ECCV 2026 Papers

Recent advances in 3D Gaussian Splatting have demonstratedunprecedented success in novel view synthesis. However, the substantialinference and storage overhead driven by high-order Spherical Harmon-ics (SH) are primary bottlenecks for mobile platforms. In this paper,we present Flux-GS, a real-time Gaussian Splatting method designedto achieve high-fidelity rendering with significantly reduced overhead forresource-constrained mobile platforms. We first propose a Monte CarloSpecular Energy Aggregator, sampling third-order radiance residualsand aggregating specular energy into a compact latent space. In thisway, our method effectively preserves visually salient lighting features inlower-order bands without expensive distillation or pre-training. To miti-gate the high-frequency details lost during compression, we introduce anAttribute-Conditioned SH Enhancement module. This module predictsGaussian-aware offsets based on intrinsic Gaussian attributes, which en-hance the first-order SH representation prior to inference, without extrainference costs. Furthermore, the original single-view gradient-based den-sification is prone to producing excessive Gaussians and overfitting to aa. Rotation Opacity Opacity Opacity b.Scale 7% 2% 4% 6%5%PositionRotation SH5%Rotation 35%18%23% 3DGS Third-orderScaleSH SH Rest52%13%SH ScalePosition Position81% 18%13% 18%61%↓ 26%↓ First-orderThird-order SH First-order SH Flux-GS Ground Truth Flux-GS SH RestFig. 2: Gaussian parameter distribution and Spherical Harmonic fidelityanalysis. a. Per Gaussian memory footprint across 3DGS variants. Flux-GS achievessignificant compression (61% and 26% reductions) by optimizing Spherical Harmonics(SH) coefficients and decoupling SH into the base and view-independent components.b. Qualitative comparison demonstrates that Flux-GS with only first-order SH canrender high-fidelity high-frequency details comparable to 3DGS.certain view. We address these limitations by proposing a Multi-viewAlpha-based Densification and Pruning strategy. By leveraging multi-view guidance, we ensure multi-view structure consistency and the pre-cise removal of redundant primitives. Extensive experiments demonstratethat Flux-GS achieves substantial parameter reduction while maintain-ing competitive visual quality, offering a robust and scalable solution forreal-time mobile rendering. Code: https://xiaobiaodu.github.io/flux-gs-project/.

Domain Adaptive Object Detection (DAOD) adapts detec-tors from labeled source domains to unlabeled target domains. Currentteacher-student methods rely solely on proposal-level pseudo-boxes assupervision, which suffer from substantial noise under domain shift. Wemake an intriguing empirical observation: image-level predictions, ob-tained by aggregating proposal scores, achieve significantly higher ac-curacy than proposal-level pseudo-boxes. This superiority stems fromaggregation’s inherent noise reduction and robustness to localization er-rors. Despite this finding, image-level predictions have been completelyoverlooked as a supervision source in DAOD, though extensively usedin weakly supervised detection. We propose Multi-Granularity Pseudo-Labeling (MGPL), which for the first time incorporates image-level su-pervision alongside proposal-level pseudo-boxes in teacher-student DAOD.Our framework supervises the student detector at both granularities:proposal-level boxes provide spatial precision while image-level predic-tions offer reliable category information. We further introduce coarse-to-fine collaborative filtering, where image-level predictions guide class-adaptive threshold selection for proposals, reducing false negatives whilemaintaining precision. Extensive experiments demonstrate state-of-the-art performance across multiple benchmarks. For example, on PascalVOC → Clipart1k, we achieve 51.4% mAP, surpassing previous best andfully-supervised Oracle by 2.3% and 6.4% respectively. Our work revealsthe untapped potential of multi-granularity learning in domain adapta-tion. Code will be made publicly available.

Web agents—autonomous systems that navigate and executetasks on the web on behalf of users—have the potential to transformhow people interact with the digital world. However, the most capableweb agents today rely on proprietary models with undisclosed trainingdata and recipes, limiting scientific understanding, reproducibility, andcommunity-driven progress. We believe agents for the open web shouldbe built in the open. To this end, we introduce (1) MolmoWebMix, alarge and diverse mixture of browser task demonstrations and web-GUIperception data and (2) MolmoWeb a family of fully open multimodalweb agents. Specifically, MolmoWebMix combines over 100K synthetictask trajectories from multiple complementary generation pipelines with30K+ human demonstrations, atomic web-skill trajectories, and GUIperception data, including referring expression grounding and screen-shot question answering. MolmoWeb agents operate as instruction-conditioned visual-language action policies: given a task instruction anda webpage screenshot, they predict the next browser action, requiring noaccess to HTML, accessibility trees, or specialized APIs. Available in 4Band 8B size, on browser-use benchmarks like WebVoyager, Online-Mind2Web, and DeepShop, MolmoWeb agents achieve state-of-the-art results outperforming similar scale open-weight-only models suchas Fara-7B, and Holo1-7B. MolmoWeb-8B also surpasses set-of-marks(SoM) agents built on much larger closed frontier models like GPT-4o.We further demonstrate consistent gains through test-time scaling viaparallel rollouts with best-of-N selection, achieving 94.7% and 60.5%pass@4 (compared to 78.2% and 35.3% pass@1) on WebVoyager andOnline-Mind2Web respectively. We release model checkpoints, train-ing data, code, and a unified evaluation harness to enable reproducibilityand accelerate open research on web agents (GitHub).

Synthetic data mitigates the data scarcity problem in au-tonomous driving perception. However, the synthetic-to-real gap leads toperformance degradation, hindering real-world model generalization. Al-though current methods leverage diffusion models for photorealistic styletransfer to bridge this gap, they critically ignore a practical asymmetry:while synthetic data possesses perfect pixel-level annotations, real-worldstyle reference images generally lack corresponding labels. Consequently,existing methods relying on symmetric semantic guidance suffer fromeither prohibitive annotation costs or severe semantic misalignment. Toaddress this dilemma, we formally propose a novel task: AsymmetricStyle Transfer for Autonomous Driving (ASTAD), which requires se-mantically consistent transfer using only labeled synthetic content andunlabeled real-world references. We further introduce the ASTModel, atraining-free two-stage framework designed to bridge this domain gapunder asymmetric constraints. ASTModel first extracts a coarse seman-tic prior from the unlabeled target, followed by dynamic prior refinementand class-consistent style injection during the denoising process. Exten-sive experiments demonstrate that ASTModel significantly outperformsexisting methods in downstream perception utility and structural fidelity,while offering a 3.2× inference speedup. This work aligns synthetic-to-real adaptation with practical constraints, holding the potential to ac-celerate the scalable deployment of robust autonomous driving systems.Code: https://github.com/Dingyi-Yao/ASTAD.

Learned image compression (LIC) is bottlenecked by theneed to store independent models for each rate-distortion operating point.Existing variable bit-rate (VBR) methods aim to reduce this overhead viadense parameter modulation, but forcing a shared backbone to approx-imate divergent mappings causes severe feature entanglement. Specifi-cally, low-rate smoothing gradients inherently conflict with the preser-vation of high-frequency textural details, leading to sub-optimal perfor-mance. To resolve this, we propose MixCompress, a unified VBR frame-work based on sparse structural specialization. While sparsely gatedMixture-of-Experts (MoE) routing successfully mitigates gradient con-flict, it operates on a fixed computational budget. To address the in-creased representational demands of higher bit-rates we introduce a Mixture-of-Depths (MoD) extension to dynamically scale model capacity. Com-bined with Conditional Auxiliary Transforms (CAT) for dynamic sub-band energy modulation, our hierarchical framework effectively dynam-ically scales capacity. Extensive evaluations demonstrate that MixCom-press not only matches individually optimized single-rate baselines butcan even surpass them, establishing a new Pareto frontier for computa-tionally efficient image coding.

Metric reasoning is a critical and challenging task for VisionLanguage Models (VLMs), playing a pivotal role in embodied AI taskssuch as robotic manipulation and autonomous navigation. However, cur-rent spatial reasoning remains bottlenecked by rigid pixel-level supervi-sion; such localized optimization often compromises general multimodalintelligence, triggering performance degradation or catastrophic forget-ting of broad reasoning capabilities. To address these limitations, weintroduce Metric-Bench, a focused benchmark designed to guide metric-spatial reasoning using contextual information. By incorporating in-imagereference objects with known physical dimensions, Metric-Bench guidesmodels to implicitly learn the 2D-to-3D mapping without camera in-trinsics. We further present MetricReasoner, a task-adapted reinforce-ment fine-tuning recipe for reference-grounded metric reasoning, usingstructured prompts and verifiable numerical rewards. Extensive experi-ments on Metric-Bench demonstrate that our approach significantly en-hances spatial metric understanding, outperforming existing and evenlarger proprietary models by 43.1%, while improving downstream em-bodied performance over a spatial-specialized counterpart by 30.4% onRoboSpatial overall accuracy and 9.3% on ERQA, and additionally deliv-ering consistent gains on general benchmarks (15.9% on V⋆Bench, 88.9%on BLINK), indicating that the proposed adaptation does not necessarilycompromise general VLM capabilities.

Medical report understanding from real-world document im-ages is essential for generating patient-facing explanations and enablingstructured information exchange in clinical systems.Existing VLMs and LLMs have shown strong performance on documentunderstanding, but structured understanding of medical reports remainsinsufficiently benchmarked. Therefore, we introduce MedRepBench, abenchmark with 1,925 de-identified Chinese medical report images span-ning diverse departments, patient demographics, and acquisition formats.In MedRepBench, we mainly focus on report-grounded interpretationrather than evaluating diagnostic reasoning, treatment recommendation,or the integration of patient history. The interpretation is defined asstructured extraction of report fields (e.g., item, value, unit, referencerange, abnormal flag) plus a patient-facing explanation grounded strictlyin the report content.The benchmark primarily evaluates end-to-end VLMs, and also includesa controlled text-only setting (high-quality OCR + LLM) to approximatean upper bound when character recognition errors are minimized. Ourevaluation framework provides two complementary protocols: (1) an ob-jective protocol measuring field-level recall of structured items, and (2)an automated subjective protocol that uses an LLM-based judge to scorefactuality, interpretability, and reasoning quality under a fixed prompt.Using the objective metric as a reward signal, we also provide a lightweightGRPO-based alignment baseline for a mid-sized VLM, which improvesfield-level recall by up to 6%.Finally, we analyze practical limitations of OCR+LLM pipelines, includ-ing layout-related errors and additional system latency, showing the needfor robust end-to-end vision-based medical report understanding. Thedataset and evaluation resources are publicly available on HuggingFace.

Human-object interaction (HOI) video generation has gar-nered increasing attention due to its promising applications in digitalhumans, e-commerce, advertising, and robotics imitation learning. How-ever, existing methods face two critical limitations: (1) a lack of effec-tive mechanisms to inject multi-view information of the object into themodel, leading to poor cross-view consistency, and (2) heavy relianceon fine-grained annotations including hand mesh and body templatesfor modeling interaction occlusions, due to limited training data. To ad-dress these challenges, we introduce ByteLOOM, a Diffusion Transformer(DiT)-based framework that generates realistic human-object interactionvideo with geometrically consistent object illustration, using simplifiedhuman conditioning and 3D object inputs. We first propose an RCM-cache mechanism that leverages Relative Coordinate Maps (RCM) as auniversal representation to maintain object’s geometry consistency andprecisely control 6-DoF object transformations in the meantime. To com-pensate HOI dataset scarcity and leverage existing datasets, we furtherdesign a training curriculum that enhances model capabilities in a pro-gressive style and relaxes the demand of hand mesh. Extensive experi-ments demonstrate that our method faithfully preserves human identityand the object’s multi-view geometry, while maintaining smooth motionand object manipulation. Project page: neutrinoliu.github.io/byteloom/

Autoregressive models have emerged as an effective paradigm for point cloud generation. However, most existing approaches rely on heuristic tokenization strategies, such as spatial sorting or stochastic downsampling, which often disrupt intrinsic point cloud topology and weaken the structural coherence of the generated shapes. In this paper, we present PointRSP, an autoregressive framework that reformulates point cloud generation as a topology-preserving tessellation process via recursive spectral partitioning. Instead of constructing token sequences heuristically, we introduce a topology-aware partitioning autoencoder that decomposes an unstructured point cloud into a nonbalanced binary tree through a hybrid recursive spectral partitioning strategy. This hierarchical representation provides a deterministic geometric blueprint that preserves topological relationships while capturing multiscale structural dependencies within a quantized latent space. To synthesize shapes in this space, we propose a dual-stream cascaded generator that jointly models structural evolution and feature synthesis. In addition, we design a geometry-calibrated positional encoding mechanism that anchors latent embeddings using multi-scale structural centers, which stabilizes cascaded generation during the early stages of structural formation. Extensive experiments show that PointRSP achieves state-ofthe-art performance in generation quality and diversity, demonstrating strong generalization across complex 3D topologies.

In this paper, we focus on motion discrete tokenization, whichconverts raw motion into compact discrete tokens—a process proven cru-cial for efficient motion generation. In this paradigm, increasing the num-ber of tokens is a common approach to improving motion reconstructionquality, but more tokens make it more difficult for generative models tolearn. To maintain high reconstruction quality while reducing genera-tion complexity, we introduce Language-Guided Tokenization (LG-Tok)for efficient motion tokenization. LG-Tok aligns natural language withmotion at the tokenization stage, yielding compact, high-level seman-tic representations. This approach not only strengthens both tokeniza-tion and detokenization but also simplifies the learning of generativemodels. Furthermore, existing tokenizers predominantly adopt convo-lutional architectures, whose local receptive fields struggle to supportglobal language guidance. To this end, we propose a Transformer-basedTokenizer that leverages attention mechanisms to enable effective align-ment. Additionally, we design a language-drop scheme, in which lan-guage conditions are randomly removed during training. This schemeprevents shortcut learning over text and enables the detokenizer to sup-port language-free guidance. On three generation benchmarks, LG-Tokoutperforms state-of-the-art methods (e.g., achieving an FID score of0.057 vs. MARDM’s 0.114 on HumanML3D). LG-Tok-mini uses onlyhalf the tokens while maintaining competitive performance, validatingthe efficiency of our semantic representations. Code and checkpoints areavailable at https://eanson023.github.io/LG-Tok/.

3D scene graphs (3DSGs) provide a compact and structuredabstraction of 3D environments. Although advances in foundation modelshave enabled open-vocabulary 3DSG generation, existing approaches re-main object-centric and encode limited relational information—restrictingtheir applicability in real-world scenarios that require fine-grained under-standing. We propose OP3DSG, an open-vocabulary part-aware 3DSGgeneration framework that constructs unified graphs that jointly modelobjects, interactive parts, spatial relations, functional relations, and af-fordances. OP3DSG integrates object-part knowledge-guided detectionwith part-aware 3D fusion to preserve small and interaction-relevantcomponents, and employs a geometry-initialized prior graph with LLM-based refinement to reduce spurious relational predictions while enablingefficient graph construction. To systematically evaluate unified 3D scenegraph construction, we introduce UniGraph3D, a benchmark designedfor part-aware perception and multi-level relational reasoning. Experi-mental results show that OP3DSG achieves state-of-the-art performanceand demonstrates its effectiveness as a perception backbone in diversereal-world robotics tasks.

Visual markups such as highlights, underlines, and bold textare common in table-centric documents. Although multimodal large lan-guage models (MLLMs) have made substantial progress in document un-derstanding, their ability to treat such cues as explicit logical directivesremains under-explored. More importantly, existing evaluations cannotdistinguish whether a model fails to see the markup or fails to reason withit. This creates a key blind spot in assessing markup-conditioned behaviorover tables. To address this gap, we introduce HighlightBench, a diagnos-tic benchmark for markup-driven table understanding that decomposesevaluation into five task families: Markup Grounding, Constrained Re-trieval, Local Relations, Aggregation & Comparison, and Consistency &Missingness. We further provide a reference pipeline that makes interme-diate decisions explicit, enabling reproducible baselines and finer-grainedattribution of errors along the perception-to-execution chain. Experi-ments show that even strong models remain unstable when visual cuesmust be consistently aligned with symbolic reasoning under structuredoutput constraints.

Domain Generalized Semantic Segmentation (DGSS) is achallenging task, as vision models often rely on low-level appearancecues that change across domains. In contrast, structural attributes ex-hibit cross-domain stability, motivating the use of structural priors forDGSS. Existing methods use prompt learning to transfer such priorsinto DGSS models, but typically encode each class as a single holisticprompt. Moreover, these methods apply prompts uniformly to all pixels,offering no mechanism to adapt when only a subset of object regions isvisible due to viewpoint changes, occlusion, and environmental variation.We address this with Spatial Hierarchical Prompts (SHP) that en-rich each class with region-level geometric anchors capturing structuralappearance from distinct viewing angles, ensuring complementary cov-erage under arbitrary viewpoints. Additionally, we propose the Hier-archical Prompt Injector (HPI), which enables spatially adaptiveprompt injection in foundation models. HPI spatially grounds promptsby modeling their semantic relevance and spatial influence with visualfeatures. Considering the difficulty of learning spatially and semanti-cally aware prompt injection, we further introduce auxiliary supervi-sion to align hierarchical prompts with their corresponding object re-gions. We achieve 70.62% and 72.74% mIoU on synthetic-to-real andreal-to-real benchmarks, respectively. Code and checkpoints are releasedat https://github.com/MosukFate/HPI.

The continuous emergence of high-quality video deepfakesrequires detectors that continually adapt to new forgery patterns, yetexisting approaches, which are designed for deepfake images, fail to cap-ture video-specific cues. Unlike deepfake images that contain only spatialartifacts, deepfake videos leave distinct evidence along both spatial andtemporal axes, necessitating the separate preservation of each modalityduring sequential model updates. To overcome this limitation, we intro-duce a continual deepfake video detection framework, Modality-SpecificFrequency Distillation (MSFD), that explicitly decomposes video fea-tures into spatial, temporal, and spatiotemporal modalities in the fre-quency domain. This decomposition enables independent preservationof each modality, as different deepfake video types exhibit varying re-liance on spatial and temporal cues across tasks. Furthermore, MSFDadopts a cross-modality decorrelation loss that encourages spatiotempo-ral representations to remain orthogonal to single-modality cues. Exten-sive experiments show that our framework achieves stronger adaptationand preserves performance more effectively than state-of-the-art methodsacross diverse continual deepfake video scenarios.

3D Gaussian Splatting has recently revolutionised novel viewsynthesis as well as many other 3D vision methods and applications.Drawing inspiration from this representation, we now rethink texturesto overcome the main issues of UV mapping while considerably lower-ing their memory footprint. Heat Kernel Textures (HKTex) eliminateUV unwrapping as well as their persistent issues of wasted UV space,seams, distortions, vertex-duplication, and varying resolution. Groundedin discrete Riemannian geometry and intrinsically defined on any man-ifold surface discretised as a triangular mesh, HKTex uses anisotropicheat kernels as geodesic equivalents to Gaussians. Like our kernels, alsothe optimisation of their position and the adaptive densification strate-gies were redefined to operate on the surface of the object to be tex-tureised. Our novel representation is also fully integrated with a physi-cally based renderer and can be optimised either from existing texturesor multi-view images. Our project page and code are available at circle-group.github.io/research/HeatKernelTextures.

Training embodied AI agents depends critically on the visual fidelity of simulation environments and the ability to model dynamic humans. Current simulators predominantly rely on mesh-based rasterization, for which photorealistic assets are costly to author at scale, and their support for dynamic human avatars is largely constrained to mesh representations, hindering agent generalization to human-populated realworld scenarios. We present Habitat-GS, a navigation-centric embodied AI simulator extended from Habitat-Sim that integrates 3D Gaussian Splatting scene rendering and drivable gaussian avatars while maintaining full compatibility with the Habitat ecosystem. Our system implements a 3DGS renderer for real-time photorealistic rendering and supports scalable 3DGS asset import from diverse sources. For dynamic human modeling, we introduce a gaussian avatar module that enables each avatar to simultaneously serve as a photorealistic visual entity and an effective navigation obstacle, allowing agents to learn human-aware behaviors in realistic settings. Experiments on point-goal navigation demonstrate that agents trained on 3DGS scenes achieve stronger cross-domain generalization. Evaluations on avatar-aware navigation further confirm that gaussian avatars enable effective human-aware navigation, while performance benchmarks validate the system’s scalability. Code is available at https://github.com/zju3dv/habitat-gs.

Compositional Zero-Shot Learning (CZSL) requires learningto compose primitive concepts from seen data and generalize to unseenattribute–object pairs. Existing methods typically disentangle visual fea-tures into isolated primitive components and align them with their cor-responding text concepts or full compositions. However, they often suf-fer from ambiguous vision–language alignment and lack mechanisms toexplicitly capture cross-primitive semantic interactions, ultimately hin-dering generalization to unseen compositions. To overcome these chal-lenges, we introduce Prototype-Conditioned Imagination (PCI), a two-stage framework. In the first stage, Bidirectional Alignment and Disen-tanglement (BAD) couples cross-modal attention with entropic optimal-transport regularization to effectively align visual and textual represen-tations while enforcing global disentanglement of attribute and objectprimitives, thereby mitigating ambiguous vision-language matching. Inthe second stage, Prototype-driven Compositional Modulation (PCM)explicitly models how attribute semantics modulate object features tosynthesize compositional representations, making cross-primitive rela-tions explicit and substantially improving recognition of unseen compo-sitions. Evaluated under both closed-world and open-world protocols onfour popular benchmarks, our approach achieves state-of-the-art perfor-mance, demonstrating superior generalization capabilities to unseen com-positions. Our code is available at https://github.com/YFan-Z1/PCI.

Streaming 3D reconstruction aims to recover 3D information, such as camera poses and point clouds, from a video stream, which necessitates geometric accuracy, temporal consistency, and computational efficiency. Motivated by the principles of Simultaneous Localization and Mapping (SLAM), we introduce GCT, geometric context transformer, a feed-forward 3D foundation model for reconstructing scenes from streaming data. A defining aspect of GCT lies in its carefully designed attention mechanism, which integrates an anchor context, a pose-reference window, and a trajectory memory to address coordinate grounding, dense geometric cues, and long-range drift reduction, respectively. This design keeps the streaming state compact while retaining rich geometric context, enabling stable efficient inference at around 20 FPS on 518×378 resolution inputs over long sequences exceeding 10,000 frames. Extensive evaluations across a variety of benchmarks demonstrate that our approach achieves superior performance compared to both existing streaming and iterative optimization-based approaches.

Instant reconstruction of dynamic humans and 3D human-object interactions from uncalibrated sparse-view videos is critical fornumerous downstream applications. Existing methods, however, are ei-ther limited by the slow reconstruction speeds or incapable of generat-ing novel-time representations. To address these challenges, we proposeForge4D, a feed-forward 4D human-object reconstruction and interpo-lation model that efficiently reconstructs temporally aligned represen-tations from uncalibrated sparse-view videos, enabling both novel-viewand novel-time synthesis. Our model simplifies the 4D reconstructionand interpolation problem as a joint task of streaming 3D Gaussian re-construction and dense motion prediction. For the task of streaming 3DGaussian reconstruction, we introduce learnable state tokens to enforcetemporal consistency in a memory-friendly manner. For novel-time syn-thesis, we design a novel motion prediction module to predict dense mo-tions for each 3D Gaussian between two adjacent frames. To overcomethe lack of the ground truth for dense motion supervision, we formulatedense motion prediction as a dense point matching task and introducea self-supervised retargeting loss to optimize this module. An additionalocclusion-aware optical flow loss is introduced to ensure motion consis-tency with plausible human and object movement, providing strongerregularization. Extensive experiments demonstrate the effectiveness ofour model on both in-domain and out-of-domain datasets. Code andmodels will be made publicly available.