ECCV 2026 Papers

Existing diffusion-based methods have recently made signif-icant progress in image dehazing. However, they typically neglect thephysics of haze formation and reconstruct clean images from pure Gaus-sian noise, thereby limiting their restoration potential. To address thisissue, we propose Haze-Noise Diffusion (HNDiff), a novel diffusion frame-work that embeds the atmospheric scattering model as an inductive bias.By grounding diffusion in physical principles, HNDiff ensures that therestoration aligns more closely with underlying mechanisms of haze for-mation. In its forward process, we introduce joint haze-noise diffusionwith a haze-aware noise scheduler, which progressively adds both hazeand noise to an image. Essentially, the scheduler adapts noise levels ac-cording to haze density, meaning that regions with heavier haze receivestronger noise injection to encourage content generation, while clearer re-gions receive lighter noise to better preserve details, which directly linksthe forward degradation process with the physics of haze. In the reverseprocess, we then derive a physically consistent dehazing-denoising pro-cess that simultaneously removes haze and noise to restore a clean imagein a manner aligned with the forward degradation process. To furtherenhance practicality, we propose Latent HNDiff, which compiles cleanlatent priors that can be seamlessly integrated into existing dehazing net-works to boost performance. Extensive experiments show that our worksignificantly improves leading dehazing backbones and achieves state-of-the-art results on benchmark datasets. The project page is available athttps://jin-ting-he.github.io/HNDiff/.

Reconstructing 3D scenes of objects displayed in glass enclo-sures remains a significant challenge for 3D Gaussian Splatting (3DGS),particularly in museum and exhibition hall settings where transparentprotective cases are ubiquitous. The high specular reflections (HSR) gen-erated by glass surfaces severely degrade reconstruction quality, yet theirunderlying properties and impact on 3DGS remain poorly understood.Through systematic investigation, we reveal two insights into HSR be-havior in 3DGS: (1) Geometry-inconsistency property: HSR causes 3Dpoints to exhibit drastically different appearances across viewpoints, fun-damentally violating the geometric consistency assumption that under-pins 3DGS; (2) Concept-preservation property: Despite corrupting lo-cal appearance, HSR has remarkably limited influence when learninghigh-level object concepts, as the true object structure remains intactbeneath the reflective layer. Motivated by these insights, we proposeGlassGS, a geometry and concept-aware 3DGS framework with an it-erative two-stage optimization. First, we compute geometry-consistencymaps for each input view to evaluate the appearance variations acrossviewpoints, explicitly identifying HSR-affected pixels by measuring theappearance variations of 3D points across views. Second, we introducea geometry-inconsistency-guided object concept learning module thatleverages these maps to learn robust object representations while activelysuppressing reflection interference. The two stages mutually reinforceeach other through iterative refinement, improving both inconsistencydetection and concept learning. For this new task, we collect the firstdataset specifically designed for 3DGS reconstruction of glass-enclosedobjects. Extensive experiments demonstrate that GlassGS significantlyoutperforms state-of-the-art methods, enabling high-fidelity digitizationof glass-enclosed scenes and opening new possibilities for virtual museumexperiences.

Accurate Remaining Useful Life (RUL) prediction is funda-mental to Prognostics and Health Management, yet remains challeng-ing due to stochastic degradation and non-stationary vibration patterns.Existing image-based approaches predominantly rely on time–frequencyrepresentations (e.g., spectrograms or wavelet transforms), which cap-ture spectral energy variations but overlook the intrinsic geometric struc-ture of system dynamics. In this work, we reformulate RUL predictionthrough a phase-space-inspired visual representation. We introduce thePhase Space Density Image (PSDI), a novel representation that encodesthe spatial density of time-delay embedded trajectories. Unlike spectralheatmaps, PSDI characterizes degradation as a progressive geometric dis-persion of the system attractor, revealing structural transitions invisibleto conventional representations. To ensure that these geometric changesreflect true degradation rather than coordinate drift, we further proposea Globally Anchored Reconstruction strategy that enforces consistentphase-space alignment across samples and time. We evaluate the effec-tiveness of the proposed representation across multiple vision architec-tures and integrate it with a compact knowledge distillation frameworkthat transfers useful visual priors while reducing prediction jitter throughtemporal aggregation. Experiments on benchmark datasets demonstratethat PSDI achieves competitive and robust performance compared withclassical signal representations and state-of-the-art time-series models,supporting phase-space density maps as an effective visual representa-tion for vibration-based degradation tracking.

Generating interactive and dynamic 4D scenes from a single static image remains a core challenge. Most existing generate-then-reconstruct and reconstruct-then-generate methods decouple geometry from motion, causing spatiotemporal inconsistencies and poor generalization. To address these, we present MoGe4D (Motion and GeometryAware image-to-4D Synthesis), a geometry-conditioned framework for single-image 4D synthesis that models a scene as dense 4D point trajectories. Instead of treating geometry and dynamics as two disconnected stages, our method starts from an initial geometric prior inferred from the input image and predicts future time-varying trajectories in a diffusion process, improving spatiotemporal coherence while preserving structural stability. To support this task, we first introduce TrajScene-60K, a large-scale dataset of 60,000 video samples with dense 4D point trajectories, addressing the scarcity of high-quality training data for scene-level 4D generation. Built on this, our diffusion-based 4D Scene Trajectory Generator (4D-STraG) predicts geometry-consistent and motionplausible trajectory fields conditioned on the input image, with a depthguided motion normalization strategy to reduce scale ambiguity and a Motion Perception Module (MPM) to inject motion-aware priors. We further propose a 4D View Synthesis Module (4D-ViSM) to render the generated 4D representation into videos under arbitrary camera trajectories. Experiments show that MoGe4D produces high-quality 4D scenes with strong temporal coherence, favorable geometry-aware consistency, and compelling novel-view synthesis from a single image. Code: https://github.com/Zhangyr2022/MoGe4D.

Scalable Vector Graphics are a fundamental medium for reso-lution-independent visual content, yet the deep learning community lacksa continuous, dense, and invertible latent space for vector representa-tions, the kind of foundational building block that Variational Autoen-coders and their descendants have long provided for raster images. Weintroduce SLS (SVG Latent Space), a Transformer-based autoencoderthat learns compact dense representations of individual SVG paths, theatomic visual elements from which any SVG image can be composed. Bymodeling SVG commands, coordinate data, and visual properties withina unified BPE-based token vocabulary, SLS learns fixed-size latent rep-resentations that jointly capture structure and appearance, and can bedecoded back into valid, style-consistent SVG paths with high fidelity.The resulting embedding space is robust, invertible, and structured: em-beddings lie on a unit hypersphere, enabling efficient similarity search,composition, and downstream conditioning through simple vector-spaceoperations. Finally, we demonstrate that SLS generalizes across diversetasks reducing their FLOPs by over 150× compared to token-based ap-proaches, and establishing a general-purpose latent foundation for vectorgraphics research.

Inference-time scaling for text-to-image generation has pro-gressed from simple Best-of-N (BoN) sampling to guided search meth-ods that verify and steer candidate trajectories at intermediate denoisingsteps. These approaches focus on when and how often to verify duringdenoising but largely treat the cost of generation itself as fixed. More-over, the standard practice of comparing methods by number of functionevaluations (NFEs) counts only denoising forward passes and ignores ver-ifier overhead, which can distort e!ciency rankings. We show that underwall-clock evaluation, simple BoN already matches or outperforms sev-eral guided search techniques, suggesting that compute is better spent onbroader exploration than on repeated intermediate verification. This mo-tivates Flash-BoN, which generates a large pool of inexpensive draft can-didates by combining three complementary acceleration knobs: timesteptruncation, layer skipping, and activation proxies into a single config-uration optimized once per model. An e!cient multi-stage verificationprocedure then identifies the most promising draft, which is refined at fullquality. Across three benchmarks and three model scales, Flash-BoN con-sistently outperforms all baselines under fixed wall-clock budgets, withgains that grow at larger model scales (+8% AUC). We further showthat our strategy combines well and improves existing orthogonal tech-niques such as reflection-based prompt optimization (+16% AUC). Thegains correlate with increased candidate diversity, which also enablesdraft-guided selection to accelerate RL post-training convergence.

Localizing document tampering is extremely challenging, asmanipulations are crafted to appear visually consistent and often leaveonly subtle traces that are nearly invisible to the human eye. In priorwork, evaluation has been largely dominated by synthetic benchmarksthat closely match the training distribution, and methods have shownsteady progress under this setting. However, these gains often translatepoorly to human-made forgeries and to cross-domain evaluation, whereboth the source documents and the tampering pipeline can change, lead-ing to a distribution shift. In addition, since the introduction of the Fre-quency Perception Head for the discrete cosine transform (DCT) modal-ity, it has become a standard choice, and subsequent work has largelyfocused on downstream modules and fusion strategies rather than revis-iting the backbone itself. To help close this gap in cross-domain perfor-mance and improve the DCT backbone design, we propose DiffNet, arelatively simple yet effective RGB–DCT early-fusion architecture drivenby two key design choices. First, to ensure that the decoder aggregatesmulti-scale inconsistency evidence rather than operating on raw, content-heavy activations, we apply a lightweight multi-level discrepancy trans-formation at the output of each backbone stage, replacing features withmagnitude-only responses to learned zero-sum filters. Second, we designan efficient DCT-domain backbone that relies on a lightweight frequency-index-aware DCT–quantization joint embedding. Our approach achievesstate-of-the-art performance on cross-domain and human-made docu-ment tampering localization, outperforming prior methods by around30%, with up to 7× higher throughput than the previous best model.

Modeling human-object interactions (HOI) from an egocen-tric perspective is a critical yet challenging task, particularly when re-lying on sparse signals from wearable devices like smart glasses andwatches. We present ECHO, the first unified framework to jointly recoverhuman pose, object motion, and contact dynamics solely from head andwrist tracking. To tackle the underconstrained nature of this problem,we introduce a novel tri-variate diffusion process with independent noiseschedules that models the mutual dependencies between the human, ob-ject, and interaction modalities. This formulation allows ECHO to oper-ate with flexible input configurations, making it robust to intermittenttracking and capable of leveraging partial observations. Crucially, it en-ables training on a combination of large-scale human motion datasetsand smaller HOI collections, learning strong priors while capturing in-teraction nuances. Furthermore, we employ a smooth inpainting inferencemechanism that enables the generation of temporally consistent interac-tions for arbitrarily long sequences. Extensive evaluations demonstratethat ECHO achieves state-of-the-art performance, significantly outper-forming existing methods lacking such flexibility.

Volumetric video seeks to model dynamic scenes as tem-porally coherent 4D representations. While recent Gaussian-based ap-proaches achieve impressive rendering fidelity, they primarily emphasizeappearance but are largely agnostic to instance-level structure, limit-ing stable tracking and semantic reasoning in highly dynamic scenarios.In this paper, we present Director, a unified spatio-temporal Gaussianrepresentation that jointly models human performance, high-fidelity ren-dering, and instance-level semantics. Our key insight is that embeddinginstance-consistent semantics naturally complements 4D modeling, en-abling more accurate scene decomposition while supporting robust dy-namic scene understanding. To this end, we leverage temporally alignedinstance masks and sentence embeddings derived from Multimodal LargeLanguage Models to supervise the learnable semantic features of eachGaussian via two MLP decoders, enabling language-aligned 4D repre-sentations and enforcing identity consistency over time.To enhance temporal stability, we bridge 2D optical flow with 4D Gaus-sians and finetune their motions, yielding reliable initialization and re-ducing drift. For the training, we further introduce a geometry-awareSDF constraints, along with regularization terms that enforces surfacecontinuity, enhancing temporal coherence in dynamic foreground model-ing. Experiments demonstrate that Director achieves temporally coher-ent 4D reconstructions while simultaneously enabling instance segmen-tation and open-vocabulary querying.

Current dexterous grasp planners primarily optimize for phys-ical stability, focusing on whether an object can be grasped rather thanhow it should be grasped to support downstream functional tasks. How-ever, conditioning grasp synthesis on specific human grasp taxonomiestypically requires prohibitively expensive, object-annotated datasets. Toaddress these limitations, we propose CoToGrasp, a novel generativeframework that synthesizes diverse, stable grasps strictly conditioned onspecific contact topologies. To bypass the data collection bottleneck, Co-ToGrasp is trained entirely in an object-agnostic manner. We introducea feature-based canonical workspace that projects local object featuresinto a unified gripper-centric domain, effectively decoupling the semanticfunctional intent from the arbitrary object geometry. By learning the in-trinsic contact manifold of the gripper within this workspace, our modelachieves zero-shot generalization to unseen objects at inference. Exten-sive evaluations on the large-scale DexGraspNet dataset demonstratethat CoToGrasp achieves state-of-the-art performance, outperformingexisting taxonomy-guided planners. Finally, we demonstrate the physicalviability and kinematic feasibility of our synthesized contact topologieson a physical robot platform. Code is available on our project websitehttps://cea-list.github.io/cotograspweb/.

Text-to-image models produce images that align well withnatural language prompts, but compositional generation has long beena central challenge. Models often struggle to satisfy multiple conceptswithin a single prompt, frequently omitting some concepts and resultingin partial success. Such failures highlight the difficulty of jointly op-timizing multiple concepts during reward optimization, where compet-ing concepts can interfere with one another. To address this limitation,we propose Correlation-Weighted Multi-Reward Optimization (CMO),a framework that leverages the correlation structure among concept re-wards to adaptively weight each attribute concept in optimization. Byaccounting for interactions among concepts, CMO balances competingreward signals and emphasizes concepts that are partially satisfied yetinconsistently generated across samples, improving compositional gener-ation. Specifically, we decompose multi-concept prompts into pre-definedconcept groups (e.g., objects, attributes, and relations) and obtain re-ward signals from dedicated reward models for each concept. We thenadaptively reweight these rewards, assigning higher weights to conflictingor hard-to-satisfy concepts using correlation-based difficulty estimation.By focusing optimization on the most challenging concepts within eachgroup, CMO encourages the model to consistently satisfy all requestedattributes simultaneously. We apply our approach to train state-of-the-art diffusion models, SD3.5 and FLUX.1-dev, and demonstrate consis-tent improvements on challenging multi-concept benchmarks, includingConceptMix, GenEval 2, and T2I-CompBench.

As audio-visual generative models evolve into world simula-tors, cross-modal synchronization stands as a critical proxy for assessingthe consistency of world dynamics and causality in generated content.However, existing evaluation metrics presume structural correctness, re-ducing synchronization to mere temporal alignment. Consequently, theyfail on generative outputs, especially when exhibiting structural hallu-cinations and asymmetric cross-modal relations, which currently man-date expert human annotation to assess synchronization. Thisdependency introduces a critical paradox: human evaluators rely on rel-ative, reference-dependent comparisons, whereas automated metrics re-quire reference-free, absolute scalars. We resolve this paradox by propos-ing a framework that distills relative human perception into a continu-ous, globally consistent metric. First, we introduce SynthSync, a datasetof generative failures ranked via pairwise human annotations. Second,we adapt the Omni-LLM equipped with a continuous latent projectionto translate relative human rankings into continuous absolute values.Third, we propose Real-Valued Group Relative Policy Optimization (R-GRPO) to internalize the global causal structure of synchronization vialistwise score distributions. Empirically, our metric achieves state-of-the-art human preference alignment. We leverage this estimator to establisha standardized benchmark, advancing AV-Gen assessment from low-levelsignal correlation to visually grounded causality.

Cellular vehicle-to-everything (C-V2X) enables cooperativeperception, prediction, and planning beyond the field of view of individualagents. However, existing datasets often overlook the complexities of real-world deployment, such as limited communication bandwidth and its dy-namics, heterogeneous sensing modalities, and scalability beyond a singlecooperative partner. In this paper, we introduce CooperScene, a high-fidelity cooperative autonomy dataset with real-world C-V2X communica-tion characterization. The dataset is organized into diverse scenes, includ-ing intersections, highway ramps, and parking lots. These scenes involvethree connected and autonomous vehicles (CAVs) and one infrastructureroadside unit (RSU), all equipped with multi-modal sensors and commer-cial off-the-shelf C-V2X communication radios. All scenes are annotatedwith globally consistent 3D labels at 10 Hz, totaling 344K objects across59K frames, underpinned by tight sensor- and agent-synchronization,centimeter-level localization and spatial alignment, precise cross-modalitycalibration, and 3GPP-standard-compliant C-V2X communication. Coop-erScene establishes a rigorous benchmark for evaluating multi-agentscaling and actual performance in real-world deployable settings. Projectwebsite for data and benchmark: https://cisl.ucr.edu/CooperScene.

Physics-based Human-Scene Interaction (HSI) imitation learn-ing is crucial for embodied intelligence as it bridges the gap betweenkinematic 3D motions and real-world dynamics. However, most exist-ing methods focus on simplified scene settings, leaving complex envi-ronments largely unexplored, which limits their applicability in real-world scenarios. In this paper, we focus on HSI mimicry in complexenvironments. Under this complex setting, we observe an inherent trade-off between successfully performing interaction and maintaining natu-ral, physically plausible motions. To address this challenge, we proposeComplexMimic, a framework that reconstructs diverse HSI by inter-preting imperfect MoCap data. First, we introduce a Dual Flow Strat-egy, which learns two complementary experts: an imitation expert foraccurate motion tracking and an interaction expert for collision-awareadaptation in complex scenes. Second, naive multi-expert distillation,which treats all experts equally, often under-samples challenging be-haviors, limiting effective learning. To mitigate this issue, we proposea difficulty-aware distillation strategy that adaptively weights super-vision and prioritizes hard-yet-learnable trajectories guided by failurestatistics and learning progress signals. Extensive experiments on threebenchmark datasets demonstrate that our approach outperforms cur-rent state-of-the-art methods. Our implementation is available at https://github.com/LuPan23/ComplexMimic.

Pose-free feed-forward 3D Gaussian Splatting enables novelview synthesis from uncalibrated multi-view images. Although more viewsshould improve performance, existing methods often degrade with dense-view inputs because global aggregation spreads attention over many to-kens, and naive voxel fusion averages many Gaussians into overly smoothrepresentations. We present AVSplat, a framework that turns additionalviews into reliable signals for both aggregation and representation. Be-fore global attention, each view performs a single lightweight interactionwith a small set of Assist Views chosen for relevance and diversity, andthe cached features provide a focused scene context that stabilizes corre-spondence. For representation, we use adaptive temperature-aware voxelfusion that sharpens attribution under high occupancy, guided by oc-cupancy and point confidence. Crucially, AVSplat restores positive viewscaling where performance remains stable or improves as more inputviews are added, instead of degrading in the dense-view regime. Abla-tions show that Assist View Preconditioning is primarily responsible forpreventing dense-view degradation, while Occupancy-guided Voxel Fu-sion contributes most of the single-point image-quality gains.

Existing red-teaming studies on GUI agents face two funda-mental limitations: adversarial perturbations require white-box accessunavailable in commercial deployments, while prompt injection is increas-ingly neutralized by stronger safety alignment. To study robustness undera more practical threat model, we propose Semantic-level UI ElementInjection, a black-box red-teaming paradigm that overlays safety-alignedand harmless UI elements onto screenshots to misdirect the agent’s visualgrounding. Our method couples a modular Editor–Overlapper–Victimpipeline with iterative search that samples multiple candidate edits, keepsthe best cumulative overlay, and adapts future prompt strategies basedon previous failures. Experiments across 19 victim models spanning 8model families show that strategic optimization substantially outperformsrandom injection (3.5–6.9× on the most robust victims) and transfersnear-perfectly across architectures, confirming model-agnostic visual-semantic vulnerabilities. After the first successful attack, the victim stillclicks the attacker-controlled icon in over 15% of subsequent independenttrials versus below 1% for random injection, establishing that strategicallyplaced icons act as persistent attractors that causally redirect groundingrather than introducing incidental clutter. Public code is available athttps://github.com/HashTAG00002/UI-Injection.

Autonomous personal mobility benefits greatly from the abil-ity to predict executable spatial goals from language and front view imagealone. In this study, we focus on absolute target position prediction froma language instruction and front view image, which requires referring ex-pression disambiguation and absolute target localization without depthcues. To address this, we propose AiSCREAM, a language-conditionedtarget localization model based on cross-view vision-language reason-ing. AiSCREAM constructs cross-view geometric semantic alignment be-tween the front and an Aerial Semantic view, which is an overhead-viewimage synthesized from the front image via a text-conditioned image-to-image generation model. By preserving appearance-level scene charac-teristics, the alignment can leverage language-conditioned semantic cuesand enable reliable absolute target localization from a single image. Toevaluate AiSCREAM, we constructed DRAMATiST, a benchmark thatintroduces road and traffic diversity. The experimental results demon-strated that AiSCREAM achieved a mean absolute position error of4.03 m on DRAMATiST, which outperformed baseline methods and hu-man performance. The project page is available at https://ai-scream-project-page.vercel.app/ .

Vision-and-Language Navigation (VLN) requires an embod-ied agent to ground complex natural-language instructions into long-horizon navigation in unseen environments. While Vision-Language Mod-els (VLMs) offer strong 2D semantic understanding, current VLN sys-tems remain constrained by limited spatial perception, 2D–3D repre-sentation mismatch, and monocular scale ambiguity. In this paper, wepropose AgentVLN, a novel and efficient embodied navigation frameworkthat can be deployed on edge computing platforms. We formulate VLNas a Partially Observable Semi-Markov Decision Process (POSMDP)and introduce a VLM-as-Brain paradigm that decouples high-levelsemantic reasoning from perception and planning via a plug-and-playskill library. To resolve multi-level representation inconsistency, we de-sign a cross-space representation mapping that projects perception-layer3D topological waypoints into the image plane, yielding pixel-alignedvisual prompts for the VLM. Building on this bridge, we integrate acontext-aware self-correction and active exploration strategy to recoverfrom occlusions and suppress error accumulation over long trajectories.To further address the spatial ambiguity of instructions in unstructuredenvironments, we propose a Query-Driven Perceptual Chain-of-Thought(QD-PCoT) scheme, enabling the agent to actively query and acquiregeometric depth information according to task demands. Finally, we con-struct AgentVLN-Instruct, a large-scale instruction-tuning dataset withdynamic stage routing conditioned on target visibility. Extensive experi-ments show that AgentVLN consistently outperforms prior state-of-the-art (SOTA) methods on long-horizon VLN benchmarks, offering a prac-tical paradigm for lightweight deployment of next-generation embodiednavigation models. Code: https://github.com/Allenxinn/AgentVLN.

Human demonstrations are widely considered the corner-stone of end-to-end (E2E) autonomous driving despite human demon-stration’s scarcity for long-tail and safety-critical scenarios. Nonetheless,current E2E autonomous driving (AD) training paradigms continue torely on human demonstrations. Imitation learning (IL) requires humandemonstrations for training, whereas reinforcement learning (RL) hasemerged as a promising alternative to reduce this dependency. How-ever, most existing RL methods for E2E AD still rely implicitly on hu-man demonstrations. A pure rewards-based RL method can overcomethe need for human demonstrations, but general RL policy gradientmethods suffer from the cold-start problem. In this paper, we proposeZTRS (Zero-human demonstration end-to-end autonomous driving withTRajectory Scorer) — a complete RL-based E2E planning paradigmtrained solely on real-world images and rule-based rewards, entirely with-out human demonstration. Through our proposed Exhaustive PolicyOptimization (EPO), a policy gradient variant tailored for enumer-able trajectory actions and dense supervision, ZTRS enables the modelto generalize better to long-tail driving scenarios. We demonstrate thisgeneralization through our SOTA performance against IL approaches onboth long-tail Navhard and closed-loop HUGSIM datasets. Project page:https://zhenxinli.net/ZTRS/.