ECCV 2026 论文列表

Despite recent progress, direct text-driven 4D object gen-eration remains challenging yet highly desirable. In this paper, we in-troduce DiTex4D, a native text-to-4D generation framework that en-ables both text-driven 4D generation from scratch and 3D animationfrom static mesh. Built upon large-scale pre-trained 3D generation mod-els, our framework DiTex4D avoids intermediate text-to-video pipelinesand costly per-object optimization. Specifically, (i) we achieve 4D spa-tiotemporal consistency via inflating 3D attention with mixed-4D RoPEand tailored correlated noise injection strategy. (ii) To enable 3D anima-tion, we introduce a mask-based diffusion model conditioned on multi-view global context to maintain strict consistency with the initial frame.We further fine-tune the framework for 4D interpolation to synthesizehigh-frame-rate sequences with smoother motion. Extensive experimentsdemonstrate that DiTex4D can achieve higher-quality, semantically align-ed, and spatiotemporally coherent 4D object generation, surpassing mostexisting state-of-the-art text-to-4D generation methods.

Pre-trained vision-language models excel across varied per-ception tasks, but adapting them to novel downstream settings with-out sacrificing generalization remains non-trivial. Existing parameter-efficient prompt learning method often yields inconsistent representa-tions and fails to account for semantic distribution shifts. In this work,we present DiscoVL, a disentangled cross-modal representation learningframework that couples orthogonal adversarial regularization with struc-tured cross-modal alignment for vision-language models. To address theinsufficient cross-modal interaction, our DiscoVL designs a multi-branchlow-rank residual aligner that decomposes representations into subspacesand enables bidirectional cross-modal feedback between visual and tex-tual streams at each layer. Furthermore, while conventional triplet con-straints overfit features to class centroids, we design an orthogonal reg-ularization for adversarial triplet loss, which prevents centroid collapseand substantially boosts generalization. Evaluations on 15 benchmarksdemonstrate that DiscoVL delivers consistent improvements over state-of-the-art methods for base-to-novel generalization, cross-dataset evalu-ation, and few-shot learning.

Text-to-image diffusion models deliver high-quality images,yet aligning them with human preferences remains challenging. We re-visit diffusion-based Direct Preference Optimization (DPO) for thesemodels and identify a critical pathology: enlarging the preference mar-gin does not necessarily improve generation quality. In particular, thestandard Diffusion-DPO objective can increase the reconstruction er-ror of both winner and loser branches. Consequently, degradation ofthe less-preferred outputs can become sufficiently severe that the pre-ferred branch is also adversely affected even as the margin grows. Toaddress this, we introduce Diffusion-SDPO, a safeguarded update rulethat preserves the winner by adaptively scaling the loser gradient ac-cording to its alignment with the winner gradient. A first-order analysisyields a closed-form scaling coefficient that guarantees the error of thepreferred output is non-increasing at each optimization step. Our methodis simple, model-agnostic, broadly compatible with existing DPO-stylealignment frameworks and adds only marginal computational overhead.Across standard text-to-image benchmarks, including validations on thelarge-scale FLUX.1-dev model, Diffusion-SDPO delivers consistent gainsover preference-learning baselines on automated preference, aesthetic,and prompt alignment metrics. Our code is available at https://github.com/AIDC-AI/Diffusion-SDPO.

Federated remote sensing (RS) semantic segmentation en-ables privacy-preserving collaborative training of dense prediction mod-els. However, its performance is severely hindered by the structured het-erogeneity inherent in RS image: sensor diversity causes distinct spec-tral shifts, while geographic variability induces spatial shifts. Existingmethods treat client heterogeneity as an unstructured distributional mis-match, failing to accommodate these coupled, RS-specific variations.To address this, we propose FedDap, a federated prototype adapta-tion framework that disentangles domain-conditioned variations from in-variant class semantics via Domain-Adaptive Prototypes (DAPs). DAPstransform globally shared semantic anchors through two specialized mod-ulation branches: a spectral branch for radiometric alignment and a spa-tial branch for multi-scale structural adaptation. A shift-aware gatingmechanism dynamically fuses these branches, producing class-conditionalprototypes tailored to local sensing and regional contexts. By employ-ing a pixel-to-prototype contrastive objective, we align local representa-tions with both the shared semantic references and the DAPs, effectivelymitigating cross-client mismatch. Extensive experiments across diversefederated RS benchmarks demonstrate that our method consistently out-performs state-of-the-art baselines.

Remote sensing change detection is caught in a long-standingtrade-off: reliable suppression of pseudo changes requires global contex-tual reasoning, while attention-based models incur prohibitive quadraticcomplexity, hindering deployment on resource-constrained UAV and satel-lite platforms. This paper proposes ChangeGCC, a fully convolutional,linear-time framework that delivers efficient global context for change de-tection. ChangeGCC core features two key components: Globally Condi-tioned Convolution that enable convolutional global routing, and a dual-temporal fusion scheme that enhances semantic consistency across scaleswhile highlighting genuine structural changes. Extensive experiments onmultiple benchmarks show that ChangeGCC achieves state-of-the-artperformance, reaching 85.28% IoU and 92.06% F1 on the LEVIR-CD benchmark, while substantially reducing parameters, FLOPs, andinference latency compared to attention-based counterparts.

Unsupervised outlier detection that automatically identifies whether visual systems involve anomalous images is a significant research topic. However, most approaches are limited to natural images because they rely on frozen and pre-trained feature extractors. So that their performance cannot be maintained in practical scenarios, especially for industrial inspection and medical imaging. In this work, we first revisit this task and then introduce a versatile unsupervised outlier detection framework to enrich the application domains. The core idea of this framework is to improve feature discriminativess via exploiting intrinsic distribution priors. Evaluated on 3 domains and 14 benchmark datasets, our proposed solution achieves state-of-the-art performance and significantly outperforms existing methods. More importantly, we show its plug-andplay property that can be integrated into diverse visual applications to improve their robustness, such as image classification and 3D reconstruction. Our code is available at https://github.com/zhliu-uod/VUOD.

The recently established ‘Out of Sight, Not out of Mind’ (OSNOM) task for egocentric videos focuses on tracking objects that are moved by the camera wearer, online, maintaining knowledge of instance locations throughout the video even when they leave the field of view or become heavily occluded. In this paper, we propose the first learningbased solution to the OSNOM task: Whareformer, a transformer-based model with two components: an updatable memory of established tracks and a track assignment module that associates observations with existing tracks in a feed-forward manner. Whareformer jointly reasons over evolving object appearance (what) and updated 3D location (where), and employs a dedicated New Track token to reason about novel objects. Thanks to its design choices of using relative distances and evolving track representations, Whareformer is trained on a small set of 56 videos but achieves SOTA performance on 260 long test videos from three datasets: EPIC-KITCHENS-100 (unseen videos), IT3DEgo, and HD-EPIC, with significant absolute improvements over prior work.

Despite significant advances in Medical Report Generation(MRG), the reliability remains constrained by the prevalence of factualerrors. While Direct Preference Optimization (DPO) has emerged as apromising post-training paradigm to enhance the performance of Super-vised Fine-Tuned (SFT) MRG models, existing DPO-based MRG meth-ods typically adopt a naive preference construction that directly pairsmodel-generated reports with ground truth reports. This strategy inad-vertently entangles critical clinical findings with clinically irrelevant lin-guistic characteristics, and fundamentally lacks explicit vision-languagealignment. To address these challenges, we propose DPO-Clin, a novelpost-training framework that focuses preference optimization on clinicalfindings and cross-modal alignment. First, we introduce the Entity-levelClinical Diagnostic (ECD) module to perform a precise entity-level fac-tual diagnosis. ECD guides the generation of linguistically-aligned re-port preference pairs, isolating clinical discrepancies from linguistic vari-ations. Second, to achieve fine-grained cross-modal alignment, we developM2 DPO, a retrieval-augmented multi-modal DPO variant that enforcestextual preference inversion triggered by visual context switches. Third,we locate correct yet highly uncertain predicted entities and apply coun-terfactual modifications to construct targeted preference data for latentrisk mitigation, thereby further enhancing the model reliability. Exten-sive experiments on two public chest X-ray datasets (MIMIC-CXR andIU X-Ray) and an in-house endoscopy dataset demonstrate that DPO-Clin significantly improves the SFT baselines on clinical-aware metrics.Furthermore, it achieves superior performance over existing DPO-basedMRG methods, exhibiting robust generalizability across distinct baselinearchitectures and diverse medical imaging modalities.

With advancements in AI, appearance-based gaze estimationmethods have improved benchmark performance, but practical deploy-ment still lags behind commercial eye-trackers. Factors like model size,inference time, and privacy often go unaddressed. Meanwhile, webcam-based eye-tracking methods remain sensitive to head movement. To tacklethese issues, we introduce WebEyeTrack, a browser-native frameworkthat integrates a novel lightweight gaze model, metric head pose estima-tion, and on-device few-shot learning with as few as nine per-user calibra-tion samples (k ≤ 9). WebEyeTrack adapts to new users, achieving com-petitive performance with an error margin of 2.32 cm on GazeCaptureand real-time inference latency of 2.4 milliseconds on an iPhone 14. Ouropen-source code is available at github.com/RedForestAI/WebEyeTrack.

What if a video generation model could not only imaginea plausible future, but the correct one – accurately reflecting how theworld changes with each action? We answer this by presenting the Ego-centric World Model (EgoWM), a simple, architecture-agnostic methodthat transforms any pre-trained video diffusion model into an action-conditioned world model, enabling precisely controllable future predic-tion. Rather than training from scratch, we repurpose the rich world pri-ors of Internet-scale video models by injecting appropriately compressedmotor commands through lightweight conditioning layers. This allowsour model to follow actions faithfully while preserving generalizationand realism. Our approach scales naturally across embodiments and ac-tion spaces – from 3-DoF mobile robots to 25-DoF humanoids, wherepredicting egocentric joint-angle driven dynamics is substantially morechallenging. The model produces coherent rollouts for both navigationand manipulation, requiring only modest fine-tuning. To evaluate phys-ical correctness independent of appearance, we introduce the StructuralConsistency Score (SCS), which measures whether stable scene elementsevolve consistently with the provided actions. Our method improves SCSby up to 65% over the prior state of the art, Navigation World Models;applies seamlessly to three different video diffusion model architectures;and effectively utilizes Internet priors to generalize to unseen environ-ments, including navigation and manipulation inside paintings. Finally,we demonstrate the applicability of EgoWM to robotic planning.

Touch supplies the physical grounding needed to perceive intrinsic material properties, such as friction and compliance, that vision alone often cannot resolve. Recent efforts for equipping multimodal LLMs with this tactile sense, however, expose a zero-sum trade-off: the limited parameter budget of compact models forces a choice between acquiring the new sensory modality and preserving the established visionlanguage reasoning. We present Splash, a mask-isolated tactile alignment learning framework for MLLMs. Splash quantifies the significance of each pretrained parameter, and partitions the parameter space into a dormant and critical subspace. While the frozen critical subspace acts as a stable anchor to safeguard general visual knowledge, Splash updates the isolated dormant subspace to internalize tactile alignment towards LLMs. This selective, non-destructive expansion effectively prevents catastrophic forgetting and ensures non-destructive modality expansion. Extensive experiments show that Splash effectively achieves tactile reasoning without additional inference overhead in the LLM part, demonstrating state-of-the-art performance on visuo-tactile benchmarks, including SSVTP, TVL, and TacQuad, while preserving its original generalpurpose capabilities.

Pre-trained Vision Transformers (ViTs) have shown strongperformance across diverse tasks, creating increasing demand for efficientfine-tuning and compression for downstream deployment. Existing VectorQuantization (VQ) methods achieve high compression ratios, yet strug-gle to balance fine-tuning efficiency and downstream performance. In thiswork, we propose Vector Quantization Tuning (VQT), a novel frameworkfor efficient fine-tuning and compression of pre-trained ViTs. We iden-tify Coupled Vector Quantization Noise (CVQN) as a key factor limitingperformance. It arises from the interaction between Low-Rank Adapta-tion (LoRA) and the Straight-Through Estimator (STE), which degradesoptimization effectiveness under vector quantization. To address this,VQT employs a VQ-aware fine-tuning process that stochastically replacesvector-quantized weights with continuous counterparts to reduce CVQN,while progressively aligning the model toward fully vector-quantized rep-resentations to ensure consistency between fine-tuning and deployment.Furthermore, VQT enables efficient storage for multi-task deploymentby storing only lightweight task-specific parameters in the cloud, withless than 1 bit per parameter needed to reconstruct compressed models.Extensive experiments show the advantages of VQT over state-of-the-art baselines in performance and efficiency at high compression ratios.For example, at 1-bit compression, VQT surpasses the advanced baselineby 9.5% on the VTAB-1K benchmark while fine-tuning only 0.8% of theparameters. The code is available: https://github.com/leenas233/VQT.

Pretraining Vision-Language-Action (VLA) policies on internet-scale video is appealing, yet current latent-action objectives often learnthe wrong thing: they remain anchored to pixel variation rather thanaction-relevant state transitions, making them vulnerable to appearancebias, nuisance motion, and information leakage. We introduce VLA-JEPA,a JEPA-style pretraining framework that sidesteps these pitfalls by design.The key idea is leakage-free state prediction: a target encoder produceslatent representations from future frames, while the student pathwaysees only the current observation—future information is used solely assupervision targets, never as input. By predicting in latent space ratherthan pixel space, VLA-JEPA learns dynamics abstractions that are ro-bust to camera motion and irrelevant background changes. This yieldsa simple two-stage recipe—JEPA pretraining followed by action-headfine-tuning—without the multi-stage complexity of prior latent-actionpipelines. Experiments on LIBERO, LIBERO-Plus, SimplerEnv and real-world manipulation tasks show that VLA-JEPA achieves consistent gainsin generalization and robustness over existing methods. Code is availableat https://github.com/ginwind/VLA-JEPA.

Zero-shot anomaly detection (ZSAD) requires detecting andlocalizing diverse defects in unseen categories. Recent vision-languagemodel (VLM) based approaches improve ZSAD via prompt learning, yettheir prompts are often weakly grounded to transferable visual evidence,making them prone to semantic drift under domain shifts and limit-ing the reuse of normal/abnormal primitives. We propose a VariationalPrototype-conditioned prompting framework (VarProtoAD) that ex-plicitly conditions prompt learning on visual prototypes discovered in thefeature space. Specifically, multi-layer patch tokens are ℓ2 -normalized andmodeled with a variational von Mises–Fisher mixture to learn layer-wisenormal and abnormal prototype banks. These prototypes form multi-cluster semantic anchors on the hypersphere and are naturally alignedwith cosine-based VLM matching, where anomalies correspond to direc-tional deviations. Conditioned on these anchors, learnable prompt con-text tokens interact with the prototype banks via confidence-gated cross-attention, producing explicit prototype-conditioned prompts as well asimplicit deviation prompts derived from normal prototypes to bettercover unseen anomaly semantics. Finally, multi-layer image–text simi-larities are fused within a unified framework to jointly produce image-level anomaly scores and pixel-level anomaly maps. Experiments on 15industrial and medical datasets demonstrate strong cross-domain per-formance, indicating that VarProtoAD yields more stable and robustprompt learning.

Video foundation models now reach human-level accuracy onphysical-reasoning benchmarks, yet such tasks require predicting unob-served physical outcomes. Do these models perform human-like forwardsimulation, or do they exploit statistical regularities in visible scenes?Accuracy alone cannot distinguish these strategies. We introduce a distri-butional evaluation framework that treats model seeds and human ratersas populations, enabling comparison of consensus, uncertainty, and strat-egy. On the Physion benchmark, we evaluate three ViT-L architectures(V-JEPA2, VideoMAEv2, DINOv2). V-JEPA2 narrows the accuracy gapto ∼1 percentage point (73.2% vs. 74.2%), yet model–human disagree-ment reaches 26.4%, far exceeding human–human disagreement (4.8%),with substantially lower agreement (κ ≈ 0.48 vs. 0.91). The divergenceis scenario-structured: models outperform humans on geometric reason-ing (linking, +11.8 pp), while human advantages appear on causal-chainand other future-dynamics scenarios in the released split (e.g., dominoes,−10.5 pp). Strategy fingerprinting confirms all three architectures sharenon-human strategies while none aligns with humans. Attribution analy-sis suggests that unobservable outcome features, rather than visible sceneproperties, predict this divergence, consistent with models relying moreon scene-level statistical regularities than on explicit forward simulation,a systematic divergence that accuracy alone cannot reveal. Code is avail-able at https://github.com/fanhong-li/model-human-divergence.

We present URHead, a unified representation for high-fidelityand animatable head avatars that fundamentally redefines mesh-Gaussianintegration. While mesh-based methods offer precise geometric controlbut lack photorealistic detail, and Gaussian-based approaches achievephotorealism but suffer from poor structural consistency, existing hybridsolutions fail to fully leverage their complementary strengths. Our keycontribution is a UV-space unification where both representations sharea common UV parameterization. Through joint optimization with adap-tive gaussian sampling, our method automatically learns to disentangleand allocate appropriate roles to each component. URHead maintainsfull parametric controllability while preserving subject-specific details,outperforms existing state-of-the-art methods in reconstruction qualityand animation consistency.Our code is available at https://github.com/Lseonghak/URHead

Unsupervised registration of large-scale LiDAR point cloudsremains challenging due to the geometric ambiguity inherent in out-door scenes, which degrades pseudo-label quality and leads to subop-timal convergence, particularly for sparse, low-resolution scans such asthose from nuScenes. We reveal that registration models intrinsicallyencode semantic awareness that strongly correlates with registration ac-curacy, albeit without explicit semantic supervision. However, this nativeawareness is fragile: noisy supervision arising from geometric ambigu-ity in unsupervised settings rapidly erodes the learned semantic struc-ture, causing performance collapse. To this end, we propose CAESAR,a teacher-student framework guided by an off-the-shelf 3D segmenta-tion model exclusively during training. We observe that potential in-lier matches are often buried just beneath a few spurious neighbors inthe noisy feature space, motivating Dual-Cue Guided Re-Matching torecover them through reselection rather than simply rejecting. Build-ing on this, a train-only Semantic-Geometric Label Mining performsCorresponding author.lightweight, batch-specific teacher refinement and mines reliable pseudo-labels under semantic guidance. We further introduce Semantic Predic-tive Distillation to consolidate the student’s semantic awareness in thefeature space. Extensive experiments on KITTI and nuScenes demon-strate state-of-the-art performance, with pronounced gains on the chal-lenging nuScenes benchmark. Crucially, CAESAR incurs zero inferenceoverhead and requires no semantic annotations on the registra-tion data.

Scene understanding requires simultaneous prediction aboutgeometry, appearance, and semantics. However, existing task-specific an-notations are fragmented across incompatible, domain-specific datasets.Current unified systems circumvent this by restricting training to fullyco-annotated data, or by incurring the large computational cost of pseudo-labeling. To mitigate this, we introduce UniD, a unified video model thatjointly predicts eight dense scene properties—depth, surface normals, se-mantic segmentation, boundaries, human parts, albedo, shading, and ma-terials—all learned from disjoint, domain-specific datasets. We propose asimple yet effective distillation step in which per-task experts supervisea unified backbone through lightweight task projectors, eliminating theneed for annotation overlap or pseudo-labeling. Our key insight is thatthe strong visual priors of a pretrained diffusion model are sufficientto bridge the domain gaps introduced by disjoint training sources, en-abling robust generalization to scene-task combinations never seen dur-ing training. UniD achieves competitive performance against per-taskspecialists and multi-task baselines, with strong generalization to out-of-distribution scenarios and enhanced temporal and cross-task consistency.Code and video results are available at https://unid-video.github.io.

Ensuring safe content generation in text-to-image diffusionmodels remains a critical challenge. Existing safety mechanisms focuson model editing or trajectory steering, yet the design of the blacklist,the set of unsafe concepts these methods steer away from, is an equallypivotal but largely under-explored dimension. Through controlled exper-iments on blacklist-guided concept erasure, we uncover three findingsthat overturn common assumptions: longer keyword phrases do not nec-essarily improve safety, enlarging the blacklist yields diminishing returns,and mixing in irrelevant concepts degrades performance. Together theypoint to a single principle: blacklist effectiveness is governed by how wellits keywords cover unsafe regions of the embedding space, not by individ-ual keyword quality or list size. Building on this principle, we introduceMAPS, a framework for automated blacklist optimization. MAPS usesreinforcement learning to discover a prompt token that, when insertedinto an LLM prompt, elicits semantically diverse safety keywords, andthen assembles category-specific blacklists for distinct unsafe concepts.At inference, it selects the most relevant blacklist for each prompt viaembedding similarity, maximizing coverage of the threat while avoidinginterference from unrelated concepts. MAPS acts as a drop-in enhance-ment for existing trajectory-steering methods, reducing attack successrates by 25–52% (relative) across five red-teaming benchmarks on thenudity concept while preserving generation quality, all without modelediting or modifications to the steering procedure. Our code is availableat https://github.com/juyangbai/MAPS.