ECCV 2026 Papers

Vision-Language Models (VLMs) enable autonomous GUInavigation, but agents still struggle to process and learn from dense,continuous visual histories. This bottleneck hinders both immediate errorcorrection within a single episode (intra-trial) and experience distilla-tion across multiple attempts (cross-trial). We trace these challenges toan empirical informational asymmetry in GUI navigation: while expectedtransitions can often be compressed into lightweight textual summaries,unexpected outcomes benefit from preserved screenshots as causal evi-dence for accurate diagnosis. Building on this insight, we propose An-chorGUI, a unified framework driven by the Cognitive State Anchor(CSA). The CSA acts as a per-step primitive that actively comparesexpected and observed transitions, converting passive multimodal tra-jectories into explicit prediction-error signals. These signals orchestratea dual-scale learning mechanism via an asymmetric memory. Forintra-trial correction, a sliding window selectively retains visual evidencefor detected mismatches, providing immediate, visually-grounded feed-back. For cross-trial distillation, this asymmetric memory focuses thecomputationally expensive credit assignment search space on likely fail-ure steps. Experiments across four benchmarks validate the effectivenessof our approach. On AndroidWorld, AnchorGUI achieves a 57.3% successrate with a 2.4× token reduction per step. Furthermore, cross-trial distil-lation reaches 69.2% success (+11.9% gain), significantly outperformingstandard reflection methods while maintaining sub-linear context scaling.

Small Object Detection (SOD) is a fundamental yet chal-lenging problem in computer vision due to its limited spatial resolu-tion and weak visual cues. Although recent approaches have achievedremarkable advances, the background distractors in different frequencyspectra still degrade the performance. In this paper, we propose a novelsmall object detection framework termed SFDNet, which is capableof detecting small objects via efficient spectrum-aware feature disentan-glement. Specifically, we propose an Adaptive Spectrum Disentangle-ment (ASD) module that decomposes backbone features into multiplecomplementary spectral components, aiming to construct discriminativeobject-relevant representations by discarding the background distrac-tors for each component. Afterwards, to strengthen the semantic con-sistency of the similar objects in the same class, we propose a Class-Wise Prototype Distillation (CPD) procedure, which establishes classprototypes for the object instances and enforces the compact repre-sentation by efficient prototype distillation. Extensive experiments onmultiple challenging benchmarks show that SFDNet outperforms exist-ing state-of-the-art methods by a large margin. Our code is available athttps://github.com/ManOfStory/SFDNet.

Our objective is to build an embedding model that capturesthe nuanced relationship between a search query and candidate videos.We cover three aspects of nuanced retrieval: (i) temporal, (ii) negation,and (iii) multimodal. For temporal nuance, we consider chiral actionsthat need distinguishing between temporally opposite actions like “open-ing a door” vs. “closing a door”. For negation, we consider queries withnegators such as “not”, “none” that allow a user to specify what theydo not want. For multimodal nuance, we consider the task of composedretrieval where the query comprises a video along with a text edit in-struction. The goal is to develop a unified embedding model that handlessuch nuances effectively. To that end, we repurpose a Multimodal LargeLanguage Model (MLLM) trained to generate text into an embeddingmodel. We fine-tune it with a contrastive loss on text alone with carefullysampled hard negatives that instill the desired nuances in the learned em-bedding space. Despite the text-only training, our method achieves stateof the art performance on all benchmarks for nuanced video retrieval. Wealso analyze how this improvement is achieved, and show that text-onlytraining reduces the modality gap between text and video embeddingsleading to better organization of the embedding space.

While video streaming understanding has made significantstrides, real-world applications, such as live sports broadcasting, au-tonomous driving, and multi-screen collaboration, inherently demandcontinuous, multi-stream interactions. However, existing benchmarks areconfined to single-stream paradigms, leaving a critical gap in evaluatingonline, cross-stream reasoning. To bridge this, we introduce X-Stream,the first benchmark dedicated to multi-stream streaming understand-ing. Comprising 4,220 rigorously curated QA pairs across 932 videos,X-Stream evaluates 11 subtasks across multi-window, multi-view, andmulti-device scenarios. Crucially, our dataset is constructed using a noveldual-verification pipeline that prevents over-reliance on a single stream.Furthermore, we pioneer the conceptualization of multi-modal large lan-guage models (MLLMs) as naive multiplexers, systematically evaluat-ing their performance through the lens of Signal Multiplexing Theory. Ourextensive online inference experiments reveal a stark reality: state-of-the-art MLLMs struggle significantly with concurrent streams, achieving only∼50% score and exhibiting a poor proactive ability. Ultimately, X-Streamexposes the trade-off of current multiplexing schemes, providing both apractical evaluation protocol and empirical guidance for next-generationmulti-stream agents. Code and data are released at Œhomepage.

Learned image compression (LIC) operates under a rate–distortion (RD) trade-off, where representation quality is constrained bythe target bitrate. We revisit knowledge distillation in LIC through thelens of bitrate asymmetry and introduce a self-teacher distillation frame-work, where a high-rate instance of a codec supervises multiple lower-rateencoders of identical architecture. Because allocating more bits naturallyleads to richer latent representations, the high-rate model provides infor-mative supervision across rate levels. Direct latent matching, however,is problematic under tight rate budgets. We therefore propose variance-normalized latent distillation (VNLD), a rate-aware alignment strategythat scales channel-wise supervision by the teacher’s variance, selectivelytransferring stable, informative structure while suppressing componentsthat cannot be reliably reproduced at lower rates. Across different dis-tillation objectives, networks, and bitrate levels, self-teacher distillation,particularly with VNLD, improves RD performance and yields consistentBD-rate gains over RD-only training. Our method remains compatiblewith fixed-decoder deployments, such as those targeted by JPEG AIstandards.

Online Video Large Language Models (VideoLLMs) playa critical role in supporting responsive, real-time interaction. Existingmethods focus on streaming perception, lacking a synchronized logicalreasoning stream. However, directly applying test-time scaling methodsincurs unacceptable response latency. To address this trade-off, we pro-pose Video Streaming Thinking (VST), a novel paradigm for streamingvideo understanding. It supports a thinking-while-watching mechanism,which activates reasoning over incoming video clips during streaming.This design improves timely comprehension and coherent cognition whilepreserving real-time responsiveness by amortizing LLM reasoning latencyover video playback. Furthermore, we introduce a comprehensive post-training pipeline that integrates VST-SFT, which structurally adapts theoffline VideoLLM to causal streaming reasoning, and VST-RL, whichprovides end-to-end improvement through self-exploration in a multi-turn video interaction environment. Additionally, we devise an auto-mated training-data synthesis pipeline that uses video knowledge graphsto generate high-quality streaming QA pairs, with an entity–relationgrounded streaming Chain-of-Thought to enforce multi-evidence reason-ing and sustained attention to the video stream. Extensive evaluationsshow that VST-7B performs strongly on online benchmarks, e.g. 79.5%on StreamingBench and 59.3% on OVO-Bench. Meanwhile, VST remainscompetitive on offline long-form or reasoning benchmarks. Comparedwith Video-R1, VST responds 15.7× faster and achieves +5.4% improve-ment on VideoHolmes, demonstrating higher efficiency and strong gen-eralization across diverse video understanding tasks. Code, data, andmodels have been released at https://github.com/1ranGuan/VST.

Recovering dynamic environment maps from a single in-the-wild video is crucial for photorealistic rendering, yet remains a challenge.Recent video generation models can produce photorealistic scenes withcomplex lighting, possessing an inherent understanding of lighting. Inthis paper, we introduce V-LITE (Video generation models are inherentlighting estimators), a framework that unlocks this internal knowledge byreframing lighting estimation as a guided video inpainting task. Inspiredby visual effects (VFX) industry practices, we insert a synthetic chromeball into the scene to compel the model to generate physically plausiblereflections from the surrounding spatio-temporal context. To bridge thegap from LDR-native models to the HDR domain, we design an HDR-aware VAE and employ an efficient LoRA-based fine-tuning strategy. Wethen construct a mixed dataset comprising high-fidelity HDR images toprovide realistic HDR priors, and in-the-wild HDR videos to provide dy-namic spatio-temporal context. Extensive experiments demonstrate thatV-LITE produces temporally coherent HDR environment maps, reveal-ing that modern video diffusion models are not merely synthesizers butalso powerful, inherently capable estimators of physical scene lighting.

Parameter-efficient fine-tuning (PEFT) has become a prac-tical solution for adapting large pretrained vision transformers (ViTs)to downstream tasks while updating only a small subset of parame-ters. However, existing adapter-based methods perform adaptation in-dependently for each token, implicitly assuming that token refinementsshould be learned in isolation. This token-wise formulation overlooksthe structured relationships among tokens that naturally arise in visualscenes, potentially leading to redundant updates and spatially inconsis-tent feature refinement. In this work, we revisit the design of parameter-efficient adapters and propose to perform adaptation in hyperedge spacerather than token space. We introduce HyperAdapter, a hypergraph-based adapter architecture that enables structured, group-aware adap-tation through soft token routing. HyperAdapter constructs a soft hy-pergraph over ViT tokens using prototype-based assignments, aggregatestoken features into latent hyperedge representations, applies lightweightbottleneck adaptation at the hyperedge level, and diffuses the result-ing updates back to tokens via the hypergraph incidence structure. Thisdesign injects an explicit structural inductive bias into PEFT while pre-serving the modularity and efficiency of standard adapters. Extensiveexperiments across diverse visual benchmarks demonstrate that struc-tured hyperedge adaptation consistently outperforms strong PEFT base-lines under comparable parameter budgets, with particularly pronouncedgains on tasks requiring structured reasoning. Our results suggest thatthe choice of adaptation space is a critical yet underexplored dimensionin parameter-efficient transfer for ViTs.

Unified Multimodal Models (UMMs) consolidate visual un-derstanding, image generation, editing, and interleaved image-text in-teraction, yet their evaluation increasingly depends on VLM-as-a-judgescoring that shifts as judges, prompts, or model versions evolve — aproblem we term evaluation drift. We present SpecV, a specification-verification framework for robust unified multimodal evaluation. SpecVreplaces holistic scoring with the Specification Verification Proto-col, which decomposes each prompt into atomic, binary specificationsand verifies them against model outputs, improving cross-judge agree-ment and stability while preserving alignment with human judgments.To produce reliable specifications, we propose Specification Ensem-ble and Refinement, a multi-model pipeline that aggregates candidatespecifications, deduplicates them semantically, and filters for verifiabil-ity and relevance. We also introduce SpecV-Bench, a 1,200-instancebenchmark covering six core UMM tasks with sub-tracks of increasingconstraint complexity, enabling fine-grained analysis of capability trade-offs and failure modes. Across extensive experiments with multiple judgemodels, SVP consistently reduces ranking flips and improves evalua-tion reproducibility. Benchmark is available at https://github.com/yuyouxixi/SpecV.

Sign language models are typically trained on datasets cap-tured under constrained conditions, with limited viewpoint, background,and signer-identity diversity, leading to poor robustness under real-worlddistribution shifts. We introduce SignNet-1M, a large-scale augmenteddataset spanning ASL, CSL, and German Sign Language (DGS).SignNet-1M synthesizes realistic variations along three axes: (i) novel-view rendering (rotation and zoom) via 3D Gaussian Splatting (3DGS),(ii) scene/ identity editing via diffusion models for background replace-ment and signer substitution while preserving sign motion and linguisticcontent, and (iii) post-rendering augmentations that emulate captureand compression artifacts (e.g., geometric transforms, photometric shifts,mild temporal resampling, and compression) to better match in-the-wildrecordings. Beyond data release, we provide a unified benchmark suiteacross downstream tasks (e.g., translation and recognition) and abla-tions that isolate each augmentation component. Experiments acrossbackbones show that training with SignNet-1M consistently improvesgeneralization under cross-view, cross-background, cross-identity, andpost-rendering shifts, while maintaining strong in-distribution perfor-mance. The dataset, released augmentation components, metadata, andbenchmark resources are available at https://signnet.chatsign.ai/.

Compositions of frozen foundation models offer a trainingfree route to instance segmentation, yet a clear gap remains between current systems and their empirical upper bounds. We introduce progressive oracle replacement, a diagnostic procedure that decomposes this gap into stage-level components, and find that the scoring stage accounts for most of it. Across diverse encoders and mask generators, we trace this gap to Quality Misalignment: a mismatch between semantic matching confidence and mask geometric fidelity that yields ranking errors when the two signals are used independently. This pattern persists across all tested configurations, suggesting that it is intrinsic to decoupled scoring. We therefore propose Semantic-Geometric Quality Alignment (SGQA), which replaces decoupled scoring with a single quality score formed by the geometric mean of the two signals and adds no hyperparameters to the scoring formula. Under a fixed backbone of DINOv2 ViT-L paired with SAM2 Hiera-L, SGQA outperforms existing training-free methods and matches or exceeds fine-tuning-based methods with comparable or smaller backbones.

Monocular depth estimation remains challenging for trans-parent objects, where refraction and transmission are difficult to modeland break the appearance assumptions used by depth networks. As a re-sult, state-of-the-art estimators often produce unstable or incorrect depthpredictions for transparent materials. We propose SeeClear, a novelframework that converts transparent objects into diffusion-generated opaqueimages, enabling stable monocular depth estimation for transparent ob-jects. Given an input image, we first localize transparent regions andtransform their refractive appearance into geometrically consistent opaqueappearances using a diffusion-based generative opacification module. Theprocessed image is then fed into an off-the-shelf monocular depth esti-mator without retraining or architectural changes. To train the opaci-fication model, we construct SeeClear-396k, a synthetic dataset con-taining 396k rendered images across 66k paired transparent-opaque con-figurations. Experiments on both synthetic and real-world datasets showthat SeeClear significantly improves depth estimation for transparentobjects.

State-of-the-art post-hoc out-of-distribution detection meth-ods rely on intermediate layer activation editing. However, they exhibitinconsistent performance across datasets and models. We show that thisinstability is driven by differences in the activation distributions, andidentify a failure mode of scaling-based methods that arises when penul-timate layer activations are not rectified. Motivated by this analysis,we propose RAS, a hyperparameter-free post-hoc method that replacessorted activation magnitudes with a fixed in-distribution reference profile.Our simple plug-and-play method shows strong and consistent perfor-mance across datasets and architectures without assumptions on thepenultimate layer activation function, and without requiring any hyperpa-rameter tuning, while empirically preserving in-distribution classificationaccuracy. We further analyze what drives the improvement, showing thatboth inhibiting and exciting activation shifts independently contribute tobetter out-of-distribution discrimination1 .

Although recent trajectory prediction and end-to-end au-tonomous driving methods improve robustness in urban environments,they still lack meaningful controllability. Existing benchmarks either pro-vide no persona-conditioned annotations or support only a single urgencyspectrum (i.e., emergency, normal, relaxed), which cannot distinguishpersonas that share the same urgency level but require different drivingdynamics. To address this, we propose (i) the Persona-Conditioned Tra-jectory (PCT) dataset, which decomposes driving personas along twoaxes—Temporal Urgency and Ride Comfort—and combines three levelsof each to form a grid of nine personas, each paired with natural-languagedescriptions and trajectories, and (ii) PersonaDrive, a framework thatcan learn driving personas from language and can generate persona-specific trajectories. PersonaDrive incorporates Persona-Conditioned An-chor Transform (PCAT), which hierarchically reshapes anchors alongboth axes, and Persona-Conditioned Multi-Modal Fusion (PCMF) forBEV-level persona fusion. Training is supervised by a Hierarchical GuideLoss enforcing axis-aligned physical orderings and an Axis-DecomposedDiversity Loss preventing diagonal mode collapse. Experimental resultsshow that PersonaDrive consistently improves over the compared base-lines across multi-dimensional scenarios. The code and PCT dataset areavailable at https://github.com/VisualAIKHU/PersonaDrive.

Modern smartphones capture Live Photos—short video bursts sur-rounding a still image—offering a dynamic and engaging photographic experi-ence. However, the cover photo and video components are generated by two dis-tinct imaging pipelines: the photo stream undergoes full computational photog-raphy processing, while the video stream is constrained by real-time efficiencyand heavy compression. This intrinsic separation produces a substantial qual-ity gap in resolution, color fidelity, and dynamic range between the cover photoand video frames. When users reselect an alternative frame from the video toreplace an imperfect cover, the chosen frame often suffers from severe degrada-tion, making direct replacement visually unsatisfactory. Restoring such framesrequires simultaneous enhancement of spatial detail and color appearance, a taskconsiderably more challenging than ordinary super-resolution or color enhance-ment. To address this, we define the Live Photo Cover Frame Reselection andEnhancement (LPRE) task, which leverages the intrinsic cues available withineach Live Photo: the high-quality cover image as a structural and color reference,the user-reselected low-quality frame as the reconstruction target and several ad-jacent video frames providing temporal cues. Building upon this formulation, weconstruct Live2K, a real-world dataset of 2,042 Live Photos, and develop a unifiedone-stage baseline that integrates multi-frame fusion, guided color enhancementand super-resolution—establishing the first benchmark for Live Photo enhance-ment research.

Omnidirectional stereo matching (OSM) estimates 360◦ depthfrom multi-view fisheye images, where inherent fisheye distortion causessparse matching cues in seam regions. However, existing methods rely onspatially uniform update policies, degrading performance in these chal-lenging regions. To address this limitation, we propose PDF-Omni, ageometry-aware OSM framework built on a Poincaré Dual Disk distor-tion field that models spatially varying geometric uncertainty across seamregions. We introduce a distortion-informed GRU with distortion-awareattention to selectively expand the receptive field in high-uncertaintyregions while preserving efficiency in low-uncertainty ones. We furtherpresent a local entropy cumulative loss to supervise high-entropy costdistributions in ambiguous regions. Experimental results demonstratestate-of-the-art performance with consistent gains especially in seam re-gions. The code is available at https://github.com/ynskyang/PDF-Omni

Distilled autoregressive di!usion models facilitate real-timeshort video synthesis but su!er from severe error accumulation duringlong-sequence generation. While existing Test-Time Optimization (TTO)methods prove e!ective for images or short clips, we identify that they failto mitigate drift in extended sequences due to unstable reward landscapesand the hypersensitivity of distilled parameters. To overcome these limi-tations, we introduce Test-Time Correction (TTC), a training-free alter-native. Specifically, TTC utilizes the initial frame as a stable referenceanchor to calibrate intermediate stochastic states along the sampling tra-jectory. Extensive experiments demonstrate that our method seamlesslyintegrates with various distilled models, extending generation lengthswith a slight overhead while matching the quality of resource-intensivetraining-based methods on 30-second benchmarks.

Bird’s-eye view (BEV) perception fuses multi-camera imagesinto a unified top-down representation for autonomous driving. Despiterecent progress, state-of-the-art methods remain confined to closed-setscenarios, making them vulnerable to unpredictable real-world environ-ments. In this work, we introduce open-vocabulary BEV segmentation(OVBS), which leverages vision-language models (VLMs) to recognizecategories beyond the training set while maintaining precise BEV percep-tion and real-time efficiency. A key challenge in OVBS lies in the 3D geo-metric inconsistency inherent in the ill-posed lifting of 2D VLM semanticsinto BEV. To address this, we propose OVBEVSeg, a geometry-awareOVBS framework that enhances efficient Gaussian splatting (GS)-basedunprojection by leveraging robust 3D geometric constraints across threeprogressive stages: (1) 2D-to-BEV pseudo-labeling via reliable 3D pro-jection for OV generalization; (2) joint 2D–BEV per-scene optimizationwith BEV structural constraints for 3D geometric consistency; and (3)3D geometric distillation for online efficiency. On the nuScenes dataset,OVBEVSeg achieves state-of-the-art performance, outperforming closed-set methods by 15.3 mIoU on unseen categories. Remarkably, even withno novel-class ground-truth labels, it remains competitive with self- andsemi-supervised baselines trained with up to 40% of ground-truth anno-tations. Furthermore, it achieves 2.5× faster inference with only 0.22×the memory consumption of projection-based methods.

We tackle the problem of sparse novel view synthesis (NVS)using video diffusion models; given K (≈ 5) posed images of a scene,we predict the view from a target camera pose. Many prior approachesleverage generative image priors encoded via diffusion models. However,models trained on single images lack multi-view knowledge, and we in-stead argue that video models contain implicit multi-view knowledge andare therefore easier to adapt for NVS. Our key insight is to formulatesparse NVS as a low-frame-rate video completion task. However, onechallenge is that sparse NVS is defined over an unordered set of inputimages, often too sparse to admit a meaningful order, so the modelsshould be invariant to permutations of that input image set. To thisend, we present FrameCrafter, which adapts video models (naturallytrained with coherent frame orderings) to permutation-invariant NVSthrough simple but effective architectural modifications, including per-frame latent encoding, query-centered camera conditioning and removalof temporal positional embeddings. Our results suggest that video modelscan be easily modified to “forget” about time with minimal supervision,producing state-of-the-art performance on sparse-view NVS benchmarks.