ECCV 2026 Papers

Generating realistic 3D city environments from remote sens-ing data is important for simulation, urban planning, and mixed reality,yet existing point cloud generation methods are limited to single ob-jects or bounded indoor scenes and cannot handle the scale, seamlesstiling, and partial observability challenges of city-scale generation. Wepresent GridFlow, a multi-stage framework that generates dense, col-ored point clouds (105 points per 150m×150m tile) at city scale, con-ditioned on satellite imagery, semantic segmentation maps, and digitalsurface models (DSM). A Grid-Aligned VAE encodes each tile into atopology-preserving latent grid where tokens correspond to fixed spatialregions, enabling spatially coherent multi-modal conditioning and com-pact latent-space edge consistency that implicitly aligns thousands ofboundary points for seamless cross-tile generation. A conditional rectifiedflow model synthesizes geometry latents from the fused multi-modal con-ditions, and an orientation-aware diffusion colorizer separately handlessatellite-visible horizontal surfaces and occluded vertical façades. To sup-port standardized evaluation, we build on public 3D data sources to in-troduce City3D-MultiGen, a benchmark of 163K densely annotated tilesfrom Melbourne and London with aligned point clouds, satellite images,semantic maps, and elevation data. Experiments show that GridFlowoutperforms adapted point cloud generation baselines across all geometrymetrics and produces visually coherent colored point clouds with seam-less boundaries over arbitrarily large urban extents. Our benchmark de-tails are available at https://huggingface.co/datasets/e32/City3D-MultiGen.

Generalized 3D hand-object pose estimation from a singleRGB image remains challenging due to the large variations in objectappearances and interaction patterns, especially under heavy occlusion.We propose GenHOI, a framework for generalized hand-object pose esti-mation with occlusion awareness. GenHOI integrates hierarchical seman-tic knowledge with hand priors to enhance model generalization underchallenging occlusion conditions. Specifically, we introduce a hierarchicalsemantic prompt that encodes object states, hand configurations, andinteraction patterns via textual descriptions. This enables the model tolearn abstract high-level representations of hand-object interactions forgeneralization to unseen objects and novel interactions while compen-sating for missing or ambiguous visual cues. To enable robust occlu-sion reasoning, we adopt a multi-modal masked modeling strategy overRGB images, predicted point clouds, and textual descriptions. Moreover,we leverage hand priors as stable spatial references to extract implicitinteraction constraints. This allows reliable pose inference even undersignificant variations in object shapes and interaction patterns. Exten-sive experiments on the challenging DexYCB and HO3Dv2 benchmarksdemonstrate that our method achieves state-of-the-art performance inhand-object pose estimation.

Clinical decision-making in oncology requires forecasting howdisease evolves under treatment, yet most AI systems remain static pre-dictors that cannot model longitudinal, treatment-conditioned progres-sion. Although generative and world models have demonstrated strongcapabilities in general domains, their adaptation to medicine remainslimited and insufficient for capturing complex, treatment-induced phys-iological dynamics across temporal scales. To address these gaps, weintroduce CLARITY, a medical world model that enables counterfac-tual simulation of treatment-conditioned disease trajectories for clini-cal decision-making. By jointly encoding imaging-derived latent states,temporal intervals that capture irregular follow-ups, and patient-specificclinical context, CLARITY learns smooth and interpretable representa-tions of disease progression, allowing the model to simulate how alter-native treatments reshape future disease dynamics. Because treatmentoptimization is inherently sequential and uncertain, requiring evaluationof long-term outcomes across multiple possible interventions, we furtherpropose an entropy-regularized, computationally efficient long-horizonprediction-to-decision framework that plans treatment strategies overimagined disease trajectories and iteratively refines therapy proposalsthrough survival-aware feedback, forming a closed-loop simulation-to-decision framework for treatment planning. CLARITY achieves state-of-the-art performance in treatment planning and survival predictionacross three cancer datasets, including two brain tumor cohorts (MU-Glioma-Post and zero-shot on UCSF-ALPTDG) and one breast can-cer dataset (ISPY-2), demonstrating strong generalization across can-cer types while consistently outperforming prior generative methods andmedical-domain large language model baselines.

Closed-loop simulation for autonomous driving requires synthesizing free-viewpoint driving scenes. However, existing datasets and generative pipelines rarely provide consistent off-trajectory observations, limiting large-scale evaluation and training. While recent generative models demonstrate strong visual realism, they struggle to simultaneously achieve interpolation consistency and extrapolation realism without perscene optimization. To address this, we propose FreeGen, a feed-forward reconstruction-generation co-training framework for free-viewpoint driving scene synthesis. The reconstruction model provides stable geometric representations to ensure interpolation consistency, while the generation model performs geometry-aware enhancement to improve realism at unseen viewpoints. Through co-training, generative priors are distilled into the reconstruction model to improve off-trajectory rendering, and the refined geometry in turn offers stronger structural guidance for generation. Experiments demonstrate that FreeGen achieves state-of-the-art performance for free-viewpoint driving scene synthesis.

Semi-supervised crowd analysis is a prominent area of re-search, as unlabeled data are typically abundant and inexpensive toobtain. However, traditional point-based annotations constrain perfor-mance because individual regions are inherently ambiguous, and con-sequently, learning fine-grained structural semantics from sparse anno-tations remains an unresolved challenge. In this paper, we first pro-pose an Exclusion-Constrained Dual-Prompt SAM (EDP-SAM), basedon our Nearest Neighbor Exclusion Circle (NNEC) constraint, to gener-ate mask supervision for current datasets. With the aim of segmentingindividuals in dense scenes, we then propose Exclusivity-Guided MaskLearning (XMask), which enforces spatial separation through a discrim-inative mask objective. Gaussian smoothing and a differentiable centersampling strategy are utilized to improve feature continuity and train-ing stability. Building on XMask, we present a semi-supervised crowdcounting framework that uses instance mask priors as pseudo-labels,which contain richer shape information than traditional point cues. Ex-tensive experiments on the ShanghaiTech A, UCF-QNRF, and JHU++datasets (using 5%, 10%, and 40% labeled data) verify that our end-to-end model achieves state-of-the-art semi-supervised segmentation andcounting performance, effectively bridging the gap between counting andinstance segmentation within a unified framework. Code can be found athttps://github.com/JoyceeH0127/ECCV2026XMask.

Continual fine-tuning of large vision–language models (VLMs)such as CLIP is crucial for adapting foundation models to evolving real-world distributions. However, such adaptation often disrupts pretrainedrepresentations and severely harms zero-shot generalization. Despite re-cent progress, existing methods overlook the underlying optimization ge-ometry, failing to anchor convergence within regions robust to distribu-tion shift. While recent studies suggest that flatter loss minima improvegeneralization, we empirically demonstrate that seeking local flatnesson a single task is insufficient to bridge the distributional gap betweenpre-trained and task-specific representations. Motivated by this obser-vation, we propose Dual-Generalization-aware Minimization (DGM), aplug-and-play optimization framework for continual adaptation of VLMs.DGM extends flatness-aware training by redefining the perturbation ob-jective to jointly promote pre-trained knowledge preservation and task-specific robustness. By incorporating these objectives into the inner max-imization step, DGM steers optimization toward solutions that remainstable under both zero-shot evaluation and task adaptation. Extensiveexperiments demonstrate that DGM effectively maintains zero-shot gen-eralization while achieving strong performance on sequential tasks.

Low-light image enhancement (LLIE) aims to restore visi-bility and structural details from severely degraded images under insuf-ficient illumination. However, real-world low-light images often exhibitspatially heterogeneous degradations, where mildly under-exposed areasare easy to recover while severely dark, noisy, or color-unstable regionsremain challenging. Existing LLIE methods typically apply a globallyuniform enhancement strategy, which often leads to suboptimal trade-offs such as over-smoothing textures in hard regions or over-enhancementin easier areas. To address this, we propose a Difficulty-ConditionedAttribute-Specific Restoration (DCASR) framework that explicitly mod-els restoration difficulty and allocates enhancement capacity accordingly.DCASR incorporates a difficulty-aware enhancement network with twoprior-driven modulators: a Global Intensity Modulator (GIM) for co-herent illumination correction and a Structural Refinement Modulator(SRM) for detail recovery, together with an adaptive weighting schemethat emphasizes hard regions during optimization. Since residual-deriveddifficulty priors are unavailable at test time, we adopt a teacher–studentscheme in which a privileged teacher derives difficulty priors from resid-ual cues only during training, while a latent conditional diffusion studentgenerates aligned priors conditioned on a single low-light input at infer-ence. Experiments demonstrate that DCASR consistently outperformsstate-of-the-art methods on multiple benchmarks. Code will be madepublicly available at github.com/mingzhuzhang1/DCASR-LLIE.

Industrial anomaly inspection is severely hindered by thescarcity of real anomalous data. Zero-shot industrial anomaly genera-tion addresses this by generating anomalies on specific products withoutrequiring any of their real anomalous images. However, existing meth-ods suffer from two critical limitations, i.e., inaccurate anomaly informa-tion acquisition and uncontrolled anomaly-product fusion. To overcomethese challenges, we propose DeCo, which decouples the anomaly struc-ture from its source product, and explicitly recouples it with the nor-mal textures of the target product. During anomaly information acquisi-tion, Dual-Routing Flow (DR-Flow) binds the texture-invariant anomalystructure to an abnormal token, while a parallel constraint, Product-Invariant Flow (PI-Flow), prevents the abnormal token from binding thesource product. During anomaly-product fusion, we propose a hybrid in-jection to recouple the acquired anomaly structure with the target prod-uct, and Product Compatibility Correction (PCC) to compensate for theincompatibility between the acquired anomaly structure and the product.Extensive experiments demonstrate that DeCo establishes a new state-of-the-art. Training downstream detection models on our generated datayields massive pixel AP improvements of 5.1% on MVTec AD and 8.2%on VisA. Code is available at https://github.com/HUST-SLOW/DeCo.

While autoregressive (AR) Multimodal Large Language Mod-els (MLLMs) excel at complex document-to-markup generation, theirsequential decoding causes severe latency. Conversely, existing paralleland diffusion-based methods accelerate inference but often struggle tomaintain strict structural dependencies, resulting in significantly higherparsing errors. To bridge this gap, we propose DARL (Diffusion AutoRe-gression with Look-ahead), a novel hybrid decoding framework. DARLintroduces a Sliding Diffusion Block (SDB) that maintains a verifiedAR prefix for strict syntactic correctness while speculatively generat-ing a localized window of future tokens in parallel. To optimize thesediscrete diffusion trajectories, we introduce Online Monte Carlo Tra-jectory Generation (OMTG) and Diffusion Trajectory Preference Opti-mization (DTPO). OMTG dynamically samples candidate paths basedon the model’s real-time state, effectively mitigating the exposure biasinherent in static heuristics. DTPO integrates immediate and look-aheadrewards to optimize current accuracy and facilitate future correctness,ensuring stable and rapid convergence of the predictive trajectory space.Experiments on OmniDocBench-1.5 and olmOCR-Bench demonstratethat DARL establishes a new Pareto frontier in document parsing. Itachieves up to a 2.3× speedup while matching or exceeding the state-of-the-art accuracy of AR baselines. The code and model are available athttps://github.com/SCUT-DLVCLab/DARL.

Recent advances in neural decoding have enabled the direct reconstruction of visual stimuli from non-invasive brain signals. However, reconstructing continuous video streams remains highly challenging due to the continuous spatial and semantic transformations inherent in dynamic scenes. Most methods rely on a "reconstruct-thendescribe" paradigm, where captions are generated from reconstructed video keyframes, which are highly prone to semantic drift. To overcome these challenges, we propose a novel decoding framework, CortexVideo, to reconstruct video from functional magnetic resonance imaging (fMRI). Inspired by the visual dual-stream hypothesis, CortexVideo introduces dual-guidance decoding strategy. Specifically, we leverage subject-adaptive semantic information to guide video reconstruction, while concurrently integrating keyframe perceptual weights. Experiments demonstrate that CortexVideo provides a cognitive enhancement to existing hierarchical models along two critical dimensions: video semantic understanding and spatial localization. The reconstructed videos show greater fidelity to the visual stimuli in both semantic and spatial aspects, achieving state-ofthe-art results.

3D scene inpainting is essential for reconstructing areas cor-rupted by occlusions or limited viewpoints. While recent methods lever-age Gaussian Splatting (GS) for efficient 3D editing, they often de-pend on precise multi-view segmentation masks and are inherently con-strained to object removal tasks. We propose CoIn, a novel frameworkthat bridges 2D inpainting models and 3DGS through a multi-stage con-sistency pipeline. Our approach first generates initial inpainted imagesusing a diffusion model, enabling the use of arbitrary-shaped masks anddiverse tasks like object insertion. We then introduce Reference Adap-tive GS with Feature Attention to reconstruct a coarse 3D scene byadaptively weighing towards a reference view (2D → 3D). This 3D rep-resentation provides geometric guidance to the diffusion process via GS-based Reference Feature Warping, ensuring multi-view consistency (3D→ 2D). Finally, a Texture-Enhancing Discriminator refines the 3D sceneto achieve high photometric realism (2D → 3D). Experiments show thatCoIn, effectively leveraging bidirectional information flow, achieves state-of-the-art performance and effectively handles both object removal andobject insertion with flexible mask input.

Personalization in emotion recognition (ER) is essential foran accurate interpretation of subtle and subject-specific expressive pat-terns. Recent advances in vision–language models (VLMs) such as CLIPdemonstrate strong potential for leveraging joint image–text represen-tations in ER. However, CLIP-based methods either depend on CLIP’scontrastive pretraining or on LLMs to generate descriptive text prompts,which are noisy, computationally expensive, and fail to capture fine-grained expressions, leading to degraded performance. In this work, weleverage Action Units (AUs) as structured textual prompts within CLIPto model fine-grained facial expressions. AUs encode the subtle muscleactivations underlying expressions, providing localized and interpretablesemantic cues for more robust ER. We introduce CLIP-AU, a lightweightAU–guided temporal learning method that integrates interpretable AUsemantics into CLIP. It learns generic, subject-agnostic representationsby aligning AU prompts with facial dynamics, enabling fine-grained ERwithout CLIP fine-tuning or LLM-generated text supervision. AlthoughCLIP-AU models fine-grained AU semantics, it does not adapt to subject-specific variability in subtle expressions. To address this limitation, wepropose CLIP-AUTT, a video-based test-time personalization method thatdynamically adapts AU prompts to videos from unseen subjects. By com-bining entropy-guided temporal window selection with prompt tuning,CLIP-AUTT enables subject-specific adaptation while preserving tempo-ral consistency. Our experiments on three challenging video-based subtleER datasets — BioVid, StressID, and BAH — indicate that CLIP-AUand CLIP-AUTT outperform state-of-the-art CLIP-based FER and TTAmethods. Code: https://github.com/osamazeeshan/CLIP-AUTT.

Temporal Action Segmentation (TAS) assigns action labelsto each frame by modeling the temporal structure of actions. While of-fline TAS leverages full temporal context, online TAS cannot directlyaccess future information during inference. Existing online TAS methodsmainly focus on modeling historical dependencies, but temporal ambigu-ity still arises due to the incomplete temporal dependencies, leading tofragmented predictions. To tackle this challenge, we propose Dual-PathTemporal Modeling (DPTM) for online TAS, a two-stage framework thatexplicitly models and transfers the future dependencies that are inacces-sible during online inference. In the first stage, we adopt an offline-styletraining scheme where bidirectional dependencies are decomposed intounidirectional historical and future components via a temporal decom-position strategy. In the second stage, we employ a predictive distillationstrategy that uses recent historical frames to predict the separated futuredependencies without violating causality. These designs enable DPTMto leverage more complete temporal context while remaining suitable foronline inference. Extensive experiments demonstrate that DPTM signifi-cantly improves temporal coherence in segmentation results and achievesstate-of-the-art performance, demonstrating the effectiveness of modelingdependencies beyond historical frames for reducing the temporal ambi-guity. The code is available at: https://github.com/zczheng777/DPTM/.

While mistake understanding in procedural video is advanc-ing toward increasingly complex multi-step tasks and diverse mistaketypes, it remains unclear whether current MLLMs can even recognize andexplain mistakes at individual step levels. This study proposes CookingMistake Recognition and Explanation (Cook-MRE), a dataset for assess-ing MLLMs in-depth performance in understanding cooking mistakeswithin each single-step. The evaluation uses an LLM-based approachalong two aspects: recognition (mistake presence and type classification)and explanation (error correction). By focusing on mistakes in the formof deviations from instructions, Cook-MRE reuses an existing cookingvideo dataset and synthesizes mistake samples through modification oftext instructions. It covers four basic operation elements and includes dif-ficulty labels based on visual distinguishability. Evaluation of 13 MLLMsshowed that our comprehensive metric, Full Pass Accuracy, was sub-stantially lower than mistake presence accuracy alone, indicating thedifficulty of explaining mistakes in detail even for single steps. Analysis byutilizing our type and difficulty labels showed notably low performanceon visually challenging problems, and scaling model size yields limitedimprovement on them. The dataset and code are publicly available athttps://github.com/Shun-Takashige/cook-mre-benchmark.

We present 360CityArena, a benchmark for evaluating theurban exploration capabilities of Embodied Agents within a photorealis-tic environment constructed from 360° videos. Existing outdoor bench-marks either lack sufficient photorealism or complexity, resulting in aconsiderable gap from real-world urban environments. 360CityArena isbuilt on a realistic reconstruction of the Akihabara district in Tokyo,Japan, using 602 360° video segments covering 85 streets, and consistsof 175 meticulously human-crafted tasks. It encompasses three task cat-egories: Environment Understanding, Path Reasoning, and Spatial Rea-soning, covering fundamental abilities required for urban exploration,such as localization, landmark search, path planning, and relational spa-tial reasoning, thereby enabling comprehensive evaluation in realisticurban scenes. Our evaluation using state-of-the-art LMM-based agentsshows that even the strongest model, Gemini-2.5 Flash, performs far be-low human level (human: 77.3% vs. Gemini-2.5 Flash: 17.1%), revealingsubstantial challenges that remain in city-scale embodied navigation andreasoning. 360CityArena provides a necessary and challenging testbed forphotorealistic urban-district navigation and spatial reasoning.

Building a unified visual tokenizer is essential for bridgingthe gap between visual understanding and generation. Yet existing ap-proaches struggle with the inherent conflict between these tasks, as asingle token space is forced to support both high-level semantic abstrac-tion and low-level pixel reconstruction. We propose WinTok, a concisehybrid tokenizer that achieves a win-win performance by explicitly de-coupling the two objectives. WinTok supplements pixel tokens with aset of learnable semantic tokens, effectively mitigating cross-task inter-ference without incurring the computational overhead of dual tokenizers.To further enhance understanding capability, we introduce an asymmet-ric token distillation mechanism: the semantic tokens are guided by pre-trained semantic embeddings from any visual foundation model, enablingthem to inherit strong discriminative power while maintaining flexibility.Across 10 challenging benchmarks, WinTok delivers consistent improve-ments in reconstruction, understanding, and generation. Trained on only50M open-source data, WinTok surpasses the strong baseline UniTok by11.2% in classification accuracy and achieves a competitive reconstruc-tion rFID of 0.41, despite using substantially less training data. Code isreleased at https://github.com/WeChatCV/WinTok.

When does 3D Gaussian Splatting (3DGS) recover the truescene surface rather than just overfitting view-dependent appearance? Weanswer this by developing a mathematical framework—based on a first-hitrendering abstraction—that cleanly isolates geometry from appearance.We prove that geometric misalignment forcefully converts spatial texturesinto high-frequency angular signals via parallax. This establishes a strictidentifiability window: if angular capacity is bounded, surface-consistentsolutions are mathematically preferred; if unrestricted, the same imagescan be perfectly explained by an incorrect, opaque “billboard” geometry.Experiments on synthetic stress tests confirm this prediction, showingbillboard failures emerge precisely at high angular capacities. Conversely,in the real-world datasets we evaluate under standard capture proto-cols, reconstructions remain surface-consistent even at high SH degrees,which is consistent with the prediction that rich spatial texture can pushbillboard solutions outside the tested angular-capacity range.

Large Vision Language Models (LVLMs) have achieved re-markable success on vision–language tasks, yet fine-grained perceptionover high-resolution images and long-context videos remains challenging.As the number of visual tokens increases, the visual attention sink phe-nomenon becomes increasingly severe, causing irrelevant tokens to absorba disproportionate amount of attention mass. Recent approaches attemptto mitigate this issue by explicitly predicting bounding boxes or temporalspans and re-encoding the cropped visual regions. Such methods dependon unreliable numeric localization in the discrete token space and in-cur significant computational overhead due to additional forward passes.In this work, we propose VisReflect, a simple yet effective frameworkthat improves fine-grained perception in long visual contexts through la-tent visual reflection. Instead of decoding intermediate predictions intodiscrete tokens, the model generates continuous visual reflection thatrepresents question-relevant visual features in the latent space. Thesereflections selectively emphasize salient regions or frames, guiding at-tention towards relevant visual tokens within a single forward pass. Weconduct comprehensive evaluations on challenging high-resolution im-age benchmarks, including BLINK, V∗ , and HRBench-4K/8K, as wellas video understanding benchmarks such as MVBench, VideoMME, andMLVU. Our method consistently improves over strong baselines, achiev-ing gains of 4.1% on image benchmarks and 1.8% on video benchmarks.Compared with zooming-based methods, our model achieves compara-ble performance while reducing inference time by roughly 44% on videounderstanding.

GUI agents powered by vision-language models show strongpotential for automating digital tasks, yet frequently fail in long-horizonscenarios due to the absence of step-level verification. Existing process re-ward models verify actions through textual reasoning alone, missing thevisual nature of GUI state changes. We introduce VisCritic, a visualprocess reward framework that verifies agent actions by directly com-paring pre-action and post-action screenshots in visual feature space.VisCritic employs a Siamese vision transformer to extract change-awarerepresentations, coupled with an Action-Aware Critic Head that jointlyevaluates action success, task progress, and error type. A critic-trainingdata construction pipeline generates weakly supervised samples from ex-isting trajectories without additional human labels for critic training.Experiments and offline analyses across five benchmarks demonstratethat VisCritic serves as a plug-and-play enhancement for diverse GUIagents, generally improving benchmark metrics while providing visualdiagnostic cues.