducing the final model’s performance. To address this, weintroduce SON-GOKU, a scheduler that computes gradient interference,constructs an interference graph, and then applies greedy graph-coloringto partition tasks into groups that align well with each other. At eachtraining step, only one group (color class) of tasks are activated, and thegrouping partition is constantly recomputed as task relationships evolvethroughout training. By ensuring that each mini-batch contains onlytasks that pull the model in the same direction, our method improvesthe effectiveness of any underlying multi-task learning optimizer withoutadditional tuning. Since tasks within these groups will update in compati-ble directions, multi-task learning will improve model performance ratherthan impede it. Empirical results on six different datasets show thatthis interference-aware graph-coloring approach consistently outperformsbaselines and state-of-the-art multi-task optimizers. We provide extensivetheory showing why grouping and sequential updates improve multi-tasklearning, with guarantees on descent, convergence, and the ability toaccurately identify what tasks conflict or align.

Visual generation models can produce photorealistic videosyet violate multiview geometry, exhibiting non-rigid deformations andocclusion inconsistencies (e.g., hallucinated content in disoccluded re-gions). These failures hinder downstream applications such as video worldmodel development and 3D asset creation, and are poorly captured byexisting metrics. We introduce GeCo, a geometry-grounded metric forjointly detecting geometric deformation and occlusion-inconsistency ar-tifacts in static scenes. By fusing residual motion and depth priors, GeCoproduces interpretable, dense consistency maps that localize these arti-facts. Using GeCo, we systematically benchmark recent video generationmodels, revealing common geometric failure modes. We further showthat GeCo provides an actionable signal by applying it as a training-free guidance loss that substantially reduces geometric artifacts duringgeneration.

Multimodal medical image registration is crucial for diagno-sis and treatment planning, yet remains challenging due to significantappearance differences between modalities such as MR and CT. Themodality gap requires that the encoder in learning-based registrationmethods has cross-modal capabilities. Meanwhile, recent vision founda-tion models (VFMs) have demonstrated strong capabilities in extractingrich cross-modal representations that help bridge such modality gaps.Inspired by this, we propose a foundation-guided multimodal represen-tation alignment framework, which transfers the rich cross-modal knowl-edge from a pretrained VFM to the registration encoder. We introduce anovel Data ReAssembly strategy to transform volumetric medical imagesinto VFM input while preserving spatial information. Additionally, weimplement a Hierarchical Representation Fusion module to dynamicallyintegrate multi-level features from the VFM, yielding semantically andstructurally rich representations. Extensive experiments on two publicMR–CT datasets for brain and abdominal imaging demonstrate the ef-fectiveness of the proposed method compared with existing approaches.The code is available at https://github.com/MeggieGuo/FGRA-Reg.

Detecting out-of-distribution (OOD) data is crucial for reliable ma-chine learning deployment. Among detection strategies, post-hoc methods areparticularly attractive due to their efficiency, as they operate directly on pre-trainednetworks without requiring retraining. Within this paradigm, one promising di-rection exploits loss-landscape curvature to estimate model uncertainty; however,such methods incur substantial computational cost and rely on implicit assump-tions about how landscape flatness differs between in-distribution (ID) and OODdata. In this work, we provide the first systematic investigation of this curvaturediscrepancy and show that OOD inputs exhibit larger Hessian curvature than IDdata, with the gap widening under stronger distributional shifts. Motivated by theseobservations, we propose F OLD, a lightweight flatness-modulated OOD detectorthat leverages the feature Hessian and partial feature normalization to improveID-OOD separability while avoiding costly parameter-space curvature approxi-mations. To optimally adapt this normalization across diverse datasets, we furtherintroduce AUTO F OLD, a self-supervised tuning scheme that synthesizes pseudo-OOD samples via ID logit masking for automatic calibration without requiringexternal data. Experiments on OOD benchmarks show that F OLD outperformsprior methods, improving the average AUROC by 1.63% and reducing FPR95by 2.30%, while maintaining computational efficiency comparable to a standardforward pass. Supported by theoretical analysis and extensive ablations, F OLDprovides a principled and practical solution for robust real-world deployment.

Text-to-image diffusion models have achieved remarkable fi-delity in synthesizing images from explicit text prompts, yet exhibita critical deficiency in processing implicit prompts that require deep-level world knowledge, ranging from natural sciences to cultural com-monsense, resulting in counter-factual synthesis. This paper traces theroot of this limitation to a fundamental dislocation of the underlyingknowledge structures, manifesting as a chaotic organization of implicitprompts compared to their explicit counterparts. In this paper, we pro-pose EruDiff, which aims to refactor the knowledge within diffusionmodels. Specifically, we develop the Diffusion Knowledge DistributionMatching (DK-DM) to register the knowledge distribution of intractableimplicit prompts with that of well-defined explicit anchors. Furthermore,to rectify the inherent biases in explicit prompt rendering, we employthe Negative-Only Reinforcement Learning (NO-RL) strategy for fine-grained correction. Rigorous empirical evaluations demonstrate that ourmethod significantly enhances the performance of leading diffusion mod-els, including FLUX and Qwen-Image, across both the scientific knowl-edge benchmark (i.e., Science-T2I) and the world knowledge bench-mark (i.e., WISE), underscoring the effectiveness and generalizability.Our code is available at https://github.com/xiefan-guo/erudiff.

Recent advances in foundation models have shown greatpromise in domains such as natural language processing and computervision, and similar efforts are now emerging in the Earth Observationcommunity. These models aim to generalize across tasks with limitedsupervision, reducing the need for training separate models for eachtask. However, current strategies, which largely focus on scaling modelsize and dataset volume, require prohibitive computational and data re-sources, limiting accessibility to only a few large institutions. Moreover,this paradigm of ever-larger models stands in stark contrast with theprinciples of sustainable and environmentally responsible AI, as it leadsto immense carbon footprints and resource inefficiency. In this work,we present a novel and efficient alternative: EoS-FM1 , an Ensemble-of-Specialists framework for building Remote Sensing Foundation Models(RSFMs). Our method decomposes the pre-training process into train-ing lightweight, task-specific ConvNeXtV2 specialists that can be frozenand reused. This modular approach offers strong advantages in efficiency,interpretability, and extensibility. Moreover, it naturally supports feder-ated training, pruning, and continuous specialist integration, making itparticularly well-suited for collaborative and resource-constrained set-tings. Our framework sets a new direction for building scalable and effi-cient RSFMs.

3D Gaussian Splatting has revolutionized the field of novelview synthesis by enabling high-fidelity and real-time rendering. How-ever, its explicit representation requires millions of primitives, result-ing in massive storage overheads that hinder scalable deployment. Toaddress this challenge, we propose Explicitly Granular Gaussian Splat-ting (EGGS), a novel representation and compression framework thatachieves an ultra-compact storage footprint without compromising vi-sual quality or rendering speed. Our method is built upon two key in-novations designed to minimize attribute dimensionality. First, we intro-duce Luma-based DC Representation, which leverages the dominance ofluminance information to compress the 3-dimensional DC component ofSpherical Harmonics into a single explicit dimension. Second, we proposeVolume-Preserving Scale Factorization, which reduces the 3-dimensionalscale attribute to a single explicit dimension while ensuring training sta-bility by maintaining consistent volume. By combining these techniqueswith a lightweight neural field for implicit attribute estimation, EGGSachieves a remarkably compact representation with the smallest storagefootprint among existing methods. Extensive experiments on the Mip-NeRF360, Tanks&Temples, and Deep Blending datasets demonstratethat EGGS significantly outperforms state-of-the-art methods in rate-distortion efficiency, reducing gigabyte-scale scenes to a mere handful ofmegabytes while maintaining high-quality real-time rendering. Code isavailable here.

Recent progress in Vision–Language–Action (VLA) modelshas enabled embodied agents to interpret multimodal instructions andperform complex tasks. However, existing VLAs are mostly confined toshort-horizon, table-top manipulation, lacking the memory and reasoningcapability required for mobile manipulation, where agents must coordi-nate navigation and manipulation under changing spatial contexts. Inthis work, we present EchoVLA, a memory-aware VLA model for mobilemanipulation. EchoVLA incorporates a synergistic declarative memoryinspired by the human brain, consisting of a scene memory that main-tains a collection of spatial–semantic maps and an episodic memory thatstores task-level experiences with multimodal contextual features. Thetwo memories are individually stored, updated, and retrieved based oncurrent observations, task history, and instructions, and their retrievedrepresentations are fused via coarse- and fine-grained attention to guidebase–arm diffusion policies. To support large-scale training, we furtherintroduce MoMani, an automated benchmark that generates expert-leveltrajectories through multimodal large language model (MLLM)–guidedplanning and feedback-driven refinement, supplemented with real-robotdemonstrations. Comprehensive simulated and real-world results demon-strate that EchoVLA substantially improves overall performance, e.g., itachieves the highest success rates of 0.52 on manipulation/navigationtasks and 0.31 on mobile manipulation tasks in simulation, exceedingthe strong baseline π0.5 by +0.20 and +0.11, respectively.

While test-time adaptation (TTA) empowers vision-languagemodels to adapt without costly retraining, it remains highly vulnerable toout-of-distribution (OOD) outliers prevalent in real-world applications.This discrepancy motivates Noisy TTA (NTTA), an online task to fil-ter noisy OOD samples on the fly while maximizing in-distribution (ID)classification accuracy. Existing zero-shot NTTA approaches typicallyrely on test-time discriminative training, leading to overconfident mis-classifications and significantly degraded inference efficiency. To addressthese limitations, we propose a novel framework named Dual Distribu-tion Estimation (DDE), shifting the zero-shot NTTA paradigm frominstance-level learning to training-free Gaussian distribution modeling.DDE incorporates two novel modules: Positive Feature Distribution Es-timation (PFDE) and Negative Label Distribution Estimation (NLDE).PFDE explicitly models class-wise inclusion and exclusion Gaussian dis-tributions to formulate a calibrated contrastive score, robustly enhancingID accuracy. In parallel, NLDE improves OOD identification by explicitlymodeling the negative label distribution to mine highly discriminative la-bels, effectively mitigating spurious correlations. Extensive experimentsshow that on the large-scale ImageNet benchmark, DDE achieves an im-provement of 3.70% in harmonic mean accuracy and reduces the FPR95for OOD detection by 6.20%, while ensuring highly scalable and effi-cient online inference. Furthermore, DDE is zero-shot and training-free,demonstrating remarkable robustness in data-scarce scenarios. Codes areavailable at https://github.com/PolyU-VCLab/OpenOOD-VLM.

High-fidelity reconstruction of large-scale aerial scenes using3DGS confronts dual challenges: restricted nadir views result in facade"blind spots" and geometric collapse in textureless regions, while themassive scale of high-resolution data induces severe computational loadimbalance during parallel training. In this paper, we propose DPGS, aholistic framework that orchestrates visual foundation model priors andgenerative diffusion priors to achieve robust reconstruction. First, we in-troduce a foundation model-driven pixel-wise dense initialization to mit-igate geometric incompleteness from sparse SfM points, effectively pre-venting floating artifacts in weak-texture areas. Second, to address com-putational bottlenecks, we propose an observation density balanced par-titioning method. By formulating a graph partitioning problem weightedby feature track lengths and observation frequency, we dynamically bal-ance the rendering load across sub-blocks. Finally, we devise a diffusion-prior guided geometry completion mechanism. Leveraging a 3D-awarediffusion model with spiraling side-view sampling, we recover texturesfor unobserved facade blind spots. DPGS achieves state-of-the-art ren-dering quality, geometric completeness, and efficiency across standardbenchmarks and a self-collected dataset. Project page: https://zsddd.github.io/dpgs.

Continual Test-Time Adaptation (CTTA) aims to maintainmodel performance under evolving target domains by adapting onlinewithout labeled data. However, practical deployments often cannot re-tain the source dataset due to privacy or licensing constraints, and purelysource-free CTTA methods tend to become unstable under long-termdistribution shift, suffering from compounding self-training errors andcatastrophic forgetting. We introduce DO-ALL (Distill Once, AdaptLife-Long), a plug-and-play framework that revisits source informationin a compact and privacy-conscious form via Dataset Distillation (DD).Before deployment, DO-ALL performs DD to produce a small set of syn-thetic distilled anchors that summarize the source distribution. Duringadaptation, each target sample is matched with its most semanticallyaligned anchor, which provides a stable reference for various CTTA viasource replay, representation alignment, and manifold-smoothing regu-larization. DO-ALL can be seamlessly integrated into existing CTTA al-gorithms, consistently improving long-term robustness across CIFAR100-C, ImageNet-C, and the CCC benchmark. This demonstrates the po-tential of leveraging DD to enable stable and continuous adaptationwithout retaining raw source data. The code is available at https://github.com/blue-531/DOALL.

Recent research on Large Vision-Language Models (LVLMs)has explored visual chain-of-thought (CoT) mechanisms to enhance visual-centric reasoning, effectively mitigating the hallucination issues that plaguecurrent models. However, existing paradigms typically rely on externalvisual tools and multi-turn interactions. While recent latent visual rea-soning approaches achieve tool-free, single-turn inference by supervisingintermediate representations with auxiliary images, they incur signifi-cant annotation costs. To address this, we propose Contrastive-GuidedSelf-Supervised Latent Visual Reasoning (CoLVR), a novel framework forself-supervised training of visual latent tokens in an annotation-free man-ner. Specifically, we leverage visual contrastive principles to identify keyvisual regions and employ Reinforcement Learning (RL) to align the gen-erated latent tokens with these salient regions. Extensive experimentson hallucination benchmarks demonstrate that our method significantlyoutperforms mainstream de-hallucination approaches and even surpassesfully supervised visual reasoning baselines.

Chain-of-thought (CoT) reasoning has enabled multi-modallarge language models (MLLMs) to tackle complex visual reasoning tasksby generating explicit intermediate reasoning steps in natural language.However, this text-based reasoning paradigm is inherently slow at infer-ence time with even thousands of tokens and fundamentally constrainedby the expressiveness of natural language. In this paper, we proposeCoLT (Chain of Latent Thoughts), a novel framework that teachesmulti-modal models to reason through a chain of latent thought repre-sentations instead of verbose text tokens, which can perform thinkingwith as few as 3 steps. Naively forcing the model to think with latentstates easily produces meaningless semantics and makes training unsta-ble. To effectively regulate the latent reasoning process, we introduce alightweight external decoder that provides step-level supervision for eachlatent reasoning step in two complementary directions: a forward modethat decodes latent thoughts into the textual reasoning of the next step,and a backward mode that aligns decoder hidden states with the model’slatent thoughts given preceding textual context. We further incorporateinternal supervision that encourages coherent step-by-step latent transi-tions. The decoder and internal supervision are removed during inferenceto maintain high efficiency of latent reasoning. Extensive experiments oneight benchmarks demonstrate that CoLT not only outperforms existinglatent reasoning methods such as CODI and SIM-CoT, but also surpasseslatent visual reasoning approaches that rely on auxiliary images withcostly annotation requirements. Compared to text CoT methods, CoLTcan notably reduce the inference time by 10.1× and text decoding timeby 22.6×. Code is released at https://github.com/hulianyuyy/CoLT.

Clothes-Changing Person Re-Identification (CC-ReID) aimsto match pedestrians across non-overlapping camera views despite vari-ations in clothes. However, current research is severely constrained bytwo main shortcomings: existing datasets lack comprehensive diversityand semantic annotations, and current methods fail to globally modeland eliminate identity-irrelevant interfering factors, severely limiting CC-ReID generalization in real-world scenarios. To address these limitations,we propose a unified framework for generalizable CC-ReID. First, weintroduce CGCC, a Comprehensive dataset for Generalizable CC-ReIDfeaturing extensive diversity and fine-grained semantic annotations, com-prising 4,101 identities and 217,248 images. Additionally, we introduceText-Guided Identity Refinement (TGIR), a novel framework that uti-lizes text descriptions to construct an identity-irrelevant subspace, andemploys Singular Value Decomposition (SVD) to holistically extract andeliminate environmental and clothes interference. Together, CGCC andTGIR form a mutually reinforcing closed loop, where the diverse dataand structured annotations of the former provide the essential founda-tion for the semantic disentanglement of the latter. Extensive exper-iments demonstrate that the comprehensive diversity of CGCC effec-tively improves model generalizability, and TGIR precisely extracts ro-bust and generalizable identity representations. Dataset is available athttps://github.com/zhi-time/CGCC.

Few-shot 3D point cloud semantic segmentation aims to rec-ognize novel object categories with limited labeled examples. Existingmethods typically rely on point-based prototypes extracted from sup-port sets, but such prototypes inevitably contain background noise andlack semantic richness, leading to suboptimal segmentation quality. In-spired by chemical distillation that iteratively removes impurities, wepropose CascadeProto, which employs multi-step refinement to distillhigh-quality prototypes from noisy initializations. At its core lies theEntropy-aware Prototype Purification Module (EPPM), which lever-ages information-theoretic principles to suppress high-entropy backgroundfeatures while amplifying low-entropy foreground information. To fur-ther enrich prototype representations, we introduce Learnable ModalityAdapters (LMA) that independently align each of three CLIP modal-ities — text, audio, and image — with point cloud features throughforeground-background decoupled distribution matching, enabling flexi-ble single-modality semantic enrichment that bridges the 2D-3D domaingap. Furthermore, we propose an Attention-based Dynamic RoutingMechanism (ADRM) that adaptively aggregates predictions from mul-tiple cascade stages, allowing simple regions to benefit from early-stageoutputs while complex regions leverage deeper purification. Extensiveexperiments on S3DIS and ScanNet benchmarks demonstrate the effec-tiveness and superiority of CascadeProto across all three modalities overstate-of-the-art methods. The code is available at https://github.com/changshuowang/CascadeProto.