ECCV 2026 Papers

Understanding fine-grained eating behaviors in unconstraineddaily-life videos is essential for dietary monitoring and behavioral healthassessment, yet existing video-language benchmarks lack the temporalgranularity and domain-specific grounding required to rigorously evalu-ate models in this health-critical domain. We introduce EatVid-Bench,a large-scale, multi-dimensional benchmark designed to advance fine-grained eating behavior understanding. Our dataset comprises nearly700 real-world eating sessions (around 3,000 minutes) annotated througha novel three-tier automated pipeline that integrates 12 complementarysignals, including body pose, food detection, bite events, and facial ex-pressions, with each annotation traceable to its source signal for fullinterpretability and verifiability. Building on this foundation, we con-struct a question-answering benchmark spanning seven capability di-mensions, three difficulty levels, and five question types, with rigorousprovenance tracking throughout. Comprehensive evaluation of state-of-the-art Video-LLMs exposes a substantial perception–reasoning gap, par-ticularly in sub-second temporal grounding and cross-frame counting. Tofurther validate the effectiveness of EatVid-Bench, we propose a domain-adapted fine-tuning strategy to provide a strong open-source baselineby leveraging structured training annotations as explicit reasoning su-pervision. Code and benchmark are publicly available at GitHub andHuggingFace.

The slow iterative sampling nature remains a major bottle-neck for the practical deployment of diffusion and flow-based genera-tive models. While consistency models (CMs) represent a state-of-the-art distillation-based approach for efficient generation, their large-scaleapplication is still limited by two key issues: training instability andinflexible sampling. Existing methods seek to mitigate these problemsthrough architectural adjustments or regularized objectives, yet over-look the critical reliance on trajectory selection. In this work, we firstconduct an analysis on these two limitations: training instability orig-inates from loss divergence induced by unstable self-supervised term,whereas sampling inflexibility arises from error accumulation. Based onthese insights and analysis, we propose the Dual-End ConsistencyModel (DE-CM) that selects vital sub-trajectory clusters to achievestable and effective training. DE-CM decomposes the PF-ODE trajec-tory and selects three critical sub-trajectories as optimization targets.Specifically, our approach leverages continuous-time CMs objectives toachieve few-step distillation and utilizes flow matching as a boundaryregularizer to stabilize the training process. Furthermore, we propose anovel noise-to-noisy (N2N) mapping that can map noise to any point,thereby alleviating the error accumulation in the first step. Extensiveexperimental results show the effectiveness of our method: it achieves astate-of-the-art FID score of 1.70 in one-step generation on the ImageNet256×256 dataset, outperforming existing CM-based one-step approaches.

Safe L2/L3 driving automation requires anticipating human-in-the-loop reactions during shared-control transitions. While most driv-ing world models forecast the external environment, in-cabin intelligenceremains strictly recognition-oriented and lacks multi-step rollout capa-bilities for driver dynamics. We introduce Driver-WM, a driver-centriclatent world model that rolls out in-cabin dynamics causally conditionedon out-cabin traffic context. This formulation unifies physical kinemat-ics forecasting with auxiliary behavioral and emotional semantic recog-nition. Operating in a compact latent space constructed from frozenvision-language features, Driver-WM adopts a dual-stream architectureto separately encode external traffic and internal driver states. Thesestreams are directionally coupled via a gated causal injection mechanism,which uses a learned vector gate to modulate external contextual per-turbations while strictly enforcing temporal causality. Experiments onAIDE show robust long-horizon forecasting on reactive high-motion clips,improved driver/traffic semantic alignment, and controlled interventionsthat expose the external-to-internal mechanism.

End-to-end autonomous driving increasingly relies on gener-ative planners, broadly classified into continuous diffusion policies anddiscrete token-based vision-language-action models (VLAs). While diffu-sion policies exploit parallel iterative noise prediction to efficiently gener-ate accurate trajectories, they suffer from cross-modal misalignment, lim-ited training efficiency, and poor generalization. Conversely, token-basedVLAs generate discrete trajectory tokens autoregressively, offering bettergeneralization but being constrained by inefficient causal reasoning, cu-mulative error, and irreversible decoding. To overcome these limitations,we propose DriveFine, a vision-language-action model that exploresmasked diffusion LLMs for trajectory planning, which significantly im-proves the efficiency, flexibility, and adaptability of trajectory generation.To address the inherent irreversibility of token decoding and improve thetrajectory quality, we introduce a plug-and-play block Mixture-of-Expert(block MoE) module, which seamlessly injects the refinement capa-bility into the dLLM at minimal cost. Through explicit expert routingand gradient isolation, DriveFine decouples generation and refinement,preventing cross-task interference. We further devise a tailored hybridreinforcement learning strategy to facilitate the effective exploration ofthe refinement expert and further explore the performance ceiling. Ex-tensive experiments on the NAVSIM v1, v2, and Navhard benchmarksdemonstrate that DriveFine exhibits strong performance, robustness andefficiency. The code will be released.

Despite the success of recent novel view synthesis methods,they tend to struggle in sparse-view settings. This poor generalizationto unseen viewpoints is an inherent challenge when training with lim-ited data. To address this, we investigate the relationship between losssharpness and generalization in novel view synthesis—an underexploreddirection. Interestingly, while pursuing flatter minima is widely knownto improve generalization in deep learning, reducing loss sharpness is notalways beneficial in novel view synthesis. We demonstrate that this dif-ference arises because high-detail regions inherently require a sharp losslandscape for accurate reconstruction, whereas low-detail regions benefitfrom a flat loss landscape for improving generalization. Based on this in-sight, we introduce structure-aware sharpness, defined within structure-adaptive neighborhoods, and propose to adaptively adjust the sharpnessregularization weight according to the local image structure. This strat-egy encourages flatter minima for generalization while preserving the losssharpness necessary to reconstruct fine details. Across various datasetsand configurations, our strategy consistently improves a wide range ofbaselines. Code is available at https://bbangsik13.github.io/FASR.

We present DASAM3D, a framework that fuses the complementary strengths of SAM3D and Depth Anything 3 (DA3) for objectaware 3D scene reconstruction. A fundamental appearance–geometry trade-off exists: SAM3D yields geometrically coherent per-object Gaussian primitives but with unrealistic appearance, while DA3 delivers photorealistic 3DGS reconstructions yet lacks object-level geometric reasoning. Our key insight is to use SAM3D exclusively as a 3D geometric layout provider, guiding DA3 geometry refinement while preserving its photorealistic textures. DASAM3D proceeds through three differentiable stages—(1) Global Alignment Optimization, (2) Per-Object Alignment Refinement, and (3) Layout-Based Geometry Optimization—and inherits DA3’s flexibility for posed or pose-free inputs at any scale. Experiments on DL3DV-10K and MipNeRF-360 show consistent gains in reconstruction accuracy and novel view synthesis; a user-preference study confirms DASAM3D’s object renderings are preferred over SAM3D’s in visual fidelity.

Recent advances in world models have greatly enhanced in-teractive environment simulation. Existing methods mainly fall into two⋆ Equal contribution. † Corresponding author.categories: (1) static world generation models, which construct 3D en-vironments without active agents, and (2) controllable-entity models,which allow a single entity to perform limited actions in an otherwise un-controllable environment. In this work, we introduce CustomX, leverag-ing the realism and structural grounding of static world generation whileextending controllable-entity models to support user-specified characterscapable of performing open-ended actions. Users can provide a 3DGSscene and a character, then use natural language to direct the characterto perform diverse behaviors, ranging from basic locomotion to object-centric interactions, while freely exploring the environment. CustomXsynthesizes temporally coherent video clips that preserve visual fidelitywith the provided scene and character, formulated as a conditional au-toregressive video generation problem. Built upon a pre-trained videogenerator, our training strategy significantly enhances motion dynamicswhile maintaining generalization across actions and characters. Our eval-uation covers a broad range of aspects, including visual quality, characterconsistency, action controllability, and long-horizon coherence.

Consistency distillation has significantly accelerated diffusion-model inference, but its sampling dynamics remain underexplored. Wereveal an asymmetry: although Logit-Normal sampling priors work well forstandard iterative generation, consistency distillation exhibits a differentdifficulty profile (e.g., U-shaped), with optimization bottlenecks concen-trated at the boundary stages rather than intermediate steps. To addressthe limitations of static sampling under evolving learning demands, wepropose Curvature-Adaptive Consistency Flow Matching (CACFM). Byformulating distillation as a dynamic decision process, CACFM usesa lightweight reinforcement learning agent to probe Probability FlowODE trajectories and construct an efficiency-oriented curriculum thatprioritizes critical regions without manual scheduling. Combined withFlow-adapted DMD and adversarial consistency objectives, our RL-basedscheduler achieves state-of-the-art results on large-scale models suchas FLUX and SDXL, mitigating structural deformities and preservinghigh-frequency details in extreme few-step regimes.

Training data for class-conditional image synthesis often ex-hibit a long-tailed distribution with limited amount of images for tailclasses. Such an imbalance causes mode collapse and reduces the diver-sity of synthesized images for tail classes. For class-conditional di!usionmodels trained with imbalanced data, we aim to improve the diversityand fidelity of tail class images without compromising the quality of headclass images. We propose contrastive conditional-unconditional align-ment (CCUA), which comprises two synergistic loss functions. Our firstloss is an Alignment Loss (AL) that aligns class-conditional generationwith unconditional generation at large timesteps. Alignment loss makesthe denoising process insensitive to class conditions for the initial steps,which enriches tail classes through knowledge sharing from head classes.Secondly, we diversify unconditional generation via an UnsupervisedContrastive Loss (UCL) to increase the distance/dissimilarity amongsynthetic images. We combine the two losses to implicitly diversify condi-tional generation. Our framework is easy to implement as demonstratedon both U-Net based architecture and Di!usion Transformer. Our methodoutperforms vanilla denoising di!usion probabilistic models, score-baseddi!usion model, and alternative contrastive methods for class-imbalancedimage generation across various datasets, in particular ImageNet-LT with256→256 resolution.

Human demonstrations provide strong priors for robot ma-nipulation, yet it is non-trivial to transfer them to execute on real robotsdue to the kinematic gap. In dexterous manipulation, it remains chal-lenging to track long-horizon, contact-rich sequences even in simulators:a reference-tracking policy must keep objects on their target trajectorieswhile preserving demonstrated joint motion and contact timing. Exist-ing approaches often rely on hand-crafted reward tuning that requireper-sequence tuning and break under limited interaction budgets. We in-troduce ConTrack, a reinforcement learning (RL) framework that scaleswith tracking data. ConTrack treats object tracking as a constraint andallocates remaining control authority to motion fidelity, which allows itto adapt task–style trade-offs online using a dual-variable update. In ad-dition, ConTrack also stabilizes long-horizon learning with an adaptivemid-trajectory reset library that reuses policy-reachable simulator states.Our qualitative and quantitative results in simulation tracking and realrobot demonstrate that ConTrack improves success and object pose ac-curacy significantly over prior arts while preserving joint and contactfidelity. Website: https://www.lyt0112.com/projects/ConTrack.

Few-shot anomaly detection remains challenging due to lim-ited training data. Multi-modal anomaly detection (MAD) offers a viablesolution, leveraging 3D geometric cues to enrich 2D RGB representationsand compensate for this scarcity. However, existing MAD methods applyspatially uniform feature processing, conflating stable macroscopic struc-tures with high-frequency localized defect signals, exacerbating cross-modal misalignment and inflating false-positive rates. To overcome this,we present CMDS-AD, a Cross-Modal Dual-Stream Anomaly Detectionframework. A LoRA-guided diffusion model generates diverse RGB sam-ples to mitigate extreme data scarcity. For 3D normal augmentation, weemploy a pre-trained diffusion model as a normal estimator. Crucially,this estimator inherently acts as a non-linear low-pass filter, directly ex-tracting low-frequency normal representations from RGB inputs. Thisestablishes an auxiliary estimated stream of purely low-frequency infor-mation, anchoring robust structural templates and assisting the uncom-pressed real stream, containing coupled high- and low-frequency compo-nents, to precisely isolate micro-defects. A Coordinate-Aware Hierarchi-cal Feature Mapper adaptively aligns cross-modal semantics, while a mul-tiplicative scoring mechanism filters modality-specific noise. Under theextreme 1-shot setting, CMDS-AD achieves absolute performance gainsof 5.7% (I-AUROC) and 2.0% (AUPRO) on MVTec 3D-AD, along-side 7.7% and 5.6% improvements on EyeCandies, establishing a newstate-of-the-art. Code is available at Junhaocai27/CMDS-AD.

Mainstream 3D asset retrieval relies predominantly on category-level matching, overlooking intra-class variations in geometry, style, andtexture critical for real-world applications. To address this, we redefineretrieval from binary categorical matching to intra-class graded relevanceestimation, treating relevance as a continuous variable to capture fine-grained semantic nuances. We propose a tri-modal encoder incorporatinga Semantic Conditioned Interaction module for cross-modal fusion and aQuery-Guided Mixture-of-Experts module for dynamic modality weight-ing. This design explicitly decouples intra-class variations and preventssemantic collapse via a re-parameterization mechanism. Furthermore,we introduce an automated, scalable benchmark generation frameworkdriven by a Multimodal Large Language Model to synthesize high-qualitysoft relevance labels, proposing a new benchmark: INGRE. Extensive ex-periments demonstrate our method achieves state-of-the-art results onINGRE while maintaining superior performance and robustness on stan-dard classification tasks.

Stereotype biases in Large Multimodal Models (LMMs) per-petuate harmful societal prejudices, undermining the fairness and equityof AI applications. As LMMs grow increasingly influential, addressingand mitigating inherent biases related to stereotypes, harmful generations,and ambiguous assumptions in real-world scenarios has become essential.However, existing datasets evaluating stereotype biases in LMMs oftenlack diversity, rely on synthetic images, and often have single-actor im-ages, leaving a gap in bias evaluation for real-world visual contexts. Toaddress the gap in bias evaluation using real images, we introduce theBBQ-Vision (BBQ-V), the most comprehensive framework for assessingstereotype biases across nine diverse categories and 50 sub-categories withreal and multi-actor images. BBQ-V benchmark contains 14,144 image-question pairs and rigorously evaluates LMMs through carefully curated,visually grounded scenarios, challenging them to reason accurately aboutvisual stereotypes. It offers a robust evaluation framework featuring real-world visual samples, image variations, and open-ended question formats.BBQ-V enables a precise and nuanced assessment of a model’s reasoningcapabilities across varying levels of difficulty. Through rigorous testingof 19 state-of-the-art open-source (general-purpose and reasoning) andclosed-source LMMs, we highlight that these top-performing models areoften biased on several social stereotypes, and demonstrate that the think-ing models induce more bias in the reasoning chains. This benchmarkrepresents a significant step toward fostering fairness in AI systems andreducing harmful biases, laying the groundwork for equitable and sociallyresponsible LMMs. Dataset and evaluation code are available here.

Neural map matchers estimate an image’s 3-DoF pose rela-tive to a 2D map. These models are trained on large-scale datasets ofgeo-referenced images, whose position and heading labels often containnoise that affects the trained models. To address this, we present Auto-Compass, a supervision approach for training neural map matchers frominaccurate absolute pose labels. First, we show that heading labels areunnecessary: trained from raw GPS labels, models learn to predict accu-rate headings, automatically. Second, defining a tolerance region aroundraw GPS improves positional accuracy. Third, if available, our supervi-sion uses relative poses between training images, obtained via SLAM orSfM, which provide a more accurate training signal. Across driving andegocentric benchmarks, AutoCompass consistently outperforms counter-parts trained with the usual strong reliance on absolute pose labels.

Diffusion Transformers (DiTs) are a dominant backbone forhigh-fidelity text-to-image generation due to strong scalability and align-ment at high resolutions. However, quadratic self-attention over densespatial tokens leads to high inference latency and limits deployment. Weobserve that denoising is spatially non-uniform with respect to aestheticdescriptors in the prompt. Regions associated with aesthetic tokens re-ceive concentrated cross-attention and show larger temporal variation,while low-affinity regions evolve smoothly with redundant computation.Based on this insight, we propose AccelAes, a training-free frameworkthat accelerates DiTs through aesthetics-aware spatio-temporal reduc-tion while improving perceptual aesthetics. AccelAes builds AesMask, aone-shot aesthetic focus mask derived from prompt semantics and cross-attention signals. When localized computation is feasible, SkipSparsereallocates computation and guidance to masked regions. We further re-duce temporal redundancy using a lightweight step-level prediction cachethat periodically replaces full Transformer evaluations. Experiments onrepresentative DiT families show consistent acceleration and improvedaesthetics-oriented quality. On Lumina-Next, AccelAes achieves a 2.11×speedup and improves ImageReward by +11.9% over the dense baseline.Code is available at https://github.com/xuanhuayin/AccelAes.

3D Foundation Models (3DFMs) such as VGGT have re-cently pushed the boundaries of 3D vision by predicting rich unifiedrepresentations with feed-foward transformers. The scene representationslearned by these models enable strong performance on multiple 3D visiontasks. In this paper, we investigate using their internal representationsto infer 3D in the scene from new views. Our hypothesis is that in or-der to solve the task of 3D reconstruction, these models need to learn arepresentation that includes a large amount of general knowledge about3D scenes. After showing that it is possible to decode hidden surfacesfrom internal 3DFM representations, we propose a method, Z3D, thatestimates pointmaps in unseen views by doing latent diffusion on 3DFMrepresentation. We show that Z3D can predict realistic depthmaps fornew views across multiple datasets.

While current 6D pose estimation benchmarks have reached near-saturation on household objects, they often fail to capture the stochastic and optical complexities of industrial environments. We introduce XYZ-IBD, a high-precision benchmark for object detection and 6D pose estimation specifically designed for industrial bin-picking. XYZ-IBD addresses the domain gap by providing 75 multi-view real-world scenes containing approximately 273k annotated instances of metallic, symmetrical, and specular objects. Unlike existing datasets, our benchmark features high-density stochastic stacking and multi-instance ambiguity, reflecting authentic robotic manipulation challenges. We employ a rigorous multistage and semi-automatic annotation pipeline, ensuring sub-millimeter annotation accuracy. The annotations are validated through our designed error quantification scheme, securing the reliability of the annotation quality. In addition to real-world evaluation data, we provide a large-scale complementary synthetic training set that is rendered under realistic bin-picking simulation. Benchmarking state-of-the-art (SOTA) methods for 2D detection and 6D pose estimation reveals a significant performance degradation compared to standard household benchmarks, highlighting the unsolved challenges of industrial vision. XYZ-IBD establishes a new frontier for robust pose estimation in complex, high-occlusion, and reflective scenarios. The dataset and benchmark are publicly available at https://xyz-ibd.github.io.

Large-scale language models for autonomous driving enableenhanced global understanding and long-horizon planning. However, whendeployed in isolated vehicles, limited sensing range and occlusions re-strict reliable decision-making, and the substantial computational andlatency overhead makes on-board deployment impractical. Cooperativedriving provides a potential solution by leveraging external agents forinformation exchange, but existing methods remain limited in semanticreasoning capability under practical constraints. To address these chal-lenges, we propose DH-VLM, a dual-horizon cooperative latent reasoningframework that enables asymmetric semantic cooperation between theinfrastructure and ego vehicle. The infrastructure aggregates multi-layerhidden states to form a global-reasoning horizon latent guidance, whichis integrated into the ego model through an Infrastructure-Driven LatentEvolution mechanism for conditional latent refinement. This enables theego vehicle to leverage long-range contextual understanding while pre-serving autonomous decision-making within its local planning horizon.Furthermore, we construct a cooperation-oriented question–answer (QA)dataset covering fundamental scene understanding and ego-personalizedcomprehension to support counterfactual and safety-aware reasoning.Extensive experiments demonstrate that DH-VLM achieves state-of-the-art planning performance, outperforming the previous state of the art by14.6% in L2 error and 26.9% in collision rate. Compared with query-basedend-to-end cooperative driving methods, our approach reduces the com-munication cost by 57.3% and GPU memory usage by 25.5%, while main-taining strong robustness against infrastructure guidance errors, provid-ing a practical and robust paradigm for cooperative autonomous driving.

Diffusion Transformers (DiTs) power high-fidelity video worldmodels but remain computationally expensive due to sequential denois-ing and costly spatio-temporal attention. Training-free feature cachingaccelerates inference by reusing intermediate activations across denois-ing steps; however, existing methods largely rely on a Zero-Order Holdassumption i.e., reusing cached features as static snapshots when globaldrift is small. This often leads to ghosting artifacts, blur, and motion in-consistencies in dynamic scenes. We propose WorldCache, a Perception-Constrained Dynamical Caching framework that improves both whenand how to reuse features. WorldCache introduces motion-adaptive thresh-olds, saliency-weighted drift estimation, optimal approximation via blend-ing and warping, and phase-aware threshold scheduling across diffusionsteps. Our cohesive approach enables adaptive, motion-consistent featurereuse without retraining. On Cosmos-Predict2.5-2B evaluated on PAI-Bench, WorldCache achieves 2.3× inference speedup while preserving99.4% of baseline quality, substantially outperforming prior training-free caching approaches.