ECCV 2026 论文列表

Point-Based unsigned distance fields (UDFs) learning oftenaggregates Euclidean neighborhoods as unstructured point sets. Eachquery is typically supervised by a single distance value. This offers lit-tle relational guidance for separating nearby surfaces, reducing featurediscriminability for distance regression. We propose PrimitiveUDF, aprimitive-based UDF framework with two coupled components: projection-induced neighborhood reorganization and hybrid primitive construction.The former uses spherical projection and a sphere-to-plane mapping toobtain a stable local chart for assessing surface-relative relations. The lat-ter grows triangle primitives with continuity and non-crossing preferences,and prunes direction-consistent segments for thin or sparse structures.Primitives are then encoded and aggregated by an attention-based re-gressor for query-conditioned UDF prediction. Experimental results onsynthetic and real-world datasets demonstrate around 10% improvementon scene-level benchmarks and consistent cross-scene gains, with balancedrobustness under noise and sparsity conditions. Code: PrimitiveUDF.

Spatiotemporal prediction of complex dynamical systems de-pends not only on historical observations but also on evolving exter-nal control parameters. Crucially, for ionospheric Total Electron Con-tent, existing concatenation-based conditioning fails to capture region-dependent responses driven by space weather.We propose PMGC-SimVP, a spatiotemporal prediction framework withParametric Multi-scale Gated Convolution. PMGC injects external con-trol parameters into local dynamics via latitudinally structured mod-ulation, enabling explicit modeling of non-stationary evolution and re-gional heterogeneity. Built upon a purely convolutional backbone anda low-rank temporal adapter, PMGC-SimVP remains computationallye_x001E_cient.We evaluate PMGC-SimVP on global TEC forecasting, where dynam-ics are strongly driven by space weather and exhibit pronounced lati-tudinal variability. The proposed method achieves consistent improve-ments over recent state-of-the-art models, including Transformer-basedand frequency-domain approaches, across overall evaluation settings andduring geomagnetic disturbances. Experiments on generic video predic-tion benchmarks further verify the robustness of the proposed backboneeven without parameter injection.

Driven by the growing capacity and training scale, Text-to-Video (T2V) generation models have recently achieved substantialprogress in video quality, length, and instruction-following capability.However, whether these models can understand physics and generatephysically plausible videos remains a question. While Vision-LanguageModels (VLMs) have been widely used as general-purpose evaluators invarious applications, they struggle to identify the physically impossiblecontent from generated videos. To investigate this issue, we construct aPID (Physical Implausibility Detection) dataset, which consists of a testsplit of 500 manually annotated videos and a train split of 2,588 pairedvideos, where each implausible video is generated by carefully rewritingthe caption of its corresponding real-world video to induce T2V modelsproducing physically implausible content. With the constructed dataset,we introduce a lightweight fine-tuning approach, enabling VLMs to notonly detect physically implausible events but also generate textual expla-nations on the violated physical principles. Taking the fine-tuned VLMas a physical plausibility detector and explainer, namely PhyDetEx,we benchmark a series of state-of-the-art T2V models to assess their ad-herence to physical laws. Our findings show that although recent T2Vmodels have made notable progress toward generating physically plau-sible content, understanding and adhering to physical laws remains achallenging issue, especially for open-source models. Our dataset, train-ing code, and checkpoints are available at https://github.com/Zeqing-Wang/PhyDetEx.

Despite the remarkable capabilities of Large Vision-LanguageModels (LVLMs), they remain highly susceptible to hallucinations. Exist-ing inference-time mitigation strategies largely rely on visual attention asa diagnostic proxy. However, we demonstrate that this approach is insuf-ficient in certain cases due to the “anchoring trap,” where parameterizedlinguistic priors override correct visual grounding. In this paper, we pio-neer a representation geometry perspective, identifying hidden states asa complementary diagnostic signal beyond visual attention. We modelthe evolution of hidden states across layers and generation steps as con-strained sequences governed by a static visual subspace and a dynamiclinguistic subspace within a high-dimensional Euclidean space. We revealthat hallucinations inherently stem from abrupt structural deviations,specifically, the spontaneous departure of latent representations from thevisual subspace and their subsequent drift into the linguistic domain.To quantify this phenomenon, we propose the Geo-Semantic TurbulenceIndex (GSTI) to adaptively monitor visual misalignment and semanticvolatility. Building upon GSTI, we introduce Orthogonal Residual-BasedIntervention (ORBIT), an architecture-agnostic, plug-and-play mecha-nism that recalibrates divergent representation sequences by employinga closed-loop intervention encompassing residual extraction, visual re-anchoring, and norm-preserving injection of suppressed visual evidence.Extensive experiments across diverse LVLM architectures demonstratethat ORBIT significantly mitigates hallucinations and enhances generalvision-centric performance with exceptional computational efficiency.

Temporal Action Localization (TAL) typically relies on seg-ment annotations or offline access to full videos, limiting scalability andonline use. We introduce Point-Supervised Online TAL (POTAL), whichlocalizes actions in streaming videos using only one temporal point perinstance. To solve POTAL, we propose OnPoint, an offline-to-onlinemulti-level distillation framework that transfers knowledge from a point-supervised offline teacher to an online student via (i) pseudo-segmentinstance distillation, (ii) class-activation sequence distillation, and (iii)anticipatory window-level distillation. We further improve robustness byincorporating the original point labels into student training and by re-fining anchor decoding with actionness-guided attention calibration. Ex-periments on five datasets show OnPoint consistently outperforms strongbaselines, establishing a solid foundation for POTAL. †

Test-time scaling is a paradigm where large models use ad-ditional compute at inference to achieve better performance, withoutchanging model weights. While it has been widely studied for Large Lan-guage Models (LLMs), its applicability to Large Vision-Language Mod-els (LVLMs) remains less explored and analyzed, with limited analysis ofwhether, when, and to what extent these approaches transfer to LVLMs.In this work, we ask a simple but fundamental question: can conven-tional test-time scaling methods developed for LLMs be directly appliedto LVLMs? We present the first comprehensive study of test-time scal-ing for LVLMs, spanning multiple models and model sizes, nine test-timescaling methods, and six diverse benchmarks. Our main findings is that1) different from previous findings, small, well-performing models benefitthe most from test-time scaling, enabling performance improvements ofup to around 30%, reaching large models performance, and often out-performing them, 2) LVLMs lose focus when given more compute thannecessary, and 3) Visual information is encoded early in the reasoningchain, after which the chain is dominated by text-only reasoning andthe contribution of image tokens drops significantly. Finally, we also pro-vide a global and fine-grained analysis on the quality and informationsufficiency of the reasoning chains produced. Overall, our findings andanalysis provide practical guidance and insights into LVLMs and theirdeployment in research and industry.

Multimodal large language models (MLLMs) have achieved impressiveprogress in vision-language reasoning, yet their ability to understand temporallyunfolding narratives in videos remains largely underexplored. Narrative under-standing requires more than recognizing isolated events: models must maintaincoherent representations of who is doing what, when, and where across scene tran-sitions and temporal gaps. We introduce NARRATIVE T RACK, the first benchmarkto evaluate narrative understanding in MLLMs through fine-grained entity-centricreasoning. Unlike existing benchmarks limited to short clips or coarse scene-levelsemantics, we decompose videos into constituent entities and evaluate modelsusing a Compositional Reasoning Progression (CRP), a structured frameworkthat progressively increases narrative complexity across three dimensions: en-tity existence, entity changes, and entity ambiguity. This progression requiresmodels to move beyond local perception to reasoning about entities’ temporalpersistence, state changes, and fine-grained perceptual disambiguation. To enablescalable benchmark construction, we develop a fully automated entity-centricpipeline that extracts temporally grounded entity representations and providesthe foundation for CRP. Evaluations of state-of-the-art MLLMs reveal that ex-isting models struggle to maintain coherent entity representations under visualtransitions and temporal dynamics. Open-source general-purpose MLLMs exhibitstrong perceptual grounding but weak temporal continuity, while video-specializedMLLMs capture temporal context yet frequently hallucinate entities’ contexts.These findings uncover a fundamental trade-off between perceptual grounding andtemporal reasoning, indicating that narrative understanding emerges only fromtheir integration. NARRATIVE T RACK provides the first systematic framework todiagnose and advance temporally grounded narrative comprehension in MLLMs.

Zero-shot video moment retrieval (ZMR) aims to overcomethe limitations of traditional approaches that require large-scale datasetsannotated with text and its relevant temporal spans. Despite advances inpre-trained vision–language models (VLMs) and multimodal large lan-guage models (MLLMs), existing ZMR methods still heavily depend onquery-to-video content similarity, making them vulnerable to modal-ity and language-style gaps. These gaps lead to unreliable span pro-posals and unstable moment retrieval results. To address this issue, wepropose Self-Similarity-based Moment proposal and Scoring (Self-SiMS)that instead exploits intrinsic relationships within videos, enabling ro-bust span generation and scoring. By deriving self-similarity only fromthe video content, we circumvent the noisy and mismatched patternsof query–frame or query–caption similarities, thereby mitigating bothmodality and language-style gaps. Furthermore, we introduce a query-aware MLLM-based reasoning stage to further sharpen alignment be-tween text and video. Extensive experiments demonstrate that Self-SiMSachieves the state-of-the-art performance across ZMR benchmarks.

Unified multimodal models have demonstrated remarkableversatility in multimodal understanding and generation. However, theystruggle with image generation under complex compositional instructions,a setting frequently encountered in real-world applications. To addressthis, we present MANGO, a novel model that unleashes image generationcapability at both its performance floor and ceiling. (1) To raise theperformance floor, we introduce Functionality-Oriented Transformers,an architectural design that assigns each Transformer branch to a dedi-cated function, such as image generation or image understanding. Thisfunction-disentangled design alleviates cross-functional interference inher-ent in mainstream modality-oriented architectures, thereby unleashingthe model’s image generation capacity. (2) To elevate the performanceceiling, we propose Paint-CoT, a framework that enables image generationthrough a human-like artistic chain of thought, consisting of planning,acting, reflection, and correction. Leveraging the model’s multimodalunderstanding and reasoning capabilities, this reasoning-guided processfurther unleashes image generation capacity. Further, to overcome the dif-ficulty and high cost of collecting multi-step aligned chain-of-thought datafor end-to-end training, we develop a multi-task joint training paradigmthat decomposes the full reasoning-generation pipeline into manageablesubtasks. This strategy maximizes the utilization of existing or readilyconstructed supervision, enabling e!ective Paint-CoT-based image gener-ation without requiring fully aligned multi-step annotations. Extensiveexperiments demonstrate that MANGO consistently achieves superiorperformance across diverse image generation benchmarks, delivering sub-stantial improvements in complex image generation scenarios.

Hyperspectral-panchromatic fusion reconstructs high-resolutionhyperspectral (HRHS) images by combining low-resolution hyperspectral(LRHS) observations with high-resolution panchromatic (PAN) imagery.Although recent deep architectures achieve strong spectral–spatial re-construction performance, most models rely on stacked layers in whichglobal cross-modal interactions and local spatial refinement are updatedjointly at every stage. Consequently, the influence of PAN information isimplicitly coupled to network depth, limiting independent control overcross-modal interaction and reducing flexibility in balancing spectralpreservation and spatial enhancement. We propose the Leader-GuidedDynamics Network (LGD-Net), a framework that explicitly decouplesglobal cross-modal interaction from local multi-scale reconstruction. Aglobal leader representation first aggregates modality-specific features tosummarize global context. Instead of propagating interactions throughrepeated stacking, only the leader state evolves through a controlled,continuous-depth dynamic process, while local feature nodes focus onspatial reconstruction. The refined leader is subsequently injected backinto the multi-scale feature hierarchy via structured residual modula-tion, providing global guidance without increasing architectural depth.By separating global state evolution from local reconstruction, LGD-Net enables explicit control over cross-modal interactions while main-taining computational efficiency. Experiments on multiple hyperspectralbenchmarks demonstrate improved spectral–spatial reconstruction qual-ity compared to conventional stacked fusion architectures.

The recent success of inference-time scaling in large languagemodels has inspired similar explorations in video diffusion. In particular,motivated by the existence of “golden noise” that enhances video quality,prior work has attempted to improve inference by optimising or searchingfor better initial noise. However, these approaches have notable limita-tions: they either rely on priors imposed at the beginning of noise sam-pling or on rewards evaluated only on the denoised and decoded videos.This leads to error accumulation, delayed and sparse reward signals, andprohibitive computational cost, which prevents the use of stronger searchalgorithms. Crucially, stronger search algorithms are precisely what couldunlock substantial gains in controllability, sample efficiency and genera-tion quality for video diffusion, provided their computational cost can bereduced. To fill in this gap, we enable efficient inference-time scaling forvideo diffusion through latent reward guidance, which provides interme-diate, informative and efficient feedback along the denoising trajectory.We introduce a latent reward model that scores partially denoised la-tents at arbitrary timesteps with respect to visual quality, motion qual-ity, and text alignment. Building on this model, we propose LatSearch,a novel inference-time search mechanism that performs Reward-GuidedResampling and Pruning (RGRP). In the resampling stage, candidatesare sampled according to reward-normalised probabilities to reduce over-reliance on the reward model. In the pruning stage, applied at the finalscheduled step, only the candidate with the highest cumulative reward isretained, improving both quality and efficiency. We evaluate LatSearchon the VBench-2.0 benchmark and demonstrate that it consistently im-proves video generation across multiple evaluation dimensions comparedto the baseline Wan2.1 model. Compared with the state-of-the-art, ourapproach achieves comparable or better quality while reducing runtimeby up to 79%.

Hybrid event–frame sensors integrate an Event Vision Sensor(EVS) and an Active Pixel Sensor (APS) within a single chip, combin-ing the high dynamic range and low latency of the EVS with the richspatial intensity information from the APS. While this tight integrationoffers compact, temporally precise imaging, the complex circuit architec-ture introduces non-trivial noise patterns that remain poorly understoodand unmodeled. In this work, we present the first unified, statistics-basedimaging noise model that jointly describes the noise behavior of APS andEVS pixels. Our formulation explicitly incorporates photon shot noise,dark current noise, fixed-pattern noise, and quantization noise, and linksEVS noise to illumination level and dark current. Based on this formula-tion, we further develop a calibration pipeline to estimate noise param-eters from real data and offer a detailed analysis of both APS and EVSnoise behaviors. Finally, we propose H-ESIM, a statistically groundedsimulator that generates RAW frames and events under realistic, jointlycalibrated noise statistics. Experiments on two hybrid sensors validateour model across multiple imaging tasks (e.g., video frame interpolationand deblurring), demonstrating strong transfer from simulation to realdata. https://yunfanlu.github.io/HESIM

The e_x001E_cient spatial allocation of primitives serves as thefoundation of 3D Gaussian Splatting, as it directly dictates the synergybetween representation compactness, reconstruction speed, and render-ing _x001C_delity. Previous solutions, whether based on iterative optimizationor feed-forward inference, su_x001B_er from signi_x001C_cant trade-o_x001B_s between thesegoals, mainly due to the reliance on local, heuristic-driven allocationstrategies that lack global scene awareness. Speci_x001C_cally, current feed-forward methods are largely pixel-aligned or primitive-aligned. By un-projecting pixels into dense, view-aligned primitives, they bake redun-dancy into the 3D asset. As more input views are added, the represen-tation size increases and global consistency becomes fragile. To this end,we introduce GlobalSplat, a framework built on the principle of align_x001C_rst, decode later. Our approach learns a compact, global, latent scenerepresentation that encodes multi-view input and resolves cross-view cor-respondences before decoding any explicit 3D geometry. Crucially, thisformulation enables compact, globally consistent reconstructions withoutrelying on pretrained pixel-prediction backbones or reusing latent fea-tures from dense baselines. Utilizing a coarse-to-_x001C_ne training curriculumthat gradually increases decoded capacity, GlobalSplat natively preventsrepresentation bloat. On RealEstate10K and ACID, our model achievescompetitive novel-view synthesis performance while utilizing as few as16K Gaussians, signi_x001C_cantly less than required by dense pipelines, ob-taining a light 4MB footprint. Further, GlobalSplat enables signi_x001C_cantlyfaster inference than the baselines, operating under 78 milliseconds in asingle forward pass.

Systematic failures of vision models on semantically coherentsubsets, known as error slices, reveal limitations in robustness and evalu-ation. Existing slice discovery approaches largely model slices as clustersin representation space or combinations of predefined attributes. Whileeffective for image-level classification, such formulations are insufficientfor instance-level tasks such as object detection and segmentation, wherefailures often arise from contextual, relational, and spatially groundedvisual patterns. We propose GH-ESD (Grounded Hypothesis-Driven ErrorSlice Discovery), a generate-and-verify framework that reformulates slicediscovery as grounded hypothesis generation and statistical verification.GH-ESD constructs relational failure hypotheses using LLM priors andgrounded visual evidence, discovers hypothesis slices at the instance levelvia Vision-Language Models, and verifies them through statistical trendanalysis over instance-level errors. We also introduce GESD (GroundedError Slice Dataset), a new benchmark for instance-level error slice discov-ery, providing expert-defined and spatially grounded slices derived fromdetection and segmentation failures. Extensive experiments demonstratethat GH-ESD consistently outperforms baselines, improving Precision@10by 0.10 (0.73 vs. 0.63) on the GESD benchmark for detection tasks, whilealso supporting segmentation scenarios. GH-ESD identifies interpretableslices that facilitate actionable model improvements.

We present GenLCA, a diffusion-based generative modelfor generating and editing photorealistic full-body avatars from text andimage inputs. The generated avatars are faithful to the inputs, whilesupporting high-fidelity facial and full-body animations. The core ideais a novel paradigm that enables training a full-body 3D diffusion modelfrom partially observable 2D data, allowing the training dataset to scaleto millions of real-world videos. This scalability contributes to the supe-rior photorealism and generalizability of GenLCA. Specifically, we scaleup the dataset by repurposing a pretrained feed-forward avatar recon-struction model as an animatable 3D tokenizer, which encodes unstruc-tured video frames into structured 3D tokens. However, most real-worldvideos only provide partial observations of body parts, resulting in ex-cessive blurring or transparency artifacts in the 3D tokens. To addressthis, we propose a novel visibility-aware diffusion training strategy thatreplaces invalid regions with learnable tokens and computes losses onlyover valid regions. We then train a flow-based diffusion model on the to-ken dataset, inherently maintaining the photorealism and animatabilityprovided by the pretrained avatar reconstruction model. Our approacheffectively enables the use of large-scale real-world video data to traina diffusion model natively in 3D. We demonstrate the efficacy of ourmethod through diverse and high-fidelity generation and editing results,outperforming existing solutions by a large margin.

Despite decades of work, surveillance still struggles in searching and reasoning about specific targets across long, multi-camera videos. Existing methods—tracking, retrieval, and video LLMs—require heavy manual filtering, capture only shallow attributes, and fail at temporal understanding. Prior benchmarks are also limited to basic retrieval and question answering, without addressing real-world challenges that often involve multimodal queries and temporal grounding (e.g., “When did this person join the fight?” with the person’s image). To address this gap, we introduce ForeSeaQA, a new benchmark specifically designed for video QA with image-and-text queries and timestamped annotations of key events. The dataset consists of long-horizon surveillance footage paired with diverse multimodal questions, enabling systematic evaluation of retrieval, temporal grounding, and multimodal reasoning in realistic forensic conditions. Not limited to this benchmark, we propose ForeSea, an AI forensic search system with a 3-stage, plug-and-play pipeline. (1) A tracking module filters irrelevant footage; (2) a multimodal embedding module indexes the remaining clips; and (3) during inference, the system retrieves top-K candidate clips for a video LLM to answer queries and localize events. On ForeSeaQA benchmark, ForeSea improves accuracy by 3.1 points and temporal IoU by 10.1 points over prior retrieval-augmented baselines. To our knowledge, ForeSeaQA is the first benchmark to support complex multimodal queries with precise temporal grounding, and ForeSea is the first VideoRAG system built to excel in this setting.

Zero-shot composed image retrieval (ZS-CIR) aims to re-trieve a target image by editing a reference image with a natural-languageinstruction, without relying on domain-specific annotated triplets. Mostexisting ZS-CIR methods rely on textual inversion to translate the ref-erence image into pseudo-text tokens and then compose them with theinstruction via simple concatenation in the text space, which can belossy and brittle for fine-grained semantics. In this work, we propose anew paradigm, namely FlowCIR, that casts ZS-CIR as conditionalsemantic transport between reference and target embeddings. Leverag-ing conditional flow matching, our model learns a lightweight transportfield that maps the instruction representation toward a target-alignedquery embedding conditioned on the reference image. Since FlowCIRoperates on pre-extracted VLM embeddings and trains only a smalltransport module without updating the image or text encoder, it of-fers a computationally efficient training protocol compared with priortextual-inversion-based approaches. We further identify negation and re-moval as a major failure mode of VLM-based composition. To addressthis, we propose an inference-only Multi-Negative Steering strategy thatsteers a negation-containing relative instruction away from its negatedsemantics, mitigating the limited negation handling of VLMs and im-proving robustness on negation-heavy queries. Extensive experiments onstandard CIR benchmarks demonstrate that FlowCIR achieves strongand competitive performance compared with recent ZS-CIR methods.Project page: https://hkust-longgroup.github.io/FlowCIR

Recent advancements in Multimodal Large Language Models(MLLMs) have demonstrated impressive fine-grained perception capabil-ities. However, existing benchmarks predominantly rely on explicit tex-tual cues or low-resolution inputs, failing to evaluate a model’s ability toautonomously perceive implicit visual cues in high-resolution. To bridgethis gap, we introduce DiCoBench, a comprehensive, multi-image high-resolution benchmark designed for cross-image fine-grained perception.DiCoBench consists of 765 meticulously curated samples categorizedinto two progressive tracks: Differential Visual Cues and Commonal-ity Visual Cues, covering 8 distinct perception tasks. By formulatingthe benchmark as a multiple-choice question task and utilizing high-resolution imagery (approaching 2K), we eliminate evaluation metricbias and pose a substantial challenge to current state-of-the-art MLLMs.Our extensive evaluation of 18 diverse MLLMs reveals a striking perfor-mance gap compared to human accuracy (98.3%), with top-performingmodels struggling significantly with micro-scale detail capture. We be-lieve DiCoBench will serve as a challenging testbed to drive future re-search in autonomous, high-resolution multi-image perception. Dataset:https://huggingface.co/datasets/oking0197/DiCoBench

Visible-to-infrared translation provides a scalable pathway tosynthesize infrared imagery from abundant visible data, offering a prac-tical complement to costly cross-modal paired data acquisition. However,most existing generative approaches formulate the task as a unified pro-cess that implicitly entangles macroscopic thermal semantics with mi-croscopic spatial detail reconstruction. This entanglement often forces asub-optimal trade-off, leading to blurred target boundaries or physicallyunnatural thermal artifacts. In this paper, we reformulate the transla-tion process into two explicitly decoupled but tightly coordinated con-tinuous flows, proposing a novel framework named D3F-IR. Leverag-ing the smooth and deterministic trajectories of flow matching, D3F-IRintroduces a semantic-domain flow to model macroscopic thermal dis-tributions within a pretrained latent space, alongside an independentpixel-domain flow to generate fine-grained spatial structures directly inthe image space. To seamlessly coordinate the two domains, we design aPixel-space Velocity Predictor equipped with SEmantic-Aligned Modula-tion (SEAM) Layers, which establish explicit correspondence between se-mantic tokens and spatial patches through patch-wise alignment. Exper-iments on three benchmarks demonstrate that D3F-IR achieves strongperformance in both thermal semantic plausibility and pixel-level spa-tial fidelity. These results highlight that explicit dual-domain decouplingoffers a highly effective paradigm for high-quality visible-to-infrared gen-eration. Code is available at https://github.com/WanrenZeng/D3F-IR.