ECCV 2026 Papers

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.

Building damage assessment is crucial for post-disaster re-lief and reconstruction. However, current assessment methods utilizinghigh-resolution remote sensing imagery struggle to represent _x001C_ne-graineddetails, as subtle damage signals are easily overwhelmed by complexbackgrounds. To address this, we propose a Frequency-Adaptive FeatureInjection and Change Decoupling framework. For feature encoding, we_x001C_ne-tune a DINOv3 encoder with a frequency-adaptive guidance strat-egy to acutely capture subtle damage. For decoding, we introduce theChange Representation Decoupling and Discriminative Decoder (CRD2 ).CRD2 leverages the Mamba model to e_x001E_ciently fuse multi-scale fea-tures for spatiotemporal dynamics. To prevent minor damage from be-ing submerged, it utilizes orthogonal subspace projection to isolate purechange features and applies class-prototype contrastive learning for pre-cise damage-level identi_x001C_cation. Extensive experiments show our CRD-Net achieves state-of-the-art Joint F1 scores on the xBD (81.26%) andEBD (81.88%) datasets.

Correspondence learning aims to estimate accurate geomet-ric model parameters from the data with severe outliers and is crucial formany computer vision tasks. While recent deep correspondence learningmethods have substantially improved inlier-outlier identification, most ofthem still rely on a differentiable weighted least squares (WLS) estima-tor to recover the final model parameters. This strategy typically treatscorrespondences as conditionally independent observations and there-fore ignores the structured dependencies among inliers. In real-worldscenarios, inliers are often sampled from geometric manifolds and tendto appear in spatially dense clusters, especially in texture-rich regions.As a result, WLS may be dominated by dense inlier groups, leadingto a poorly-conditioned estimation problem that is vulnerable to noiseor degenerate configurations. To address this bottleneck, we propose astructure-aware estimator, which explicitly models inlier correlations viaa graph Laplacian matrix and integrates this prior into a maximum-likelihood framework. Our proposed estimator is fully differentiable, ar-chitecturally lightweight, and can seamlessly replace the standard WLSestimator in existing correspondence learning pipelines. Extensive exper-iments on multiple benchmarks demonstrate consistent improvementswhile introducing a modest increase in computational cost. The code ispublicly available at: https://github.com/Tianyu-Yan/SAE

While contemporary monocular depth estimation (MDE)methods achieve remarkable overall acacy, they consistently produce er-roneous depth discontinuities at object contact regions, particularly be-tween objects and supporting surfaces. In this paper, we address this crit-ical limitation by presenting a boundary-aware monocular depth estima-tion framework that enforces depth continuity at contact areas throughthe principled exploitation of contact boundaries as explicit structuralpriors. Specifically, we propose a boundary detection and filtering modulethat explicitly identifies object contact regions, yielding a novel boundary-aware representation that enables depth-consistent learning at contactareas. Furthermore, we introduce a boundary-aware feature fusion strat-egy that seamlessly incorporates contact boundary priors into the depthdecoding process, effectively rectifying the persistent discontinuities thatelude existing approaches. Our framework further supports interactiverefinement, allowing users to manually specify missing contact bound-aries for controllable depth correction. Extensive experiments on threeunseen benchmarks with dense object interactions demonstrate the ef-fectiveness of our approach, consistently outperforming baselines withparticularly pronounced gains in contact regions. Code is available athttps://github.com/abai969/contact-depth-refinement.git

Vision–Language Models (VLMs) have recently emerged asa powerful paradigm for autonomously perceiving and interacting withcomplex and dynamic Graphical User Interface (GUI) environments.However, their interaction capability remains fundamentally limited byinsufficient precision in spatial grounding. Intriguingly, simply overlay-ing visual markers on screenshots has been extensively shown to sub-stantially enhance the grounding performance of models. Despite its em-pirical effectiveness, the underlying mechanism—specifically, why theseexplicit visual cues can influence the model’s internal spatial understand-ing—remains largely unexplained. We find that the causal influence ofvisual markers on grounding performance is mediated by low-rank la-tent subspaces associated with specific neurons in the late decoder lay-ers. Based on this key insight, we propose DeCoPatch, a novel decode-time causal intervention method that precisely modulates the activa-tions of key neurons relevant to GUI grounding. DeCoPatch follows atwo-stage design: it first identifies top-k critical neurons during the pre-fill stage, effectively isolating a low-rank latent subspace, and then re-inforces these neurons during the decoding stage to causally steer themodel’s spatial grounding behavior. Extensive experiments on four chal-lenging benchmarks demonstrate that DeCoPatch consistently enhancesthe GUI grounding performance of diverse VLMs with almost no addi-tional computational overhead.

Language-aligned vision foundation models perform strongly across diverse downstream tasks. Yet, their learned representations remain opaque, making interpreting their decision-making difficult. Recent work decompose these representations into human-interpretable concepts, but provide poor spatial grounding and are limited to image classification tasks. In this work, we propose CFM, a language-aligned concept foundation model for vision that provides fine-grained concepts, which are human-interpretable and spatially grounded in the input image. When paired with a foundation model with strong semantic representations, we get explanations for any of its downstream tasks. Examining local co-occurrence dependencies of concepts allows us to define concept relationships through which we improve concept naming and obtain richer explanations. On benchmark data, we show that CFM provides performance on classification, segmentation, and captioning that is competitive with opaque foundation models while providing fine-grained, high quality concept-based explanations. Code at https://github.com/kawi19/CFM

3D Gaussian Splatting (3DGS) has emerged as a powerfulrepresentation for high-fidelity rendering. However, existing assets of-ten suffer from quality bottlenecks such as missing details and texturenoise. Prior attempts to enhance these assets via 2D image processingintroduce multi-view inconsistencies and high computational costs. Inthis paper, we propose a novel 3D-native refinement paradigm namedAnchorSplat. AnchorSplat is an end-to-end deep network operatingdirectly on 3D structures, avoiding the expensive optimization overheadof traditional 3D-2D-3D pipelines. Crucially, AnchorSplat is a strictlysource-free solution requiring no original multi-view images. Central tothe proposed method is the Point Anchor Mechanism, which enforcesgeometric consistency via local offset constraints, mitigating ill-posedmapping and gradient confounding. Furthermore, AnchorSplat replacesiterative densification with a single-pass multiplication mechanism. Tofacilitate research, we construct 3DGS-SR, the first large-scale bench-mark for this task. Experiments demonstrate state-of-the-art results onthe 3DGS-SR dataset, with throughput up to 105 times faster thanoptimization methods. Notably, AnchorSplat exhibits robust zero-shotgeneralization across diverse data distributions, including generativemodel outputs and real-world scans. The repository is available at: github

Despite rapid progress, most existing vision-language models (VLMs) built from 2D visual inputs often struggle when handling various 3D tasks that require fine-grained spatial understanding and reasoning. To bridge this gap, we present VLM-IE3D, a unified framework that enhances the 3D spatial awareness of VLMs by equipping them with both implicit and explicit 3D geometries learned from RGB videos. Our VLMIE3D introduces Implicit Geometry Tokens (IGTs) that capture high-level geometric priors from input videos, as well as complementary Explicit Geometry Tokens (EGTs) that encode detailed geometric structures from reconstructed 3D attributes. On top of that, VLM-IE3D comes with a 3D-aware adapter that effectively fuses the two types of geometric representations with 2D visual cues. This RGB-only design injects strong 3D inductive biases for fine-grained spatial understanding and reasoning without requiring any additional 3D inputs. Extensive experiments show that VLM-IE3D achieves superior performance consistently across various 3D tasks including 3D video detection, 3D visual grounding, 3D dense captioning, and spatial reasoning. Code and models are available at https://github.com/Vegetebird/VLM-IE3D.

Efficient strategies to jointly adapt and deploy large language models have seen a growing need under resource-limited conditions for downstream applications. However, when applied to visual foundation models, existing methods typically incur either high GPU memory consumption during adaptation or extra computation costs introduced by the adapters at deployment. In this paper, we propose Efficient Quantization-aware Adaptation (EQuA) that achieves high efficiency in both adaptation and deployment for visual foundation models. We observe that dominant memory consumption arises from intermediate activations cached for backpropagation in the deep backbone and activation quantizers. To address this issue, we split a lightweight sub-network from the backbone during adaptation as a side adapter branch, and tailor two adaptation strategies to eliminate these cached activations, thereby significantly reducing memory consumption. At deployment, the side adapter branch is merged back into the backbone, yielding a quantized model without any extra computation costs. Extensive experiments on representative visual foundation models and diverse downstream tasks exhibit that EQuA achieves an elegant trade-off between performance and efficiency. For example, EQuA yields over 70% GPU memory reduction compared to state-of-the-art baselines while maintaining competitive performance. Code: https://github.com/leenas233/EQuA.

Attention sinks are defined as tokens that attract dispro-portionate attention. While these have been studied in single modalitytransformers, their cross-modal impact in Large Vision-Language Mod-els (LVLM) remains largely unexplored: are they redundant artifactsor essential global priors? This paper first categorizes visual sinks intotwo distinct categories: ViT-emerged sinks (V-sinks), which propagatefrom the vision encoder, and LLM-emerged sinks (L-sinks), which arisewithin deep LLM layers. Based on the new definition, our analysis re-veals a fundamental performance trade-off: while sinks effectively encodeglobal scene-level priors, their dominance can suppress the fine-grainedvisual evidence required for local perception. Furthermore, we identifyspecific functional layers where modulating these sinks most significantlyimpacts downstream performance. To leverage these insights, we proposeLayer-wise Sink Gating (LSG), a lightweight, plug-and-play module thatdynamically scales the attention contributions of V-sink and the restvisual tokens. LSG is trained via standard next-token prediction, re-quiring no task-specific supervision while keeping the LVLM backbonefrozen. In most layers, LSG yields improvements on representative mul-timodal benchmarks, effectively balancing global reasoning and preciselocal evidence.1

Text-driven hand–object interaction (HOI) generation is gain-ing attention for immersive applications and robotics, yet producingphysically plausible interactions remains challenging. Even when individ-ual motions appear natural, small contact errors can cause conspicuousartifacts such as floating and interpenetration. Prior methods mitigatethese issues using explicit contact cues or implicit grasp priors, but typ-ically rely on multi-stage pipelines and fail to model temporally evolvingcontact. We present JointHOI, a single-stage diffusion framework thatjointly generates 3D hand–object motion and dynamic, distance-basedcontact maps from text. By treating contact as an auxiliary inner modal-ity, joint generation enables the model to learn contact–motion couplingduring training. At inference, contact-guided sampling enforces consis-tency between generated contact maps and motion-implied geometry,improving temporal stability and reducing penetration and floating. Ex-periments on GRAB and ARCTIC demonstrate consistent improvementsin text adherence and physical plausibility over prior methods.