ECCV 2026 Papers

Video generation models produce visually coherent content but struggle with tasks requiring spatial reasoning and multi-step planning. Reinforcement learning (RL) offers a path to improve generalization, but its effectiveness in video reasoning hinges on reward design—a challenge that has received little systematic study.WeinvestigatethisproblembyadaptingGroupRelativePolicyOptimization (GRPO)toflow-basedvideomodelsandtrainingthemonmaze-solvingandrobotic navigationtasks.Wefirstshowthatmultimodalrewardmodelsfailcatastrophically in this setting. To address this, we design verifiable reward functions grounded in objective task metrics. For structured game environments, we introduce a multicomponent trajectory reward that is verifiable against ground-truth optimal paths. For robotic navigation, where ground truth is given only as a reference rollout, we propose an embedding-level reference-anchored reward. Our experiments show that RL fine-tuning with verifiable rewards improves generalization. For example, on complex 3D mazes, our model improves exact match accuracy by 29.1% over the SFT baseline, and on trap-avoidance tasks by 51.4%. Our systematic reward analysis reveals that verifiable rewards are critical for stable training, while multimodal reward models could lead to degenerate solutions. These findings establish verifiable reward design as a key enabler for robust video reasoning.

Communities are fundamental spatial units that shape urbanform and social life. Whether a residential compound is spatially openor enclosed affects mobility, access to public services, and equity, yetstudies of Chinese fengbi xiaoqu remain largely qualitative or small-scale,limiting reproducible city-scale analysis. We address this gap by introduc-ing GBA-GCs, a metropolitan-scale multimodal benchmark for locallygrounded gated/open community recognition in China’s Greater BayArea, covering 37,444 residential compounds with aligned boundary poly-gons, high-resolution satellite imagery, Chinese metadata, and structuredattributes, together with expert-verified labels, inter-annotator reliability,and official evaluation splits. Built on this benchmark, we present Multi-modal Classifier for Gated Community (MCGC), a vision-centricmultimodal framework based on DINOv3-SAT that fuses imagery, text,and structured cues via modality-aware cross-attention and adaptivegating to mitigate modality imbalance. MCGC consistently outperformsstrong unimodal and multimodal baselines. Finally, we apply the validatedmodel to metropolitan-scale mapping and report equity-oriented findingsincluding spatial clustering of GCs, privatized green space, and reducedpedestrian connectivity. The benchmark, code, and release documentationare available at https://github.com/MinweiZhao/GBA-GCs.

Simulation is indispensable for autonomous driving, yet thepronounced visual Sim2Real gap severely limits the cross-domain gen-eralization of perception models. To bridge this gap, we propose Cross-Domain Control Transfer (CDCT), a diffusion-based Sim2Real frame-work for translating synthetic images into realistic counterparts withoutpaired supervision. CDCT features a novel cross-domain score composi-tion mechanism that injects domain-agnostic geometry guidance, derivedfrom synthetic-domain score differences, into a real-world appearanceprior. This ensures strict geometric and semantic consistency withoutpaired real-world spatial conditions. However, naively incorporating mul-tiple geometric conditions often incurs a prohibitive parameter overhead.To address this, our Lightweight Multi-Condition Adapter (LMCA) mod-ule processes diverse rendering buffers simultaneously, eliminating thecomputational redundancy of standard multi-branch architectures. Ex-tensive experiments on the CARLA-to-Cityscapes benchmark demon-strate that CDCT yields substantial zero-shot performance gains ondownstream perception tasks. Crucially, joint training with our generateddata surpasses the performance of real-world-only training, showcasingits immense potential for autonomous driving.

We introduce UEval, a benchmark to evaluate unified mod-els, i.e., models capable of generating both images and text. UEval com-prises 1,000 expert-curated questions that require both images and textin the model output, sourced from 8 real-world tasks. Our curated ques-tions cover a wide range of reasoning types, from step-by-step guides totextbook explanations. Evaluating open-ended multimodal generation isnon-trivial, as simple LLM-as-a-judge methods can miss the subtleties.Unlike prior methods relying on multimodal Large Language Models(MLLMs) to rate image quality or text accuracy, we design a rubric-basedscoring system in UEval. For each question, reference images and text an-swers are provided to an MLLM to generate an initial rubric, consisting ofmultiple evaluation criteria, and human experts then refine and validatethese rubrics. In total, UEval contains 10,417 validated rubric criteria,enabling fine-grained automatic scoring. UEval is challenging for currentunified models: GPT-5-Thinking scores only 66.4 out of 100, while thebest open-source model reaches merely 49.1. We observe that reasoningmodels often outperform non-reasoning ones, and transferring reasoningtraces from a reasoning model to a non-reasoning model narrows thegap. This suggests that reasoning may be important for tasks requiringcomplex multimodal understanding and generation. UEval is availableat https://huggingface.co/datasets/zlab-princeton/UEval

We present TriFlow, a new generative approach for produc-ing compact 3D meshes with artist-like triangle topology directly frominput geometry conditions such as signed distance fields. Our key insightis to represent mesh topology as a nearest-vertex vector field (NVF) de-fined over the surface, where each point encodes its association to thenearest triangle vertex in the local barycentric frame. We train a latentflow-matching model to synthesize this field, enabling topology genera-tion conditioned on the input geometry. To extract a coherent mesh, wecluster surface regions using the generated NVF and guide a constrainedquadric error metric mesh simplification with topology-aware optimiza-tion. This yields output meshes that closely match the input geometrywhile exhibiting structured, artist-like connectivity. Experiments demon-strate that TriFlow achieves stronger generalization and significantlyimproved topology quality compared to state-of-the-art learning-basedapproaches, alongside 90% lower Chamfer Distance and an 8× speedup.

Generalized face anti-spoofing (FAS) has attracted increas-ing attention due to the need for robustness in unseen scenarios. Do-main adaptation methods can improve performance by leveraging target-domain information. However, in FAS, collecting spoof samples is signif-icantly more expensive than acquiring a few bonafide samples, makingit impractical to gather diverse attack types and finetune models foreach deployment scenario. To address this challenge, we propose Text-guided RelAtionship Modeling (TRAM), a finetuning-free test-time do-main adaptation approach for generalized FAS using only a few bonafidesamples. Since samples within a target domain share common domainfactors, we treat genuine sample as the anchor and model its relation-ship with incoming inputs to mitigate domain bias. To learn discrimina-tive spoof cues from these relationships, we introduce a Relative Image-Pair-Text Contrastive Learning strategy that leverages fine-grained FASrelation prompts. Furthermore, to compensate for the lack of detailedspoof annotations, we design a Fine-grained Relation Prompt Learn-ing strategy that mines low-level photometric discrepancies through pre-computed measurements to generate robust prompts. Extensive exper-iments on eight FAS datasets with typical variations demonstrate thatTRAM achieves consistent improvements over state-of-the-art methodswhile using only a few bonafide samples from the target domain (as fewas five). Comprehensive ablation studies further validate the effectivenessof each component.

Test-time adaptation (TTA) has emerged as a prominentstrategy for adapting vision-language models to distribution shifts dur-ing inference. We conduct a per-sample analysis of model predictionsbefore and after adaptation, and observe two failure modes in existingTTA methods that echo previous work. Adaptations are frequently neg-ligible, yielding no change in the model’s predictions, and more severely,they can be detrimental by flipping previously correct predictions to in-correct ones. This naturally raises a question: Can we identify and skipsuch negligible or harmful adaptations? In this work, we introduce a newproblem of selective adaptation, which aims to determine whether agiven test sample should undergo adaptation or be skipped. To this end,we propose Cross-Augmentation Similarity (CAS), a simple baseline thatperforms adaptation only when predictions across augmented views ex-hibit low similarity. Notably, CAS not only preserves but in some casesimproves overall accuracy, even when skipping nearly 85% of the adap-tation process. We hope other researchers will explore this new directionand surpass the performance of our baseline. Our code is available athttps://github.com/sirujiang/selective-adaptation.

Textual image classification remains challenging due to thediversity of document layouts and textual structures. We propose TiCRL(Textual Image Classification with Reinforcement Learning-based Cur-riculum Learning), which combines a text-independent difficulty mea-surer with an RL-based training scheduler. The measurer estimates sam-ple difficulty using visual-textual features and loss signals without re-quiring OCR, while the RL agent dynamically adapts the curriculumbased on the learner’s state. TiCRL substantially improves both accu-racy and data efficiency. On a lightweight CNN trained from scratch,it achieves 83.67% accuracy using only 82% of the training data, out-performing curriculum learning baselines by up to 13.8%p. Applied toa strong transformer baseline (DiT), TiCRL reaches 93.02% accuracyusing only 58% of the data, surpassing transformer baselines trained on80%. The learned curriculum policy generalizes to new datasets and ar-chitectures without retraining, demonstrating that RL-based curriculumlearning is an effective and efficient training strategy for textual imageclassification.

Deep representations across modalities are inherently inter-twined. In this paper, we systematically analyze the spectral character-istics of various semantic and pixel encoders. Interestingly, our studyuncovers a highly inspiring and rarely explored correspondence betweenan encoder’s feature spectrum and its functional role: semantic encodersprimarily capture low-frequency components that encode abstract mean-ing, whereas pixel encoders additionally retain high-frequency informa-tion that conveys fine-grained detail. This heuristic finding offers a uni-fying perspective that ties encoder behavior to its underlying spectralstructure. We define it as the Prism Hypothesis, where each datamodality can be viewed as a projection of the natural world onto ashared feature spectrum, just like the prism. Building on this insight,we propose Unified Autoencoding (UAE), a model that harmonizessemantic structure and pixel details via an innovative frequency-bandmodulator, enabling their seamless coexistence. Extensive experimentsdemonstrate that UAE effectively unifies semantic abstraction and pixel-level fidelity within a single latent space, achieving state-of-the-art per-formance. Moreover, we show that UAE can be directly applied to pixel-space modeling, significantly improving both FID and IS over the vanillaJIT baseline. Code is avaliable at: https://github.com/WeichenFan/UAE.

006 Content safety is a fundamental challenge for text-to-image 006007 (T2I) models, yet prevailing methods enforce a debilitating trade-off 007008 between safety and generation quality. We argue that mitigating this 008009 trade-off hinges on addressing systemic challenges in current T2I safety 009010 alignment across data, methods, and evaluation protocols. To this end, 010011 we introduce a unified framework for synergistic safety alignment. First, 011012 to overcome the flawed data paradigm that provides biased optimiza- 012013 tion signals, we develop LibraAlign-100K, the first large-scale dataset 013014 with dual annotations for safety and quality. Second, to address the my- 014015 opic optimization of existing methods focus solely on safety reward, we 015016 propose Synergistic Preference Optimization (T2I-SPO), a novel align- 016017 ment algorithm that extends the DPO paradigm with a composite re- 017018 ward function that integrates generation safety and quality to holistically 018019 model user preferences. Finally, to overcome the limitations of quality- 019020 agnostic and binary evaluation in current protocols, we introduce the 020021 Unified Alignment Score, a holistic, fine-grained metric that fairly quan- 021022 tifies the balance between safety and generative capability. Extensive 022023 experiments demonstrate that T2I-SPO achieves state-of-the-art safety 023024 alignment against a wide range of NSFW concepts, while better main- 024025 taining the model’s generation quality and general capability. This 025026 paper contains harmful text and image examples. 026

State-of-the-art vision models process images in their en-tirety, lacking the ability to selectively zoom in on relevant regions. Thislimitation is particularly acute in scenarios where processing must beconditioned on a specific task – such as instance detection, which re-quires localizing a specific object in a high-resolution, cluttered scene. Insuch settings, critical details are easily lost as images are often resizedto match the model dimensions and computational constraints. We in-troduce Coarse-to-Fine Glimpse-based Active Perception (CF-GAP), atask-driven front-end that enhances high-resolution processing of exist-ing instance detectors. CF-GAP selectively directs a sequence of limited-view glimpses across the scene, utilizing task information to iterativelyrefine focus on the most relevant regions. These localized regions arethen processed at high resolution by a downstream instance detector.By avoiding full-image processing and eliminating irrelevant confound-ing information, CF-GAP improves Average Precision (AP) by up to20% across various state-of-the-art instance detectors on the HR-InsDetand Robotools benchmarks, while further enabling lightweight detectorsto outperform their larger counterparts.

While recent video world models can generate highly real-istic videos, their ability to perform semantic reasoning and planningremains unclear and unquantified. We introduce Target-Bench, thefirst benchmark that enables comprehensive evaluation of video worldmodels’ semantic reasoning, spatial estimation, and planning capabil-ities. Target-Bench provides 450 robot-collected scenarios spanning 47semantic categories, with SLAM-based trajectories serving as motiontendency references. Our benchmark reconstructs motion from generatedvideos with a metric scale recovery mechanism, enabling the evaluationof planning performance with five complementary metrics that focus ontarget-approaching capability and directional consistency. Our evalua-tion result shows that the best off-the-shelf model achieves only a 0.368overall score, revealing a significant gap between realistic visual genera-tion and semantic reasoning in current video world models. Furthermore,we demonstrate that fine-tuning process on a relatively small real-worldrobot dataset can significantly improve task-level planning performance.

Recent geometric foundation models enable feed-forward in-ference for SLAM, but their predictions are strongly dependent on the in-put view set, which leads to geometric inconsistencies and trajectory driftwhen results are chained over long sequences. Online deployment furtherexposes a trade-off between the low latency of two-view tracking and theconstraint richness of multi-view inference. We introduce UniSim-SLAM,an integrated system that runs lightweight two-view keyframe trackingin the frontend and performs periodic multi-view submap refinementin the backend. To combine predictions defined in heterogeneous localcoordinates with inconsistent scales, we formulate a unified multi-levelfactor graph on Sim(3) that jointly optimizes global keyframe poses andsubmap poses. The graph integrates temporal view-to-view odometryedges, view-to-submap bridge edges with depth-statistics scale anchor-ing, and submap-to-submap tie and scale constraints to enforce consis-tent similarity relations across submaps. Experiments on TUM RGB-Dand 7-Scenes show that UniSim-SLAM achieves state-of-the-art accuracyin the uncalibrated setting, reducing trajectory error by 38.5% on TUMRGB-D and 45.9% on 7-Scenes compared to prior best results. Projectpage: https://vision3d-lab.github.io/unisim-slam/.

This paper addresses the problem of omnidirectional videointerpolation, which plays an essential role in applications such as vir-tual reality and immersive video enhancement. Existing video interpo-lation methods are not well-suited for spherical videos, as they havedifficulty handling severe distortions close to the poles. To address thisissue, we propose SVI360, a dual-branch framework that combines theimage frame and its rotated orthogonal view to deal with these dis-tortions. The core methodological aspect of the approach is to rein-force equivariance of the flow displacements between the original andorthogonal views to improve intermediate frame prediction. Experimentsshow that our method outperforms state-of-the-art approaches in inter-polation quality while maintaining accurate optical flow in four differ-ent public benchmarks. Code and pre-trained models are available at:https://icb-vision-ai.github.io/video360_interpolation/

Despite the remarkable success of large-scale pre-trained im-age representation models (i.e., vision encoders) across various visiontasks, they are predominantly trained on 2D image data and thereforeoften fail to capture 3D spatial relationships between objects and back-grounds in the real world, constraining their effectiveness in many down-stream applications. To address this, we propose SpatialBoost, a scalableframework that enhances the spatial awareness of existing pre-trained vi-sion encoders by injecting 3D spatial knowledge expressed in linguisticdescriptions. The core idea involves converting dense 3D spatial infor-mation from 2D images into linguistic expressions, which is then usedto inject such spatial knowledge into vision encoders through a LargeLanguage Model (LLM). To this end, we adopt a multi-turn Chain-of-Thought (CoT) reasoning process that progressively incorporates densespatial knowledge and builds hierarchical spatial understanding. To val-idate effectiveness, we adapt SpatialBoost to state-of-the-art vision en-coders such as DINOv3, and evaluate its performance gains on a widerange of benchmarks requiring both 3D perception and general visionabilities. For instance, SpatialBoost improves DINOv3 performance from55.9 to 59.7 mIoU on ADE20K, achieving state-of-the-art performancewith 3.8%p gain over the pre-trained DINOv3. Project page.

Skill determination in video understanding remains challeng-ing due to the intricate temporal dynamics of skill performance and thecomplex cognitive processes inherently underlying human evaluation. Ex-isting approaches often overlook the progressive evolution of evaluators’cognitive states and the influence of key action stages on overall per-formance. To address this limitation, we propose Sequential Primacyand Attribution Ranking Modeling (SPAR). SPAR captures the tem-poral evolution of cognitive states throughout the evaluation process,enabling iterative refinement of evaluative understanding and producingprogressively enhanced representations of overall performance quality.The framework integrates adaptive weighting based on the global tempo-ral context, emphasizing critical segments through localized interactionmodeling and multi-scale temporal aggregation. To correct inconsisten-cies in intermediate recognition signals, SPAR employs a bias-mitigationloss to ensure a faithful and consistent reflection of intrinsic action qual-ity. Extensive experiments on five public datasets validate the effective-ness and generalization capability of the proposed framework.

Choreography is a structured creative process grounded inmusical form, rather than a mere sequence of improvised movements.Professional choreographers design choreography following a set of chore-ographic primitives: (1) organizing movements at the phrase level, (2)refining movements by aligning them with musical beats and energy,and (3) stylizing phrases through semantic text descriptions. However,existing dance generation methods overlook these primitives, leadingto dances that feel structurally incomplete and resemble improvisationrather than professional choreography. Here, we present HiChor: Hi-erarchical Choreography generation from pop music, a framework in-spired by the structured workflow of professional choreographers, whichis grounded in choreographic primitives. HiChor hierarchically generateschoreography from pop music by (1) generating phrase-level choreogra-phy aligned with beats, (2) enhancing choreography by aligning withmusic’s energy, and (3) concatenating all phrase-level choreography andenhancing it through plausibility enhancement to ensure smooth transi-tions between phrases. Furthermore, we introduce a semantic text styl-ization module that converts semantic text descriptions into low-levelchoreography feature captions using an LLM, enabling semantic control.To assess how well the generated choreography adheres to choreographicprinciples, we introduce new metrics—Phrase Diversity, Feature Align-ment, and Text Stylization Score. Experiments demonstrate that HiChoroutperforms existing methods in both choreographic quality and seman-tic stylization.

Can we recover the 3D poses of multiple people using onlysound? This paper presents the first attempt to estimate multi-person3D poses solely from acoustic signals. Estimating the poses of multipleindividuals using acoustic signals is inherently challenging due to the su-perposition of motion-dependent signal variations. Unlike single-personscenarios, the presence of multiple subjects leads to overlapping acousticsignatures, making it difficult to attribute specific signal changes to anindividual’s pose. Furthermore, the complexity is compounded by inter-person reflections, which introduce intricate propagation delays that ob-scure the temporal motion-acoustic relationship. To address these issues,we propose SoundMHPE (Sound-based Multi-person Human Pose Es-timator), a novel encoder-decoder framework consisting of two key com-ponents. First, the Acoustic Multi-scale Encoder captures diverse tem-poral and fine-grained frequency features to isolate subtle acoustic sig-natures from complex, overlapping signals. Second, the Temporal PoseDecoder employs an attention mechanism to disentangle multi-personinformation across successive frames. By jointly accounting for tempo-ral dynamics and inter-person dependencies, this component preciselyreconstructs frame-wise individual poses. To validate our approach, weconstructed the 6-hour Acoustic Multi-person Pose (AMP) dataset con-sisting of 432K synchronized frames of multi-person pose and acous-tic data, and demonstrated that our SoundMHPE outperforms baselinemodels. Project page: https://oumi03.github.io/sound-mhpe/

Domain adaptive visual object tracking under adverse weatherconditions has garnered significant attention in recent years. Despitethe impressive performance, existing methods heavily rely on the large-scale video frames from both source and target domains, which is im-practical under rigid resource constraints where source data is unavail-able. To overcome this limitation, we propose SFDATrack, a general-ized source-free domain adaptive tracker that merely leverages adverseweather samples from the target domain for robust state estimation.Specifically, SFDATrack first employs a mean-teacher backbone withDual Interactive Mamba (DIM) blocks to distill the candidate targettokens that are resilient to weather variations from classified, augmentedsamples. Afterwards, we introduce a hyperspherical prototype projec-tion (HPP) module to project these tokens onto multi-domain proto-types within a latent hyperspherical space. By enforcing both domain-specific and domain-invariant properties of the multi-domain prototypes,SFDATrack can be seamlessly adapted to diverse weather conditionswith powerful generalizability. Extensive experiments evaluated on vari-ous benchmarks demonstrate that SFDATrack achieves superior perfor-mance compared to state-of-the-art approaches. The code is available athttps://github.com/watcherBR0/sfdatrack.