ECCV 2026 论文列表

In this paper, we focus on motion discrete tokenization, whichconverts raw motion into compact discrete tokens—a process proven cru-cial for efficient motion generation. In this paradigm, increasing the num-ber of tokens is a common approach to improving motion reconstructionquality, but more tokens make it more difficult for generative models tolearn. To maintain high reconstruction quality while reducing genera-tion complexity, we introduce Language-Guided Tokenization (LG-Tok)for efficient motion tokenization. LG-Tok aligns natural language withmotion at the tokenization stage, yielding compact, high-level seman-tic representations. This approach not only strengthens both tokeniza-tion and detokenization but also simplifies the learning of generativemodels. Furthermore, existing tokenizers predominantly adopt convo-lutional architectures, whose local receptive fields struggle to supportglobal language guidance. To this end, we propose a Transformer-basedTokenizer that leverages attention mechanisms to enable effective align-ment. Additionally, we design a language-drop scheme, in which lan-guage conditions are randomly removed during training. This schemeprevents shortcut learning over text and enables the detokenizer to sup-port language-free guidance. On three generation benchmarks, LG-Tokoutperforms state-of-the-art methods (e.g., achieving an FID score of0.057 vs. MARDM’s 0.114 on HumanML3D). LG-Tok-mini uses onlyhalf the tokens while maintaining competitive performance, validatingthe efficiency of our semantic representations. Code and checkpoints areavailable at https://eanson023.github.io/LG-Tok/.

3D scene graphs (3DSGs) provide a compact and structuredabstraction of 3D environments. Although advances in foundation modelshave enabled open-vocabulary 3DSG generation, existing approaches re-main object-centric and encode limited relational information—restrictingtheir applicability in real-world scenarios that require fine-grained under-standing. We propose OP3DSG, an open-vocabulary part-aware 3DSGgeneration framework that constructs unified graphs that jointly modelobjects, interactive parts, spatial relations, functional relations, and af-fordances. OP3DSG integrates object-part knowledge-guided detectionwith part-aware 3D fusion to preserve small and interaction-relevantcomponents, and employs a geometry-initialized prior graph with LLM-based refinement to reduce spurious relational predictions while enablingefficient graph construction. To systematically evaluate unified 3D scenegraph construction, we introduce UniGraph3D, a benchmark designedfor part-aware perception and multi-level relational reasoning. Experi-mental results show that OP3DSG achieves state-of-the-art performanceand demonstrates its effectiveness as a perception backbone in diversereal-world robotics tasks.

Visual markups such as highlights, underlines, and bold textare common in table-centric documents. Although multimodal large lan-guage models (MLLMs) have made substantial progress in document un-derstanding, their ability to treat such cues as explicit logical directivesremains under-explored. More importantly, existing evaluations cannotdistinguish whether a model fails to see the markup or fails to reason withit. This creates a key blind spot in assessing markup-conditioned behaviorover tables. To address this gap, we introduce HighlightBench, a diagnos-tic benchmark for markup-driven table understanding that decomposesevaluation into five task families: Markup Grounding, Constrained Re-trieval, Local Relations, Aggregation & Comparison, and Consistency &Missingness. We further provide a reference pipeline that makes interme-diate decisions explicit, enabling reproducible baselines and finer-grainedattribution of errors along the perception-to-execution chain. Experi-ments show that even strong models remain unstable when visual cuesmust be consistently aligned with symbolic reasoning under structuredoutput constraints.

Domain Generalized Semantic Segmentation (DGSS) is achallenging task, as vision models often rely on low-level appearancecues that change across domains. In contrast, structural attributes ex-hibit cross-domain stability, motivating the use of structural priors forDGSS. Existing methods use prompt learning to transfer such priorsinto DGSS models, but typically encode each class as a single holisticprompt. Moreover, these methods apply prompts uniformly to all pixels,offering no mechanism to adapt when only a subset of object regions isvisible due to viewpoint changes, occlusion, and environmental variation.We address this with Spatial Hierarchical Prompts (SHP) that en-rich each class with region-level geometric anchors capturing structuralappearance from distinct viewing angles, ensuring complementary cov-erage under arbitrary viewpoints. Additionally, we propose the Hier-archical Prompt Injector (HPI), which enables spatially adaptiveprompt injection in foundation models. HPI spatially grounds promptsby modeling their semantic relevance and spatial influence with visualfeatures. Considering the difficulty of learning spatially and semanti-cally aware prompt injection, we further introduce auxiliary supervi-sion to align hierarchical prompts with their corresponding object re-gions. We achieve 70.62% and 72.74% mIoU on synthetic-to-real andreal-to-real benchmarks, respectively. Code and checkpoints are releasedat https://github.com/MosukFate/HPI.

The continuous emergence of high-quality video deepfakesrequires detectors that continually adapt to new forgery patterns, yetexisting approaches, which are designed for deepfake images, fail to cap-ture video-specific cues. Unlike deepfake images that contain only spatialartifacts, deepfake videos leave distinct evidence along both spatial andtemporal axes, necessitating the separate preservation of each modalityduring sequential model updates. To overcome this limitation, we intro-duce a continual deepfake video detection framework, Modality-SpecificFrequency Distillation (MSFD), that explicitly decomposes video fea-tures into spatial, temporal, and spatiotemporal modalities in the fre-quency domain. This decomposition enables independent preservationof each modality, as different deepfake video types exhibit varying re-liance on spatial and temporal cues across tasks. Furthermore, MSFDadopts a cross-modality decorrelation loss that encourages spatiotempo-ral representations to remain orthogonal to single-modality cues. Exten-sive experiments show that our framework achieves stronger adaptationand preserves performance more effectively than state-of-the-art methodsacross diverse continual deepfake video scenarios.

3D Gaussian Splatting has recently revolutionised novel viewsynthesis as well as many other 3D vision methods and applications.Drawing inspiration from this representation, we now rethink texturesto overcome the main issues of UV mapping while considerably lower-ing their memory footprint. Heat Kernel Textures (HKTex) eliminateUV unwrapping as well as their persistent issues of wasted UV space,seams, distortions, vertex-duplication, and varying resolution. Groundedin discrete Riemannian geometry and intrinsically defined on any man-ifold surface discretised as a triangular mesh, HKTex uses anisotropicheat kernels as geodesic equivalents to Gaussians. Like our kernels, alsothe optimisation of their position and the adaptive densification strate-gies were redefined to operate on the surface of the object to be tex-tureised. Our novel representation is also fully integrated with a physi-cally based renderer and can be optimised either from existing texturesor multi-view images. Our project page and code are available at circle-group.github.io/research/HeatKernelTextures.

Training embodied AI agents depends critically on the visual fidelity of simulation environments and the ability to model dynamic humans. Current simulators predominantly rely on mesh-based rasterization, for which photorealistic assets are costly to author at scale, and their support for dynamic human avatars is largely constrained to mesh representations, hindering agent generalization to human-populated realworld scenarios. We present Habitat-GS, a navigation-centric embodied AI simulator extended from Habitat-Sim that integrates 3D Gaussian Splatting scene rendering and drivable gaussian avatars while maintaining full compatibility with the Habitat ecosystem. Our system implements a 3DGS renderer for real-time photorealistic rendering and supports scalable 3DGS asset import from diverse sources. For dynamic human modeling, we introduce a gaussian avatar module that enables each avatar to simultaneously serve as a photorealistic visual entity and an effective navigation obstacle, allowing agents to learn human-aware behaviors in realistic settings. Experiments on point-goal navigation demonstrate that agents trained on 3DGS scenes achieve stronger cross-domain generalization. Evaluations on avatar-aware navigation further confirm that gaussian avatars enable effective human-aware navigation, while performance benchmarks validate the system’s scalability. Code is available at https://github.com/zju3dv/habitat-gs.

Compositional Zero-Shot Learning (CZSL) requires learningto compose primitive concepts from seen data and generalize to unseenattribute–object pairs. Existing methods typically disentangle visual fea-tures into isolated primitive components and align them with their cor-responding text concepts or full compositions. However, they often suf-fer from ambiguous vision–language alignment and lack mechanisms toexplicitly capture cross-primitive semantic interactions, ultimately hin-dering generalization to unseen compositions. To overcome these chal-lenges, we introduce Prototype-Conditioned Imagination (PCI), a two-stage framework. In the first stage, Bidirectional Alignment and Disen-tanglement (BAD) couples cross-modal attention with entropic optimal-transport regularization to effectively align visual and textual represen-tations while enforcing global disentanglement of attribute and objectprimitives, thereby mitigating ambiguous vision-language matching. Inthe second stage, Prototype-driven Compositional Modulation (PCM)explicitly models how attribute semantics modulate object features tosynthesize compositional representations, making cross-primitive rela-tions explicit and substantially improving recognition of unseen compo-sitions. Evaluated under both closed-world and open-world protocols onfour popular benchmarks, our approach achieves state-of-the-art perfor-mance, demonstrating superior generalization capabilities to unseen com-positions. Our code is available at https://github.com/YFan-Z1/PCI.

Streaming 3D reconstruction aims to recover 3D information, such as camera poses and point clouds, from a video stream, which necessitates geometric accuracy, temporal consistency, and computational efficiency. Motivated by the principles of Simultaneous Localization and Mapping (SLAM), we introduce GCT, geometric context transformer, a feed-forward 3D foundation model for reconstructing scenes from streaming data. A defining aspect of GCT lies in its carefully designed attention mechanism, which integrates an anchor context, a pose-reference window, and a trajectory memory to address coordinate grounding, dense geometric cues, and long-range drift reduction, respectively. This design keeps the streaming state compact while retaining rich geometric context, enabling stable efficient inference at around 20 FPS on 518×378 resolution inputs over long sequences exceeding 10,000 frames. Extensive evaluations across a variety of benchmarks demonstrate that our approach achieves superior performance compared to both existing streaming and iterative optimization-based approaches.

Instant reconstruction of dynamic humans and 3D human-object interactions from uncalibrated sparse-view videos is critical fornumerous downstream applications. Existing methods, however, are ei-ther limited by the slow reconstruction speeds or incapable of generat-ing novel-time representations. To address these challenges, we proposeForge4D, a feed-forward 4D human-object reconstruction and interpo-lation model that efficiently reconstructs temporally aligned represen-tations from uncalibrated sparse-view videos, enabling both novel-viewand novel-time synthesis. Our model simplifies the 4D reconstructionand interpolation problem as a joint task of streaming 3D Gaussian re-construction and dense motion prediction. For the task of streaming 3DGaussian reconstruction, we introduce learnable state tokens to enforcetemporal consistency in a memory-friendly manner. For novel-time syn-thesis, we design a novel motion prediction module to predict dense mo-tions for each 3D Gaussian between two adjacent frames. To overcomethe lack of the ground truth for dense motion supervision, we formulatedense motion prediction as a dense point matching task and introducea self-supervised retargeting loss to optimize this module. An additionalocclusion-aware optical flow loss is introduced to ensure motion consis-tency with plausible human and object movement, providing strongerregularization. Extensive experiments demonstrate the effectiveness ofour model on both in-domain and out-of-domain datasets. Code andmodels will be made publicly available.

Training Latent Diffusion Models (LDMs) within FederatedLearning (FL) has attracted increasing attention due to its ability tocombine the powerful generative capacity of LDMs with the privacy-preserving properties of FL. However, FL requires sharing the globalmodel with multiple participants, which risks unauthorized model dis-tribution or resale by malicious clients. While an intuitive approach isto adopt existing VAE-based watermarking techniques for LDMs in FL,this strategy falls short in addressing such threats due to two fundamen-tal challenges: (1) Existing methods support ownership verification butlack the ability to trace model leakage to a specific malicious client; (2)VAE-based watermarks are vulnerable, as they can be removed simplyby replacing the decoder with a clean counterpart. In this paper, we pro-pose FedOT, the first framework for ownership verification and leakagetracing in federated LDMs. Specifically, to address the first challenge, wedesign a chunked watermark, where the first part is for ownership verifi-cation, and the second part is used for client identification. Furthermore,to overcome the second challenge and secure the model against VAEreplacement attack, we introduce Latent Vector Transformation (LVT),which strengthens the connection between the VAE and U-Net latentspaces by modifying the original latent distribution of the VAE. Conse-quently, any attempt to replace the VAE for watermark removal leads tosignificant image quality degradation, making the LDM model unusable.Extensive experiments demonstrate that FedOT achieves superior per-formance in both ownership verification and traceability. Project page:https://spyzixuan.github.io/FedOT/.

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.