ECCV 2026 Papers

005 We present Lunima-OmniLV (abbreviated as OmniLV), 005006 a universal multimodal multi-task framework for low-level vision that ad- 006007 dresses over 100 sub-tasks across four major categories, including image 007008 restoration, image enhancement, weak-semantic dense prediction, and 008009 stylization. OmniLV leverages both textual and visual prompts to offer 009010 flexible, user-friendly interactions. Built on Diffusion Transformer (DiT)- 010011 based generative priors, our framework supports arbitrary resolutions 011012 — achieving optimal performance at 1K resolution — while preserving 012013 fine-grained details and high fidelity. Through extensive experiments, 013014 we demonstrate that separately encoding text and visual instructions, 014015 combined with co-training using shallow feature control, is essential to 015016 mitigate task ambiguity and enhance multi-task generalization. Our find- 016017 ings also reveal that integrating high-level generative tasks into low-level 017018 vision models can compromise detail-sensitive restoration. These insights 018019 pave the way for more robust and generalizable low-level vision systems. 019

Real-world deployment of text-to-image diffusion models re-quires continual concept removal as new privacy, copyright, or safetyobligations arise over time. Existing unlearning methods, however, aredesigned for single-step deletion and collapse after only 3–5 sequentialapplications. We trace this instability to two compounding factors: (i)coarse mapping targets that cause degradation to accumulate unnec-essarily across steps, and (ii) the absence of local protection for se-mantically neighboring concepts, whose shared internal representationsmake them the first to suffer collateral damage. Because this damageis strongest in the local semantic neighborhood of the forget concept,global replay alone cannot prevent it. Building on this analysis, we pro-pose Locality-Aware Continual Unlearning (LACU), a frameworkwith two complementary mechanisms. Locality-Aware Target Selec-tion chooses, for each forget prompt, the context-preserving mappingprompt that the diffusion model itself treats as most similar to the orig-inal prompt, measured by score-prediction distance (how differently themodel denoises the same noisy image under two text conditions), ensur-ing each update is as small and targeted as possible. Locality-AwareReplay uses the same metric to identify the retain concepts closest to theforget concept in the model’s own representation and replays them as alocal functional regularizer, directly shielding the most vulnerable neigh-borhood. Combined with teacher-student distillation and lightweight ℓ2parameter regularization, LACU maintains stable unlearning over 10 se-quential steps, preserving significantly higher related retention (RRacc )and general retention (GRacc ) than recent baselines. The code is availableat https://github.com/SonyResearch/LACU.

Spatial understanding is a fundamental cornerstone of human-level intelligence. Nonetheless, current research predominantly focuses ondomain-specific data production, leaving a critical void: the absence of aprincipled, open-source engine capable of fully unleashing the potentialof high-quality spatial data. To bridge this gap, we elucidate the designprinciples of a robust data generation system and introduce OpenSpa-tial—an open-source data engine engineered for high quality, extensivescalability, broad task diversity, and optimized efficiency. OpenSpatialadopts 3D bounding boxes as the fundamental primitive to constructa comprehensive data hierarchy across five foundational tasks: SpatialMeasurement (SM), Spatial Relationship (SR), Camera Perception (CP),Multi-view Consistency (MC), and Scene-Aware Reasoning (SAR). Lever-aging this scalable infrastructure, we curate OpenSpatial-3M, a large-scale dataset comprising 3 million high-fidelity samples. Extensive eval-uations demonstrate that versatile models trained on our dataset achievestate-of-the-art performance across a wide spectrum of spatial reasoningbenchmarks. Notably, the best-performing model exhibits a substantialaverage improvement of 19%, relatively. Furthermore, we provide a sys-tematic analysis of how data attributes influence spatial perception. Byopen-sourcing both the engine and the 3M-scale dataset, we provide arobust foundation to accelerate future research in spatial intelligence.Code: github.com/VINHYU/OpenSpatialDataset: huggingface.co/datasets/JoyAI-Image-OpenSpatial

Decomposing CAD models into sweepable sub-blocks is acritical preprocessing step for high-quality hexahedral mesh generation.Existing block decomposition methods fall into two main categories: tra-ditional geometry-driven approaches and reinforcement learning (RL)-based methods. Traditional methods rely heavily on manual expertiseand interactive operations, resulting in intensive manual labor and ex-tremely low automation efficiency. Existing RL-based methods are ei-ther only applicable to simple 2D models, or dependent on large-scaleoffline datasets and pre-training, leading to high training costs and diffi-cult industrial deployment. To address these challenges, we propose thefirst pure online RL framework for 3D CAD model block decomposition.Specifically, we formulate the task as a tree-structured Markov Deci-sion Process (MDP), and adopt two core designs: a decoupled frozengeometric-topological feature extraction architecture for lightweight RLnetwork design, and a continuous-to-discrete action projection mecha-nism to ensure geometric validity of decomposition operations. Withoutrelying on any pre-training or offline datasets, our agent learns optimalcutting strategies purely through online interaction with a single targetCAD model. Experiments on our self-built dataset and public datasetsshow that our method outperforms the state-of-the-art offline RL base-line, and achieves competitive decomposition quality comparable to thehybrid fine-tuning paradigm with no pre-training dependency.

LiDAR has increasingly been integrated into traffic cam-eras to expand coverage and mitigate occlusion in roadside cooperativeperception. However, how unimodal and camera–LiDAR fusion archi-tectures behave under variations in LiDAR point sparsity induced bysensor configurations and scene-dependent sensing conditions remainsunderexplored. We introduce RESOLVE, a large-scale real-world bench-mark dataset featuring multi-resolution roadside LiDAR and synchro-nized camera-LiDAR sensing for systematic evaluation of unimodal andfusion-based architectures in roadside 3D detection and tracking. RE-SOLVE contains over 100k images and 26k point cloud frames with 220kmanually annotated bounding boxes, captured at a real-world urban in-tersection across diverse lighting and weather conditions and spanning10 classes of traffic participants. In particular, RESOLVE enables con-trolled evaluation across three LiDAR resolution levels while keeping allother sensing and environmental factors fixed. This allows fair cross-architecture comparisons under point cloud distribution shifts resultingfrom resolution variations, sensing distance, and training–inference res-olution mismatches. Results from extensive benchmark experiments re-veal insights into how multimodal fusion can compensate for LiDARpoint sparsity, offering clues for designing cost-efficient roadside multi-modal perception. The dataset and benchmark codes are available athttps://github.com/ASU-Suo-Lab/RESOLVE.

How to accurately predict a high-fidelity future world? Whilethe visual world is inherently continuous, existing deterministic videoprediction models operate in discrete pixel space and are mainly opti-mized with pixel-wise mean squared error (MSE), which often leads toover-smoothed predictions and a lack of fine-grained visual details. Toaddress these limitations, we propose Predictive Differentiable Rendering(PDR), a novel end-to-end video prediction paradigm that bridges thegap between discrete and continuous representations. Inspired by recentprogress in 3D reconstruction with 3D Gaussian Splatting, we introducePredGS, a lightweight and plug-and-play adapter based on 2D Gaus-sian representation, which could be seamlessly integrated with existingpixel space predictors, significantly improving spatial detail preserva-tion with negligible computational overhead. Furthermore, we developpredgsplat, a CUDA-accelerated differentiable 2D Gaussian renderersupporting arbitrary channels. Each Gaussian is defined by 5 + C learn-able parameters (position, scale, rotation, and C channel amplitudes)and achieves up to 10× faster rendering than the baseline. Optimizedby a combined L1 and SSIM loss, PDR overcomes the inherent blurringtendencies of MSE Loss, significantly enhancing the prediction perfor-mance. Extensive experiments on diverse real-world benchmarks, includ-ing TaxiBJ, WeatherBench, KTH, and Human3.6M, demonstrate thatPDR consistently surpasses existing methods, delivering superior detailpreservation, visual fidelity, and predictive accuracy.

Single-Photon Avalanche Diode (SPAD)-based Light Detec-tion and Ranging (LiDAR) is emerging for autonomous vehicles due toits high sensitivity and precise depth sensing capabilities. However, flarecaused by excessive photon returns or pile-up effects can lead to incorrectdepth estimation and exaggerated boundaries in point clouds, resultingin severe distortions of geometric measurements, making flare suppres-sion essential for safety-critical applications. Existing flare mitigationmethods primarily operate at the hardware or signal-processing levels.While effective under specific configurations, they are largely rule-basedand configuration-dependent, lacking learnable representations that gen-eralize across diverse sensing scenarios. In this work, we reformulate flaresuppression as a semantic segmentation problem, enabling data-drivenlearning of geometric and photometric cues directly from SPAD mea-surements. We first benchmark representative segmentation models onthe newly introduced SPAD flare dataset and observe that they strug-gle to exploit the intrinsic multi-echo characteristics of SPAD signals.Motivated by this observation, we propose Physically-Informed seg-mentation for LiDAR Flare (PILF), a learning-based approach thattreats the first and second echoes, together with ambient illumination,as distinct modalities, aggregating cross-echo information while jointlyencoding geometric and photometric features. Experiments across mul-tiple real-world scenes demonstrate that PILF significantly outperformscompared segmentation models, achieving up to 79.32% mIoU, andproviding an effective solution for SPAD-based LiDAR flare suppression.

Single-photon LiDAR (SPL) based on single-photon avalanche diode (SPAD) sensing enables time-resolved photon measurements with extreme sensitivity, offering unique potential for active 3D perception. However, real-world single photon perception remains fundamentally challenging due to unique measurement noise and complex multi-return transient phenomena, which jointly complicate geometric reconstruction and semantic scene understanding. Despite growing interest in SPAD-based sensing, existing studies are largely limited to simulated data or smallscale controlled captures. As a result, systematic evaluation of real-world single photon perception across depth estimation, multi-view reconstruction, and 3D semantic understanding remains underexplored. To bridge this gap, we introduce SP-TransientBench (STB), a real-captured multitask benchmark for single photon perception. STB provides 256 × 192 transient data for three tasks: 10 depth estimation samples, 9 reconstruction scenes, and 27 semantic sequences with 10,297 samples. Each view provides full time-of-flight histograms with multi-return behavior, standardized metadata, and calibrated camera poses for multi-view evaluation. We further provide 13-class 3D semantic annotations and histogram-domain multi-return annotations for selected scenes, enabling the study of raw transient cues. By providing dedicated data splits and evaluation protocols for each task, STB enables consistent and reproducible benchmarking of real-world single photon perception across multiple 3D vision problems. The dataset and code are available at: https: //huggingface.co/datasets/shuinb/SP-TransientBench

Visual Place Recognition (VPR) often fails under extremeenvironmental changes and perceptual aliasing. Beyond these limita-tions, standard systems cannot perform ’blind’ localization from verbaldescriptions alone, a capability critical for applications such as emer-gency response. To address these challenges, we introduce LaVPR, alarge-scale benchmark that extends existing VPR datasets with over650,000 rich natural-language descriptions. Using LaVPR, we investigatetwo paradigms: Multi-Modal Fusion for enhanced robustness and Cross-Modal Retrieval for language-based localization. Our results show thatlanguage descriptions yield consistent gains in visually degraded condi-tions, with the most significant impact on smaller backbones. Notably,adding language allows compact models to rival the performance of muchlarger vision-only architectures. For cross-modal retrieval, we establish abaseline using Low-Rank Adaptation (LoRA) and Multi-Similarity loss,which substantially outperforms standard contrastive methods acrossvision-language models. Ultimately, LaVPR enables a new class of lo-calization systems that are both resilient to real-world stochasticity andpractical for resource-constrained deployment. Our dataset and code areavailable at https://github.com/oferidan1/LaVPR.

The success of modern text-to-image generation is largely at-tributed to massive, high-quality datasets. Currently, these datasets arecurated through a filter-first paradigm that aggressively discards low-quality raw data based on the assumption that it is detrimental to modelperformance. Is the discarded bad data truly useless, or does it hold un-tapped potential? In this work, we critically re-examine this question. Wepropose LACON (Labeling-and-Conditioning), a novel training frame-work that exploits the underlying uncurated data distribution. Insteadof filtering, LACON re-purposes quality signals, such as aesthetic scoresand watermark probabilities as explicit, quantitative condition labels.The generative model is then trained to learn the full spectrum of dataquality, from bad to good. By learning the explicit boundary betweenhigh- and low-quality content, LACON achieves superior generation qual-ity compared to baselines trained only on filtered data using the samecompute budget, proving the significant value of uncurated data. Ourproject page is available at https://zhiyangliang.github.io/LACON.

This paper focuses on a challenging setting of simultaneouslymodeling geometry and appearance of hand-object interaction sceneswithout any object priors. We follow the trend of dynamic 3D GaussianSplatting based methods, and address several significant challenges. Tomodel complex hand-object interaction with mutual occlusion and edgeblur, we present interaction-aware hand-object Gaussians with newlyintroduced optimizable parameters aiming to adopt piecewise linear hy-pothesis for clearer structural representation. Moreover, considering thecomplementarity and tightness of hand shape and object shape duringinteraction dynamics, we incorporate hand information into object de-formation field, constructing interaction-aware dynamic fields to modelflexible motions. To further address difficulties in the optimization pro-cess, we propose a progressive strategy that handles dynamic regions andstatic background step by step. Correspondingly, explicit regularizationsare designed to stabilize the hand-object representations for smooth mo-tion transition, physical interaction reality, and coherent lighting. Exper-iments show that our approach surpasses existing dynamic 3D-GS-basedmethods and achieves state-of-the-art performance in reconstructing dy-namic hand-object interaction.

Incremental scene reconstruction is essential for real-worldapplications. Although 3D Gaussian Splatting shows strong potential,most existing approaches require offline conversion of the optimized Gaus-sians into an intermediate implicit field for explicit mesh extraction,which hinders seamless integration with downstream tasks. To addressthis limitation, we propose a novel online framework that incrementallyreconstructs and updates high-fidelity explicit meshes by directly trian-gulating a dense geometric Gaussian representation, which supports bothhigh-quality rendering and incremental surface reconstruction. More-over, we present a direct meshing algorithm that efficiently extractsand updates the mesh from the Gaussian set. To ensure mesh accu-racy, we enforce a plane-based pulling constraint that dynamically aligns3D Gaussian primitives to the approximated local surface. Furthermore,our framework significantly reduces memory and computational overheadduring long-sequence processing by dynamically freezing fully optimizedhistorical regions. Experiments on public datasets demonstrate that ourmethod outperforms conventional Gaussian-based methods on both ren-dering quality and reconstruction accuracy.

Quantifying plant growth dynamics from sparse longitudinal3D observations is fundamental for agriculture and plant sciences. Yet,plants pose unique challenges: they undergo intricate non-rigid deforma-tions, exhibit changing topology as new organs emerge, and often lackexplicit temporal correspondences between consecutive data acquisitionsdue to newly formed tissue. Methods designed for general scenes struggleto model topology changes and asynchronous organ growth characteris-tic of plants. To address these challenges, we introduce GrowFields, acompositional dynamic neural field representation for organ-aware 4Dplant growth modelling from point cloud time series. Our approach de-composes a plant into its constituent organs and aligns each organ intoits own canonical coordinate frame, isolating intrinsic growth patternsfrom global plant motion. We then learn a shared continuous neural de-formation field that models temporal dynamics across all organs, condi-tioned on learnable per-organ latent codes capturing organ identity andgrowth characteristics. The resulting modular yet unified representationnaturally accommodates the asynchronous development of plant organswhile remaining grounded in the practical setting of organ-level planttracking. We evaluate GrowFields on growth sequences from four plantspecies, assessing geometric fitting and organ tracking accuracy usingmanually annotated leaf-tip trajectories. Results demonstrate consistentimprovements in spatial precision, temporal coherence, and morphologi-cal fidelity over a range of existing representations.

With the increasing development of Vision-Language Mod-els, it becomes imperative that their predictions are readily explainableto relevant stakeholders. However, the field of explainability has notkept pace with the multimodal surge. While recent Multimodal Explain-able AI (MxAI) methods generate explanations to attribute the interac-tion between different modalities, current evaluation protocols lack theground truth required to distinguish between true cross-modal reason-ing (e.g. spatial composition) and shallow cross-modal shortcuts (e.g.Bag-of-Words attribute matching). It remains unknown whether MxAImethods faithfully capture synergistic interactions or merely hallucinatereasoning on models acting as simple feature detectors. In this paper,we introduce GridVQA-X, the first diagnostic framework specificallydesigned to evaluate cross-modal explainability. Unlike natural datasets,GridVQA-X leverages a closed-world synthesis logic to generate unique,mathematically guaranteed explanations. We utilize this controlled envi-ronment to train paired ground-truth models on identical architectures:Mpure , which learns robust spatial-relational reasoning and Mspur , whichis structurally forced to rely on cross-modal shortcuts. This behavioraldivergence creates a rigorous testbed: a faithful explainer must reportdistinct reasoning pathways for each model.Our findings reveal that widely used methods fail to distinguish betweenmodels relying on genuine spatial-relational reasoning and those exploit-ing cross-modal shortcuts, highlighting a critical gap in capturing truecross-modal synergy and misrepresenting how multimodal models actu-ally make decisions. Our code is available at GitHub, and the dataset andmodels are available on Hugging Face at gridvqa-dataset and gridvqa-models, respectively.

Visible-infrared person re-identification aims to match pedes-trian images of the same identity captured by visible and infrared cam-eras, and is hindered by severe domain gaps due to the absence of colorin the infrared modality. Some methods employ text to supplement colorcues in infrared modality. However, they rely on global textual guidance,which hinders fine-grained alignment between textual color informationand local infrared semantics. To this end, this paper proposes a GranularSemantic Cognition (GSC) method, which leverages cross-modal sharedsemantics to facilitate the transfer of color features. Specifically, a Hierar-chical Refinement Module (HRM) progressively refines shared semanticprototypes via multi-stage interactions with pixel-wise features, yieldingfine-grained semantics. Based on these refined semantics, a Semantics-guided Adjustment Module (SAM) generates semantic-conditioned fu-sion weights to selectively aggregate textual color information in se-mantically relevant regions. In addition, a fusion-weighted identity con-trastive loss is proposed to alleviate semantic ambiguity across identi-ties by regularizing the semantics-guided fusion process. Extensive ex-perimental results conducted on the SYSU-MM01, RegDB and LLCMdatasets demonstrate the favorable performance.

4D occupancy forecasting models the spatio-temporal evo-lution of 3D scenes and is crucial for autonomous driving, especially forcorner-case simulation. Existing methods often rely on discrete tokeniza-tion followed by autoregressive prediction, yet struggle with geometricdistortion in static structures and inconsistent temporal coherence overthe forecasting horizon. In this work, we propose a Geometry-AwareSpatio-Temporal context modeling method (GAST) for 4D occupancyforecasting, built upon progressive explicit-implicit generation and dual-path spatio-temporal modeling. Specifically, the generation module pro-duces per-frame occupancy with high geometric fidelity and semanticplausibility through pose-driven warping, motion-aware feature modula-tion, and attention-based feature refinement. Subsequently, the spatio-temporal module enhances spatial consistency through global contextaggregation while capturing scene evolution through temporal dynam-ics extraction. This unified design enables joint optimization of histori-cal reconstruction and future forecasting in an end-to-end manner. Ex-tensive experiments on Occ3D-nuScenes demonstrate the superiority ofour method, outperforming the state-of-the-art by 7.67% in mIoU and6.44% in IoU with a 2.84× speedup, while maintaining strong perfor-mance in long-term forecasting. Our source code is publicly available athttps://github.com/chenst27/GAST.

Generative models like Di_x001B_usion Models and Flow Matchinghave demonstrated remarkable capabilities in synthesizing high-_x001C_delitydriving videos, but are severely constrained by high inference latencydue to the requirement of extensive sampling steps. We argue that thisine_x001E_ciency stems from the prevailing reliance on a standard Gaussiansource distribution, where consecutive frames are initialized as indepen-dent Gaussian noise. This paradigm disregards the rich spatiotemporalcorrelations inherent in driving videos, compelling the model to regener-ate deterministic scene structures existing in previous frames from noise,which is both computationally redundant and prone to geometric incon-sistency. To address this problem, we propose GeoFlow, a novel frame-work designed to achieve e_x001E_cient driving video generation by harnessingexplicit geometric priors. Instead of sampling from standard Gaussiannoise, we leverage multi-view geometry and spatially-adaptive noise in-jection to construct a Geometry-Aligned Prior (GAP) distribution asstarting point. This initialization bridges the gap between source dis-tribution and data distribution, yielding a signi_x001C_cantly straighter andshorter sampling trajectory. Extensive experiments demonstrate that Ge-oFlow can achieve remarkable e_x001E_ciency of both training and inference:merely several hours of _x001C_ne-tuning on baseline models can signi_x001C_cantlyboost few-step generation quality, while fully converged training drasti-cally reduces number of inference steps required for state-of-the-art videogeneration.

Lane topology reasoning aims to construct a lane graph fromonboard sensor observations. Existing methods follow a detection and as-sociation paradigm that treats each lane instance independently, leadingto geometric inconsistency at connected endpoints and incomplete graphsdue to visual occlusions. To address these issues, we propose TopoGPT,a generative framework that learns the geometry prior from typical lanegraph structures through autoregressive sequence modeling. Specifically,we construct a large-scale map dataset comprising 3.3M scenes. For eachlane graph, a lane tokenizer serializes it into discrete tokens, while a scenecontext encoder converts it into a rasterized image and extracts globalfeatures as scene tokens. We pre-train an autoregressive lane sequencetransformer via scene-conditioned next-token prediction, endowing themodel with the geometry prior over lane graph structures. Building uponthis prior, a perception adapter aligns BEV features from multi-view im-ages with the pre-trained scene condition, transferring the learned geom-etry prior to sensor-based lane graph prediction. On the OpenLane-V2benchmark, TopoGPT outperforms existing methods by an average of+6.4 on lane-level and +11.6 on point-level metrics, and produces geo-metrically consistent and structurally complete lane graphs. Our projectpage is available at https://buaa-colalab.github.io/topogpt_page.

We introduce a novel learning problem: decoding gaze intonatural language descriptions of human goals across diverse visual tasks.Unlike prior work, which frames gaze decoding as a discriminative taskover predefined categories, we formulate it as a generative learning prob-lem: training a model to produce free-form descriptions that capture therich nuances and open-ended nature of human intentions beyond fixedlabels. To this end, we introduce Gazette, the first gaze-to-text decod-ing framework. Based on multimodal large language models (MLLMs),Gazette learns to decode gaze scanpaths into natural language for goalsthat may extend beyond categorical labels and require articulation innatural language. To help Gazette filter out individual differences in gazebehavior and learn the goal-specific spatiotemporal dynamics crucial forgenerating accurate natural language goal descriptions, we propose anovel strategy that leverages the encyclopedic knowledge and reason-ing abilities of a large language model to synthesize natural languageexplanations of goal-directed attentional behavior called think-aloud tran-scripts. Instruction tuning on these synthetic narratives allows Gazetteto achieve state-of-the-art performance in gaze decoding across mul-tiple tasks, demonstrating its generalizability and versatility, therebyenabling gaze to serve as a powerful, non-intrusive cue for inferringhuman goals and intentions in diverse scenarios. Code is available athttps://github.com/cvlab-stonybrook/Gazette.