ECCV 2026 Papers

We introduce LaMP, a dual-expert Vision-Language-Actionframework that embeds dense 3D scene flow as a latent motion prior forrobotic manipulation. Existing VLA models regress actions directly from2D semantic visual features, forcing them to learn complex 3D physicalinteractions implicitly. This implicit learning strategy degrades underunfamiliar spatial dynamics. LaMP addresses this limitation by aligninga flow-matching Motion Expert with a policy-predicting Action Expertthrough gated cross-attention. Specifically, the Motion Expert generatesa one-step partially denoised 3D scene flow, and its hidden states condi-tion the Action Expert without full multi-step reconstruction. We eval-uate LaMP on the LIBERO, LIBERO-Plus, and SimplerEnv-WidowXsimulation benchmarks as well as real-world experiments. LaMP consis-tently outperforms evaluated VLA baselines across LIBERO, LIBERO-Plus, and SimplerEnv-WidowX benchmarks, achieving the highest re-ported average success rates under the same training budgets. On LIBERO-Plus OOD perturbations, LaMP shows improved robustness with an av-erage 9.7% gain over the strongest prior baseline. Our project page isavailable at https://summerwxk.github.io/lamp-project-page/.

Masked generative models have made significant advancesin text-to-motion synthesis. Previous approaches adopt a non-factorized,whole-body tokenization, limiting compositional and fine-grained con-trol. We present K-Mask, a kinematic-aware generative framework thatfactors motion into anatomically grounded groups and learns temporaland cross-group interactions through masked modeling. In the proposedapproach, a kinematic-group residual VQ-VAE (KG-RVQ) encodes mo-tion into disentangled, group-aligned latents. In particular, a latent-awarekinematic dropout (LAKD) loss is proposed to promote within-groupreconstruction and suppress inter-group leakage. Then, we introduce twobidirectional spatiotemporal transformers: (i) a base model that generatescoarse tokens over a time-group lattice with axis-factorized attention, and(ii) a residual model that predicts deeper quantizer layers for refinement.Finally, We introduce kinematic group masking, which randomly occludesentire group-specific subtrees, to enhance inter-group reasoning. On theHumanML3D and KIT-ML text-to-motion benchmarks, K-Mask obtainscompetitive FID scores of 0.041 and 0.154, respectively, while providingan anatomically factorized token interface for localized motion control.

Modern video generation models, despite high-quality frame synthesis, still struggle with motion fidelity and temporal consistency, causing artifacts and semantic drift; current solutions typically require extensive backbone retraining. We introduce a training-free framework for inference-time motion calibration that intervenes in the sampling trajectory of pretrained models to correct nascent inconsistencies efficiently. Our framework has two complementary components. First, Dynamic Latent Control (DLC) at selected steps estimates a clean latent, evaluates a VFM-based motion–semantic loss on proxy frames, and applies a corrective update to steer the trajectory. To control overhead, Amortization Trajectory Rectification (ATR) interleaves sparse calibration with fast amortized steps by caching a rectification field. Second, Dynamic RoPE Control (DRC) modulates temporal RoPE in attention blocks from the same motion signals, strengthening long-range coherence or sharpening short-range dynamics when flicker is detected. Extensive experiments on VBench and VBench 2.0 show that our method improves advanced motion fidelity metrics while incurring a small trade-off on superficial temporal smoothness for some backbones.

We present Generative Relightable Avatars (GRA), a personspecific method for photorealistic free-view rendering and environmentmap relighting of full-body humans. We postulate that modeling finegrained appearance details is inherently a one-to-many problem that can benefit from a generative formulation. In contrast to fully regressive relightable avatar methods, GRA follows a hybrid approach that combines controllable, physics-grounded relighting with probabilistic refinement. Starting from a tracked animated mesh, we optimize material parameters in UV-space and render a coarse relit appearance under a target HDR environment map. Next, we refine the textures with a feedforward model to capture pose-dependent texture dynamics and illumination effects beyond simplified reflectance assumptions. Finally, a finetuned video-to-video diffusion model transforms the physically grounded renderings into temporally coherent, high-detail videos while preserving 3D control, with an error-recycling strategy for generating long videos. Experimental evaluations demonstrate our method’s improved perceptual quality over prior relightable avatar baselines. We urge the readers to watch the supplementary video. See the project page for more details.

Diffusion models have made significant progress in both text-to-image (T2I) generation and text-guided image editing. However, thesemodels are typically built with billions of parameters, leading to high la-tency and increased deployment challenges. While on-device diffusionmodels improve efficiency, they largely focus on T2I generation and lacksupport for image editing. In this paper, we propose DreamLite, a com-pact unified on-device diffusion model (0.39B) that supports bothT2I generation and text-guided image editing within a single network.DreamLite is built on a pruned mobile U-Net backbone and unifies con-ditioning through in-context spatial concatenation in the latent space.It concatenates images horizontally as input, using a (target | blank)configuration for generation tasks and (target | source) for editing tasks.To stabilize the training of this compact model, we introduce a task-progressive joint pretraining strategy that sequentially targets T2I, edit-ing, and joint tasks. After high-quality SFT and reinforcement learning,DreamLite achieves GenEval (0.72) for image generation and ImgEdit(4.11) for image editing, outperforming existing on-device models and re-maining competitive with several server-side models. By employing stepdistillation, we further reduce denoising processing to just 4 steps, en-abling our DreamLite could generate or edit a 1024 × 1024 image in lessthan 3s on iPhone 17 pro. To the best of our knowledge, DreamLiteis the first unified on-device diffusion model that supports both imagegeneration and image editing. Our code and model are available at here.

Reasoning segmentation aims to predict pixel-wise masks fortargets given complex language queries. Existing approaches leverageMultimodal Large Language Models (MLLMs) for vision-language rea-soning and generate intermediate target cues (e.g., points or boxes) toguide a segmentation model. However, compressing rich reasoning intosparse cues often introduces ambiguity and noise, preventing these cuesfrom accurately preserving the reasoning intent. While multiple comple-mentary cues can enrich target information, existing methods typicallyfeed them jointly into a single segmentation process, allowing ambiguousor erroneous cues to affect the entire prediction. Therefore, we proposeDGSeg, a reasoning segmentation framework that learns to fuse pre-dictions guided by semantic and spatial cues. Specifically, the MLLMjointly reasons about both target identity and spatial location, produc-ing complementary semantic and spatial cues that are fed into separatesegmentation branches. Their predictions are adaptively integrated by alightweight dynamic gating module trained with relative branch-qualitysupervision to suppress noisy or conflicting regions. Extensive experi-ments demonstrate that DGSeg consistently outperforms strong base-lines on multiple benchmarks and achieves 69.6% and 67.3% gIoU on thechallenging ReasonSeg validation and test splits.

Object detection models deployed in safety-critical applica-tions remain vulnerable to backdoor attacks that cause targeted misbe-haviors when a hidden trigger is present. Existing detection methods ei-ther rely on trigger inversion or exploit architecture-specific assumptions,and critically, representative existing methods fail to generalize reliablyto scene-level attacks, where a single trigger induces anomalous behavioracross all objects in the scene simultaneously. We present DistScan, abackdoor detection framework based on a simple but previously unex-ploited observation: backdoor injection systematically shifts a model’spre-NMS prediction class distribution away from its training class fre-quencies, even on clean inputs without any trigger present. DistScanaggregates intermediate class predictions over a clean validation set andflags a model as backdoored if the resulting distribution deviates sig-nificantly from the training class frequencies, requiring no model weightaccess, no trigger knowledge, and no additional training. Extensive exper-iments on MS-COCO and PASCAL VOC across two architectures andthree scene-level attack scenarios demonstrate that DistScan substan-tially outperforms existing methods, improving average detection accu-racy over the best-performing applicable baseline by 27.32 percentagepoints.

From a Bayesian perspective, score-based di_x001B_usion solves in-verse problems through joint inference, embedding the likelihood withthe prior to guide the sampling process. However, this formulation failsto explain its practical behavior: the prior o_x001B_ers limited guidance, whilereconstruction is largely driven by the measurement-consistency term,leading to an inference process that is e_x001B_ectively decoupled from thedi_x001B_usion dynamics. We show that the di_x001B_usion prior in these solversfunctions primarily as a warm initializer that places estimates near thedata manifold, while reconstruction is driven almost entirely by measure-ment consistency. Based on this observation, we introduce DAPS++,which fully decouples di_x001B_usion-based initialization from likelihood-drivenre_x001C_nement, allowing the likelihood term to guide inference more directlywhile maintaining numerical stability and providing insight into whyuni_x001C_ed di_x001B_usion trajectories remain e_x001B_ective in practice. By requiringfewer function evaluations (NFEs) and measurement-optimization steps,DAPS++ achieves high computational e_x001E_ciency and robust reconstruc-tion performance across diverse image restoration tasks.

Hair simulation and rendering are challenging due to complex strand dynamics, diverse material properties, and intricate light–hair interactions. Recent video diffusion models can generate high-quality videos, but they lack fine-grained control over hair dynamics. We present ControlHair, a hybrid framework that integrates a physics simulator with conditional video diffusion to enable precise and controllable dynamic hair rendering. ControlHair adopts a three-stage pipeline: it first encodes physics conditions into per-frame geometry using a simulator, then extracts per-frame control signals, and finally feeds control signals into a video diffusion model to generate videos with desired hair dynamics. This cascaded design decouples physics reasoning from video generation, supports diverse physics, and makes training the video diffusion model easy. Trained on a curated 10K video dataset, ControlHair outperforms textand pose-conditioned baselines, delivering precisely controlled hair dynamics. We also demonstrate three use cases of ControlHair, including dynamic hairstyle try-on, bullet-time effects, and cinemagraphic. Project page: https://linwk20.github.io/controlhair-web.

Blind image deblurring demands the recovery of high-fidelitydetails and coherent structures from complex, unknown degradations.Current blind image deblurring methods struggle with real-world, spa-tially varying degradations, and lack the semantic awareness necessary toreliably differentiate valid textures from artifacts. To bridge this gap, wepropose CogSENet, a dynamic, semantic-aligned reconstruction frame-work inspired by the eagle’s visual system. By mimicking the eagle’sactive saccadic scanning, we devise a Semantic-Driven State Space Mod-ule (SDSSM) with semantic-aware token regrouping via differentiablerouting, enabling prompt-conditioned long-range dependency modeling.To ensure physically interpretable recovery of textures and structures,a BiFreqFusionBlock (BFFB) mirrors functional differentiation of theeagle’s retina by decomposing features into high and low frequenciesusing wavelet transforms. Finally, we estimate a continuous Blur Field(CBF) from blur image and fuse it with CLIP semantic priors to modu-late the deepest latent features, emulating focal adaptation and enablingadaptive restoration under spatially non-uniform blur. Extensive exper-iments demonstrate that CogSENet outperforms state-of-the-art deblur-ring methods in both visual quality and structural fidelity with fewerparameters, while also performing favorably on dehazing, deraining, anddenoising tasks.

Recent advances in Large Language Model (LLM)-based agentshave shown remarkable progress in code generation. However, currentagent methods mainly rely on text-output-based feedback (e.g. command-line outputs) for multi-round debugging and struggle in graphical userinterface (GUI) that involve visual information. This is mainly due totwo limitations: 1) GUI programs are event-driven, yet existing methodscannot simulate user interactions to trigger GUI element logic. 2) GUIprograms possess visual attributes, making it difficult for text-based ap-proaches to assess whether the rendered interface meets user needs. Tosystematically address these challenges, we first introduce InteractGUIBench, a novel benchmark comprising 984 commonly used real-worlddesktop GUI application tasks designed for fine-grained evaluation ofboth interaction logic and visual structure. Furthermore, we propose VF-Coder, a vision-feedback-based multi-agent system for debugging GUIcode. By perceiving visual information and directly interacting with pro-gram interfaces, VF-Coder can identify potential logic and layout issuesin a human-like manner. On InteractGUI Bench, our VF-Coder approachincreases the success rate of Gemini-3-Flash from 21.68% to 28.29% andraises the visual score from 0.4284 to 0.5584, indicating the effectivenessof visual feedback in GUI debugging.

Contrastive Language-Image Pre-training (CLIP) relies onVision Transformers whose attention mechanism is susceptible to spu-rious correlations and scales quadratically with resolution. To addressthese limitations, we present CLIMP, the first fully Mamba-based con-trastive vision-language model that replaces both the vision and text en-coders with state-space architectures. VMamba’s cross-scan mechanismcaptures spatial inductive biases that reduce reliance on spurious correla-tions, producing an embedding space with tighter cross-modal alignmentand lower hubness - geometric properties that translate to superior re-trieval and out-of-distribution robustness, surpassing even CLIP-ViT-Btrained on a dataset 167× larger on ImageNet-O. CLIMP naturally sup-ports variable input resolutions without positional encoding interpolationor specialized training, achieving up to 6.6% higher retrieval accuracyat 16× training resolution while using 5× less memory and 1.8× fewerFLOPs. Mamba’s autoregressive nature further enables processing of ar-bitrarily long text, overcoming CLIP’s fixed 77-token context limitationfor dense captioning retrieval. Our scaling experiments across model sizesand dataset sizes show consistent, unsaturated improvements - indicat-ing that CLIMP’s architectural advantages are not limited by trainingscale. These results demonstrate that Mamba is a compelling alternativeto Transformers for vision-language pre-training. The code and modelsare publicly available at https://github.com/NimrodShabtay/CLIMP

Implicit neural representation (INR) has emerged as a pow-erful prior for multi-dimensional data (e.g., multispectral images andvideos). However, most INR methods employing periodic activation func-tions (e.g., Sine) predominantly rely on function composition. This mech-anism introduces optimization instability as network depth increases,thereby limiting their performance. Meanwhile, these methods fail to in-corporate proper physical priors to effectively alleviate spectrum bias. Toaddress these issues, inspired by the commonalities between deep periodicnetworks and generalized Fourier series, we propose a novel CalibratedHarmonic Overlaid Implicit Neural Representation (CHOIR). Specifi-cally, we utilize Coordinated Harmonic Superposition (CHS) to replacethe conventional function composition used in most INRs, thereby en-suring optimization stability when scaling network depth. Furthermore,we introduce a Perceptual Spectrum Calibration (PSC) to mitigate spec-trum bias. This calibration embeds the ubiquitous power-law spectrumprior of natural images and adjusts the globally fixed spectrum towards aphysically plausible log-uniform distribution. Extensive experiments onvarious multidimensional data recovery problems demonstrate that ourmethod achieves superior performance over state-of-the-art approaches.Code is available at https://github.com/chorl0229/CHOIR.

Existing imitation learning methods for end-to-end autonomousdriving predominantly learn from successful demonstrations by minimiz-ing geometric deviations from expert trajectories. This paradigm implic-itly assumes that spatial proximity implies behavioral safety, leading toa critical objective mismatch: trajectories with nearly identical imita-tion losses may exhibit drastically different safety outcomes, where oneremains recoverable while the other results in collision. To address thislimitation, we propose BeyondDrive, a failure-aware imitation learn-ing framework that jointly learns from successful and failed driving be-haviors. First, we introduce a flow matching-based negative trajectory† Corresponding author. Code link: https://github.com/wjl2244/BeyondDrive.This work was supported by Beijing Natural Science Foundation-Xiaomi InnovationJoint Fund L253007, by Beijing Nova Program (202604841268) and Beijing NaturalScience Foundation under Grant 4242052.generator that synthesizes safety-critical yet expert-proximate trajecto-ries, enabling explicit modeling of safety asymmetry. Second, we developa diversity-aware sampling strategy that mitigates mode collapse andimproves coverage of diverse failure modes during negative trajectorygeneration. Third, we propose a Repulsive Distance Loss that simultane-ously attracts predictions toward expert demonstrations while repellingthem from hard negative trajectories, thereby establishing discrimina-tive safety boundaries in trajectory space. Applied to the uni-modalbaseline Latent TransFuser, BeyondDrive achieves 89.7 PDMS on theNAVSIMv1 closed-loop benchmark, outperforming prior state-of-the-artmethods. Moreover, BeyondDrive generalizes effectively across differentautonomous driving architectures, including multi-modal planners, andfurther demonstrates strong zero-shot transferability on the HUGSIMbenchmark.

Road surfaces serve as the only contact medium between ve-hicle wheels or robot feet and the physical world. Reconstructing the roadsurface is crucial for unmanned vehicles and mobile robots. Recent stud-ies on Neural Radiance Fields (NeRF) and Gaussian Splatting (GS) haveachieved remarkable results in scene reconstruction. However, they typi-cally rely on multi-view image inputs and require prolonged optimizationtimes. In this paper, we propose BEV-GS, a real-time single-frame roadsurface reconstruction method based on feed-forward Gaussian splatting.BEV-GS consists of a prediction module and a rendering module. Theprediction module introduces separate geometry and texture networksfollowing the Bird’s-Eye-View (BEV) paradigm. Geometric and textureparameters are directly estimated from a single frame, avoiding per-sceneoptimization. In the rendering module, we utilize grid Gaussian for roadsurface representation and novel view synthesis, which better aligns withroad surface characteristics. Our method achieves state-of-the-art per-formance on the real-world dataset RSRD. The road elevation error isreduced to 1.73 cm, and the PSNR of novel view synthesis reaches 28.36dB. The prediction and rendering FPS are 26 and 2061, respectively,enabling high-accuracy and real-time applications. Our code is availableat https://github.com/IRMVLab/BEV-GS.

Open-set action recognition (OSAR) requires a model toclassify known actions while rejecting unseen ones. Existing methodstypically use predictive uncertainty only at inference time, treating it asa rejection score after representation learning is complete. This leaves theunderlying representation unchanged and allows spurious background orcontext cues to remain embedded in the model, which weakens known–unknown separation. We propose a staged uncertainty-guided fine-tuningframework that uses ensemble-derived epistemic uncertainty not only forrejection, but also as a training signal for representation refinement. Theframework proceeds in three stages: motion-guided ensemble training tobias learning toward action-relevant evidence, in-distribution epistemicregularization to improve uncertainty reliability on known samples, anduncertainty attribution mask-based input attention (UAM-IA), whichconverts uncertainty attribution maps into input attention masks for re-finement. To improve deployment efficiency, we further distill the refinedensemble into a single student model trained to approximate both theensemble predictive distribution and the ensemble-derived epistemic un-certainty. Under a controlled OSAR protocol with UCF101 as the knownset and HMDB51 and MiT-v2 as unknown sources after removing over-lapping classes, the proposed method improves known–unknown separa-tion across multiple backbones. The distilled student preserves much ofthis benefit while enabling efficient single-model inference.

Inspired by text prompts in large language models, visualprompts have been explored to enhance the perceptual capabilities oflarge vision-language models (LVLMs). However, performance tends tosaturate under single visual prompt designs, making further prompt engi-neering increasingly ineffective. To address this limitation, we shift fromprompt engineering to prompt retrieval and propose AutoV, a lightweightframework for instance-adaptive visual prompt identification. Given aninput image and a textual query, AutoV automatically locates the mostsuitable visual prompt from a diverse candidate pool. Training such a re-trieval framework requires prompt-level supervision, yet prompt qualityis inherently ambiguous and difficult to assess reliably, even for humans.To enable automatic supervision, we evaluate visual prompts using a pre-trained LVLM and label them according to their prediction losses. Usingthe loss-oriented ranking as a robust training signal, AutoV learns to re-trieve the query-aware optimal prompt for each instance without manualannotation. Experiments indicate that AutoV enhances the performanceof various LVLMs on image understanding, captioning, grounding, andclassification tasks. For example, AutoV improves LLaVA-OV by 10.2%on VizWiz and boosts Qwen2.5-VL by 3.8% on MMMU, respectively.

Segmenting thin structures such as infrastructure cracks andanatomical vessels is hampered by topology-sensitive geometry, highannotation cost, and poor cross-domain generalization. We propose FMS2 ,a unified flow-matching framework with two complementary modules.SegFlow is a 2.96M-parameter encoder–decoder segmentation modelthat recasts prediction as continuous image→mask transport: it learnsa time-indexed velocity field with a single flow-matching regression lossand obtains the mask through deterministic ODE integration, ratherthan supervising only end-state logits. This trajectory-level supervisionimproves thin-structure continuity and sharpness over topology-awareloss baselines, without auxiliary topology heads, post-processing, or multi-term loss engineering. SynFlow is a mask-conditioned mask→imagegenerator that produces pixel-aligned image–mask pairs by injectingmask geometry at multiple scales, emphasizing boundary bands withedge-aware gating, and expanding sparsity, width, and branching througha controllable mask generator. On five crack and vessel benchmarks,SegFlow alone outperforms CNN, Transformer, Mamba, and generativebaselines, improving mean IoU from 0.511 to 0.599 (+17.2%) and reducingBetti matching error from 82.145 to 51.524 (−37.3%). With limitedlabels, SynFlow recovers near-full performance using 25% real annotationsand improves cross-domain IoU by 0.11. Unlike classical augmentation,SynFlow provides pixel-aligned supervision with controllable structuralshifts (e.g., sparsity, width, branching), making it effective under domainshift. We release 10k crack and 1k vessel image–mask pairs. The projectpage is available at FMS2 .

Localising multiple sound sources in visual scenes remains afundamental challenge in multimodal perception due to an inherent cir-cular dependency: separating mixed audio requires knowing source loca-tions, while identifying sound-producing regions requires separated audiosignals. In this paper, we focus on the dual-source setting and discover aselective convergence in self-supervised audio-visual learning: when pre-sented with multiple sound sources, contrastive models naturally con-verge to the most salient audio-visual correspondence rather than at-tempting to represent all sources equally. This emergent phenomenon,analogous to human selective auditory attention, enables us to breakthe above circular dependency through a progressive two-stage frame-work: first, leveraging selective convergence to identify dominant sources,and then exploiting these learned priors to uncover remaining sources.Our self-supervised approach achieves the best performance among self-supervised methods on dual-source benchmarks without requiring anymanual annotations, and even surpasses some weakly-supervised ap-proaches on certain metrics. Furthermore, we identify a fundamentalevaluation inconsistency in existing benchmarks: comparing continuouslocalisation heatmaps against bounding-box annotations creates system-atic biases, particularly for non-axis-aligned objects where the boundingbox includes substantial background regions. To address this, we intro-duce pixel-level segmentation masks to the existing benchmark, enablingspatially-aligned evaluation. Together, these results suggest that embrac-ing rather than suppressing selectivity offers a scalable, annotation-freeroute to multi-source localisation.