ECCV 2026 论文列表

Streaming 4D reconstruction has been demonstrated onlyindoors, on dense camera rigs surrounding subjects that move at hu-man pace. Outdoor 4D reconstruction exists but relies either on camerasmounted on the moving vehicle itself, or on limited-coverage arrays ob-serving quasi-static subjects offline. The case that actually matters forspectators is a fast-moving subject, watched from a sparse ring of allocen-tric cameras, streaming. No method targets this, and no benchmark ex-ists to evaluate one. To this end, we introduce FastFlowGS, a streaming4D Gaussian Splatting method for reconstructing fast-moving subjectsfrom a small set of fixed external cameras, and Monaco4D, a photoreal-istic Unreal Engine 5 benchmark for high-speed outdoor reconstruction.FastFlowGS fuses sparse matches, semi-dense tracks, and dense opticalflow by lifting each signal to 3D with geometric uncertainty and combin-ing them through a Kalman-style temporal update. Monaco4D providesFormula 1 sequences under varied illumination from trackside, onboard,and drone viewpoints with dense ground truth. On CMU-Panoptic (apublic dataset), FastFlowGS exceeds the strongest baseline by 12.6%VMAF at 35% greater efficiency. On Monaco4D, where existing stream-ing methods degrade severely, it improves dynamic-region PSNR by upto 18.6% with 28.3% lower per-frame optimization time.

In this work, we focus on photorealistic sign avatar modeling,which is crucial for effective communication with the Deaf community andis characterized by complex hand gestures and nuanced facial expressions.To this end, we introduce MVSign, the first multi-view Chinese signlanguage dataset co-designed with Deaf experts, featuring diverse gesturesand rich annotations. For precise SMPL-X annotation, we develop a hybridfitting pipeline that produces accurate body, hand, and facial parametersand can also be applied to the monocular setting. Building on MVSign,we propose a decoupled sign avatar representation that isolates body,head, and hand components to capture complex articulations, togetherwith a motion-aware sampling strategy to handle motion blur and balancegesture diversity. Extensive experiments demonstrate that our methodachieves high-fidelity visual results on MVSign, particularly in detailedhand and facial regions, and generalizes well to in-the-wild monocularsign language videos. Project page: https://naaapi.github.io/PHOSA.

Hyperspectral imaging (HSI) promises material-level perception beyond RGB vision, yet current models remain tightly coupled to sensing hardware. Minor changes in spectral discretization or spatial resolution can invalidate trained networks, despite all sensors observing the same material reflectance physics. We introduce OneHSI, a unified hyperspectral foundation model that aligns representations across heterogeneous sensing configurations through physically motivated inductive biases. Rather than treating bands and resolutions as arbitrary tensor dimensions, OneHSI reparameterizes spectral channels in a shared continuous wavelength space, separates scale-invariant relational attention from scale-conditioned feature transformation, and balances cross-domain contributions to learn sensor-invariant representations. Cross-domain experiments demonstrate state-of-the-art performance and strong robustness under sensor and scale shifts, supporting physics-aligned representation learning as a principled path toward cross-sensor generalization.

In recent years, diffusion stochastic differential equation (SDE)inversion and inversion-free methods have become prevalent for training-free image editing, as they can achieve faithful reconstruction withouttuning. However, existing approaches remain inefficient, exhibit limitedplasticity, and struggle to accurately preserve unedited regions. To ad-dress these issues, we propose MIEdit, a training-free editing frame-work based on SDE inversion. MIEdit introduces a predictor–correctormulti-history-step scheme to achieve superior editing quality with fewersteps. We further mitigate heterogeneity and conflict between the multi-conditioned noise residuals and gradient terms during sampling, improv-ing stability and editing plasticity under large edits. MIEdit also includesInversion-Time Automatic Semantic Angle Masking (IASM); it leveragesclassifier-free guidance to automatically generate semantic angle masksduring inversion and applies them throughout the sampling process forregional constraints, without extra user inputs. We additionally constructEditEval++ (30 fine-grained tasks, 1,000+ image–text–mask triplets) forcomprehensive evaluation; experiments show that MIEdit outperformsstate-of-the-art techniques. Project page: https://whywwwzzzg.github.io/MIEdit/.

Volume microscopy, including electron and light microscopy,suffers from severe anisotropic resolution due to physical axial section-ing. Existing self-supervised axial super-resolution (ASR) methods facea trilemma bounded by overly smoothed regression textures, structuralhallucinations of pure diffusion models, and prohibitive inference latency.In this paper, we propose Skeleton-refinE Microscopy (SkelEM), a self-supervised framework that decouples ASR at the training-signal level:a frozen topological network and a diffusion refiner are optimized bydisjoint objectives, separating low-frequency topology formulation fromhigh-frequency detail enhancement. Building on this deterministic skele-ton, we exploit a unified cycle-consistent mechanism on input sparseslices to simultaneously extract a real-domain residual prior and bidi-rectionally align the diffusion refiner, washing away cross-plane arti-facts without synthetic bias. By truncating the reverse diffusion processwith this physical prior, SkelEM achieves high-fidelity detail restorationin merely ≤ 5 steps. To rigorously assess cross-instrument generaliza-tion, we further introduce BRAVE-ASR, a new benchmark of co-alignedanisotropic and isotropic volumes acquired on a Plasma-FIB instrument.Across public benchmarks, SkelEM achieves the most favorable balanceacross the fidelity-perception trade-off among self-supervised methods,with state-of-the-art downstream membrane segmentation performanceand robust zero-shot generalization across distinct modalities.

Continual Learning (CL) enables models to learn sequentially while retaining prior knowledge. Recently, CL under the CLIP framework has gained increasing attention for its remarkable cross-task generalization capability. Despite recent progress, existing approaches overlook modality-asymmetric forgetting, where the visual modality undergoes severe degradation during CL. In this paper, we empirically identify this phenomenon and provide evidence that the lack of a stable structural anchor in the visual branch may drive high-level visual features to over-adapt to new tasks, thereby exacerbating catastrophic forgetting. To address this issue, we propose Visual Anchor-based Structural Transfer (VAST) to stabilize visual representations during continual learning. Specifically, we integrate a cross-layer representation enhancement module to construct the stable visual anchor, and further employ an information-theoretic structural distillation mechanism to transfer them across tasks, promoting stable adaptation and mitigating catastrophic forgetting. Extensive experiments across multiple benchmarks demonstrate that our method consistently surpasses SOTA approaches without relying on any replay data. Code is available at VAST.

Despite the rapid progress of Multimodal Large Language Models (MLLMs), their ability to perform reliable visual grounding in high-stakes clinical software environments remains underexplored. Existing GUI benchmarks largely focus on isolated, single-step grounding queries, overlooking the sequential, workflow-driven reasoning required in real-world medical interfaces, where tasks evolve across interdependent steps and dynamic interface states. We introduce MedSPOT, a workflow-aware sequential grounding benchmark for clinical GUI environments. Unlike prior benchmarks that treat grounding as a standalone prediction task, MedSPOT models procedural interaction as a sequence of structured spatial decisions. The benchmark comprises 216 task-driven videos with 597 annotated keyframes, in which each task comprises 2–3 interdependent grounding steps within realistic medical workflows. This design captures interface hierarchies, contextual dependencies, and fine-grained spatial precision under evolving conditions. To evaluate procedural robustness, we propose a strict sequential evaluation protocol that terminates task assessment upon the first incorrect grounding prediction, explicitly measuring error propagation in multistep workflows. We further introduce a comprehensive failure taxonomy, including edge bias, small-target errors, no prediction, near miss, far miss, and toolbar confusion, to enable systematic diagnosis of model behavior in clinical GUI settings. By shifting evaluation from isolated grounding to workflow-aware sequential reasoning, MedSPOT establishes a realistic and safety-critical benchmark for assessing multimodal models in medical software environments. Code and data are available at https://github.com/Tajamul21/MedSPOT

Although large vision-language models (LVLMs) have demonstrated remarkable capabilities, they are prone to hallucinations in multiimage tasks. We attribute this issue to limitations in existing attention mechanisms and insufficient cross-image modeling. Inspired by this, we propose a structured hallucination mitigation framework involving Cross-Image Attention calibration and Preference Learning (CAPL). CAPL explicitly enhances inter-image interactions at the architectural level while reinforcing reliance on genuine cross-image evidence during training, thereby improving the model’s perception and modeling of cross-image associations. Specifically, we (i) introduce a selectable image token interaction attention mechanism to establish fine-grained crossimage entity alignment and information flow; (ii) design a cross-image modeling–based preference optimization strategy that contrasts reasoning outcomes under full inter-image interaction and those obtained when images are mutually invisible, encouraging the model to ground its predictions in authentic visual evidence and mitigating erroneous inferences driven by textual priors. Experimental results demonstrate that CAPL consistently improves performance across multiple model architectures, achieving stable gains on both multi-image hallucination and general benchmarks. Notably, performance on single-image visual tasks remains stable or slightly improves, indicating strong generalization capability. The code is available at https://github.com/yyyxcleo/CAPL.

Feedforward geometric foundation models achieve strongshort-window reconstruction, yet scaling them to minutes-long videos isbottlenecked by quadratic attention complexity or limited effective mem-ory in recurrent designs. We present LoGeR (Long-context GeometricReconstruction), a novel architecture that scales dense 3D reconstruc-tion to extremely long sequences without post-optimization. LoGeRprocesses video streams in chunks, leveraging strong bidirectional priorsfor high-fidelity intra-chunk reasoning. To manage the critical challengeof coherence across chunk boundaries, we propose a learning-based hybridmemory module. This dual-component system combines a parametric Test-Time Training (TTT) memory to anchor the global coordinate frame andprevent scale drift, alongside a non-parametric Sliding Window Attention(SWA) mechanism to preserve uncompressed context for high-precisionadjacent alignment. Notably, this memory architecture allows LoGeR togeneralize to sequences of thousands of frames during inference, despitebeing trained on only 128 frames. When evaluated on standard bench-marks and a repurposed VBR dataset with sequences up to 19k frames,LoGeR substantially outperforms current state-of-the-art feedforwardmethods. It reduces the ATE on KITTI by over 74% while achievingrobust, globally consistent reconstruction across unprecedented horizons.

Cloud removal (CR) is essential for optical remote sensing,serving as a prerequisite for reliable downstream interpretation, suchas semantic segmentation and change detection. However, existing CRapproaches often prioritize visual realism while overlooking their im-pact on subsequent analytical tasks, leading to semantic drift and de-graded downstream performance. To address this issue, we propose Geo-Anchored Cloud Removal (GACR), a unified framework that jointlyensures faithful reconstruction and robust interpretability. At its core,GACR incorporates Observation-Anchored Residual Flow (OAR-Flow),which reformulates CR as a physically grounded residual inversion pro-cess. By anchoring the generative trajectory to the cloudy observationrather than pure noise, OAR-Flow enables fast, stable, and faithful re-construction. To further preserve semantic structures critical for down-stream interpretation, GACR integrates Geo-Contextual Prior Align-ment (GCPA) to constrain the reconstruction within a semantic manifoldinduced by a Vision Foundation Model (VFM). Consequently, GACRstrictly maintains the spatial-semantic integrity of complex landscapes.Extensive experiments across six CR datasets and twelve downstreamtasks demonstrate that GACR produces superior reconstruction qual-ity while consistently improving downstream task accuracy. The code isavailable at https://github.com/wzy6055/GACR.

Recent breakthroughs in 3D generation have advanced no-tably with the development of text-to-image diffusion model. However,existing methods remain two practical challenges: (1) They primarilygenerate single 3D object, but struggle to generate multi-object compo-sitional 3D assets due to the lack of the modeling for Gaussian prim-itives in reasonable interactions. (2) They often suffer from cross-viewinconsistency during 3D optimization, as Score Distillation Samplinginherently performs on each single view, inevitably resulting in cross-view hallucinations. To solve above issues, we propose I2 C-3D, a noveloptimization-based method to generate multi-view consistent composi-tional 3D assets with reasonable interactions. Specifically, we proposean Inclusive Interactive Collisions strategy to guide Gaussian primitivesappearing in reasonable interaction regions naturally, thereby ensuringobjects in the compositional scene interact in a physically plausible andvisually coherent way. Additionally, to enhance multi-view consistency,Multi-View Adaptive Score Distillation Sampling is devised to distillmulti-view consistency prior and layout prior from pre-trained diffusionmodel by modulating attention map of instance token and spatial to-ken across viewpoints. Benefiting from above elaborate designs, I2 C-3Dnot only generates high-fidelity multi-view consistent compositional 3Dassets but also supports 3D editing flexibly, facilitating complex scenegeneration. Extensive experiments demonstrate our I2 C-3D outperformsexisting methods in generation quality and multi-view consistency.

In this work, we propose Image-to-Image Rectified Flow Re-formulation (I2I-RFR), a practical plug-in reformulation that recastsstandard I2I regression networks as continuous-time transport models.While pixel-wise I2I regression is simple, stable, and easy to adapt acrosstasks, it often over-smooths ill-posed and multi-modal targets, whereasgenerative alternatives often require additional components, task-specifictuning, and more complex training and inference pipelines. Our methodaugments the backbone input by channel-wise concatenation with a noise-corrupted version of the ground-truth target and optimizes a simplet-reweighted pixel loss. This objective admits a rectified-flow interpre-tation via an induced velocity field, enabling ODE-based progressiverefinement at inference time while largely preserving the standard su-pervised training pipeline. In most cases, adopting I2I-RFR requiresonly expanding the input channels, and inference can be performed witha few explicit solver steps (e.g., 3 steps) without distillation. Exten-sive experiments across multiple image-to-image translation and videorestoration tasks show broad perceptual improvements, with possibleperception–distortion trade-offs in highly pixel-aligned settings such assuper-resolution.

We investigate the token mixer in vision backbones by revisit-ing clustering, one of the most classic approaches in machine learning. Aneffective token mixer is a fundamental component of modern vision back-bones like vision Transformers, facilitating information exchange betweenimage patches. Mainstream token mixers, which rely on convolution, at-tention, MLP, or their hybrids, primarily focus on navigating the trade-offbetween accuracy and computational cost. However, a significant draw-back of these methods is their black-box nature; their encoding processis opaque and lacks interpretability. Diverging from these opaque designs,we introduce ClusterMixer, a transparent token mixer that is grounded ina clustering paradigm and interpretable by design. ClusterMixer explicitlyformulates the token mixing process through a hierarchical clusteringmechanism. To model the natural, tree-like relationships inherent in visualdata, the clustering is performed in hyperbolic space, which is well-suitedfor embedding hierarchies with low distortion. Building on this innovation,we present HCFormer, a new backbone architecture that integratesClusterMixer with a series of meticulously designed clustering strategiesto ensure robust performance across tasks. Extensive experiments demon-strate that HCFormer consistently outperforms its counterparts acrossdiverse tasks, including image classification, object detection, instancesegmentation, and semantic segmentation. Considering its transparencyand efficacy, we hope HCFormer can facilitate a paradigm shift towardinterpretable backbones.

Monocular 3D lane detection plays a critical role in autonomous driv-ing, yet recovering reliable 3D geometry from a single image remains challeng-ing due to inherent depth ambiguity. Prior methods project image features intoBird’s-Eye-View (BEV) space under a flat-ground assumption, causing geomet-ric distortion on real-world roads. Recent methods instead predict explicit heightmaps to capture non-planar surfaces, but still rely on sparse anchor-based regres-sion and exploit the recovered geometry merely for spatial transformation ratherthan semantic understanding. To overcome these limitations, we propose HSDF-Lane, which implicitly models the road surface as a Height-aligned Signed Dis-tance Field (HSDF) over a densely sampled 3D feature volume. Through dif-ferentiable rendering, the HSDF jointly produces an accurate height map andsurface-aligned features. We further introduce Lane-aware Semantic PositionalEncoding (LSPE), which injects a lane-existence prior derived from the surface-aligned features into the transformer queries, coupling geometric structure withsemantic guidance. Extensive experiments on the OpenLane benchmark showthat HSDF-Lane achieves state-of-the-art performance in both 3D lane detectionand height map estimation. The code is available at https://github.com/JiyongBoo/HSDF-Lane.

Zero-shot object counting aims to count objects in images from givenclass names. Existing methods typically select representative image patches orextract text embeddings as exemplars to match image regions for counting. How-ever, a limited number of selected patches is insufficient to cover intra-class di-versity, and the accumulated errors in exemplar detection and selection substan-tially degrade counting accuracy. While the text embeddings remain at a genericsemantic level, blind to the specific instance diversity of target objects in animage. To handle these issues, this paper proposes the Hyper-Exemplar Repre-sentation (HER) counting framework (HER-Count) to represent the objects ofa specific category in a given image. Rather than image patches or text embed-dings, HERs are synthesized from holistic images and textual class names viaan multi-modal large-language model. By jointly integrating all target objects ineach image, HERs avoid exemplar generation errors and are more customizedrepresentative, thus yielding more accurate counting results. To effectively ex-ploit HERs in counting, HER-Count introduces a hierarchical fusion strategy thatinjects HERs into multiple intermediate layers of the vision encoder when pre-dicting density maps. This enables the model to progressively refine countingcues across layers, significantly enhancing the performance. A global discrimi-nation enhancement constraint at both image sample and semantic class levels isdesigned to further improve the capacities for both representation and discrimi-nation of HERs. Extensive experiments on widely used benchmarks demonstrateboth the state-of-the-art performance and generalization ability of the proposedHER-Count. Benefiting from the compact end-to-end design, HER-Count alsoachieves promising inference speed.

While Multimodal Large Language Models demonstrate im-pressive semantic capabilities, they often suffer from spatial blindness,struggling with fine-grained geometric reasoning and physical dynamics.Existing solutions typically rely on explicit 3D modalities or complexgeometric scaffolding, which are limited by data scarcity and generaliza-tion challenges. In this work, we propose a paradigm shift by leveragingthe implicit spatial prior within large-scale video generation models. Weposit that to synthesize temporally coherent videos, these models inher-ently learn robust 3D structural priors and physical laws. We introduceVEGA-3D (Video Extracted Generative Awareness), a plug-and-playframework that repurposes a pre-trained video diffusion model as a La-tent World Simulator. By extracting spatiotemporal features from inter-mediate noise levels and integrating them with semantic representationsvia a token-level adaptive gated fusion mechanism, we enrich MLLMswith dense geometric cues without explicit 3D supervision. Extensiveexperiments across 3D scene understanding, spatial reasoning, and em-bodied manipulation benchmarks demonstrate that our method outper-forms state-of-the-art baselines, validating that generative priors providea scalable foundation for physical-world understanding. Code is publiclyavailable at https://github.com/H-EmbodVis/VEGA-3D.

Recalling that diffusion models map noisy images to cleanones, the choice of noise addition (i.e., corruption) critically influencesthe denoising process. However, prior diffusion-based image restoration (IR)methods often employ naïve corruption strategies—such as randomlysampling noise, fixing the timestep, or even omitting corruption alto-gether—which may not accurately reflect the actual degradation of eachimage. To handle this, we propose Input-Aware Corruption for IR (IAC-IR) framework, which maps each low-quality (LQ) input into an optimalnoisy sample that lies on the pretrained diffusion trajectory. Instead ofchoosing corruption heuristically, we predict the timestep and noise us-ing supervision derived from the properties of the pretrained diffusionmodel. Specifically, the predicted timestep aligns the corrupted samplewith the Gaussian noise corrupted distribution, while the predicted noisepreserves the recoverable content of the input. Moreover, we use theseinput-aware corruption factors to improve conventional score-based dis-tillation. Rather than relying on random corruption, which often produceunreliable target scores and weak gradients, we perform distillation withinput-aware corruption, yielding more reliable score estimates and morestable distillation. By modeling input-aware corruption and integratingit into distillation, our method better leverages the pretrained diffusionprior, achieving superior perceptual quality in image restoration.

Real-world image restoration (RWIR) is a highly challeng-ing task due to the absence of clean ground-truth images. Many recentmethods resort to pseudo-label (PL) supervision, often within a Mean-Teacher (MT) framework, where a teacher network generates targets fora student network. However, these methods face a critical paradox: un-conditionally trusting the often imperfect, low-quality PLs forces the stu-dent model to learn undesirable artifacts, while discarding them severelylimits data diversity and impairs model generalization. In this paper, wepropose QualiTeacher, a novel framework that transforms pseudo-labelquality from a noisy liability into a conditional supervisory signal. In-stead of filtering, QualiTeacher explicitly conditions the student modelon the quality of the PLs, estimated by an ensemble of complementarynon-reference image quality assessment (NR-IQA) models spanning low-level distortion and semantic-level assessment. This strategy teaches thestudent network to learn a quality-graded restoration manifold, enablingit to understand what constitutes different quality levels. Consequently,it can not only avoid mimicking artifacts from low-quality labels but alsoextrapolate to generate results of higher quality than the teacher itself.To ensure the robustness and accuracy of this quality-driven learning,we further enhance the process with a multi-augmentation scheme to di-versify the PL quality spectrum, a score-based preference optimizationstrategy inspired by Direct Preference Optimization (DPO) to enforcea monotonically ordered quality separation, and a cropped consistencyloss to prevent adversarial over-optimization (reward hacking) of theIQA models. Experiments on standard RWIR benchmarks demonstratethat QualiTeacher can serve as a plug-and-play strategy to improve thequality of the existing pseudo-labeling framework, establishing a newparadigm for learning from imperfect supervision. Code will be releasedat https://github.com/fengyang1399-pixel/QualiTeacher.git .6.87.52 F. Xiao et al.(a) Unconditional Trust (b) Aggressive Filtering (c) QualiTeacher5.2 6.8 7.5 5.2 6.8 7.5Aug AugTeacher Teacher NR-IQA Teacher NR-IQAInput PL Input Optimal PL Input Quality-Conditioned PLsArtifacts are replicated rather than removed Limited data diversity, over-smoothed and blurry Better generalization via quality-conditioned pseudo-labelsDegraded Unconditional Trust Aggressive Filtering QualiTeacherFig. 1: Comparison of pseudo-label (PL) utilization strategies in the mean-teacherframework. (a) Unconditional Trust: PLs are used without filtering, causing degra-dation artifacts to be replicated by the student. (b) Aggressive Filtering: Low-quality PLs are discarded via NR-IQA filtering, yet over-smoothed outputs that receivedeceptively high scores survive, introducing blurriness. (c) QualiTeacher (Ours):Nearly all PLs are retained (with only extreme outliers discarded), with their qualityscores injected as continuous conditioning signals, enabling the student to leverage fulldata diversity while remaining aware of each sample’s reliability.

Visual intelligence requires anticipating the future behaviorof agents, yet vision systems lack a general representation for motion andbehavior. We propose dense point trajectories as visual tokens for behavior,a structured mid-level representation that disentangles motion fromappearance and generalizes across diverse non-rigid agents, such as animalsin-the-wild. Building on this abstraction, we design a diffusion transformerthat models unordered sets of trajectories and explicitly reasons aboutocclusion, enabling coherent forecasts of complex motion patterns. Toevaluate at scale, we curate 300 hours of unconstrained animal motion fromvideo through robust shot detection and camera-motion compensation.Experiments show that forecasting trajectory tokens achieves category-agnostic, data-efficient prediction, outperforms state-of-the-art baselines,and generalizes to rare species and morphologies, providing a foundationfor predictive visual intelligence in the wild.