ECCV 2026 Papers

Fine-grained robotic manipulation depends on understand-ing parts, not only whole objects. Existing 3D foundation models tendto be either generalized but object-aware, or part-aware but limited toclosed-set taxonomies, which weakens zero-shot transfer. We study text-conditioned 3D part segmentation, where a free-form phrase selects afunctional part on point cloud. We introduce UniPart, a feed-forwardcross-modal 3D Transformer that conditions CLIP text embedding. Toscale supervision, we build LangPart-1M with 160K+ Objaverse assetsand 8M text to part pairs using multi-view consistent part generation.We further manually label a high-quality subset, LangPart-4K, for fine-tuning and evaluation. UniPart achieves strong zero-shot results on open-vocabulary part benchmarks and transfers to language-conditioned partgrasping in real world.

Diffusion models have achieved remarkable success across di-verse domains, with performance closely related to the denoising back-bones that parameterize the score function. In this paper, we presenta systematic, phase-aware analysis of diffusion components and showthat abrupt, early-stage fluctuations in deep latents are strongly asso-ciated with artifacts. Guided by these findings, we introduce DUNE(Diffusion Unified Network refiNEr), a training-free refinement frame-work that detects abrupt deviations in deep low-noise internal latentsusing a shared EMA-based criterion, and applies backbone-specific sup-pression to the detector-selected entries. Although derived from U-Net,the same detect–suppress principle extends naturally to Transformer-based diffusion models by acting on the latents of deep self-attentionblocks. Extensive experiments across multiple backbones indicate thatDUNE improves fidelity while reducing hallucinations, offering new in-sight into where and when diffusion backbones should be controlled.

Though Gaussian splatting has achieved impressive resultsin novel view synthesis, it requires millions of primitives to representhighly textured scenes — even when the underlying geometry is simple.We propose a representation that decouples geometry and appearanceto achieve a compact and computationally efficient representation. Ourmethod uses surfels for geometry and combines a global neural field withper-primitive colors for appearance. While existing approaches that tex-ture primitives incur significant overhead in either memory or renderingspeed, our method strikes a better balance by leveraging the neural fieldto implicitly compress texture information. We bound the computationsrequired in the rendering pass by querying the neural field for a fixednumber of primitives per pixel, enabling real-time rendering speeds. Ourrepresentation matches the perceptual quality of 3D Gaussian splattingwhile using 9.7× fewer primitives and 5.5× less memory on outdoorscenes, and 31× fewer primitives and 3.7× less memory on indoor scenes.

Large Vision-Language Models (LVLMs) remain bottleneckedby massive computational footprints, precluding their deployment onresource-constrained edge devices. While efforts to compress LVLMs fo-cus heavily on vision token reduction or smaller language models, thevision encoder is largely overlooked, typically deployed as a monolithic,computationally heavy feature extractor. Moreover, there is no previouseffort that designs a vision encoder for LVLMs directly optimized foron-device latency. In this paper, we present UltraViT, a vision encoderfor LVLMs, explicitly designed and optimized for on-device performance.Specifically, by taking into account real on-device latencies, we system-atically design a pyramidal architecture that strategically integrates andadapts heterogeneous spatial mixers at the macro-block level. Further-more, to pre-train UltraViT, we propose a novel two-stage generativepre-training strategy: cultivating rich spatial features via dense distil-lation, followed by direct generative supervision from a capacity-mixedfrozen LLM. Compared to standard contrastive and SSL, we show thatour pre-training is much more effective for achieving high-level seman-tic grounding for UltraViT needed for the subsequent generative multi-modal alignment of LVLM training. Extensive experiments demonstratethat our on-device latency-informed design combined with our tailoredtraining strategy establishes a new state-of-the-art for efficient LVLMencoding, significantly outperforming existing encoder-centric baselineswhile operating on-device at nearly 1.7× the speed.

Native 3D generative models have achieved remarkable _x001C_-delity and speed, yet they su_x001B_er from a critical limitation: inability toprescribe precise structural articulations, where precise structural con-trol within the native 3D space remains underexplored. This paper pro-poses SK-Adapter, a simple yet e_x001E_cient and e_x001B_ective framework thatunlocks precise skeletal manipulation for native 3D generation. Mov-ing beyond text or image prompts, which can be ambiguous for precisestructure, we treat the 3D skeleton as a _x001C_rst-class control signal. SK-Adapter is a lightweight structural adapter network that encodes jointcoordinates and topology into learnable tokens, which are injected intothe frozen 3D generation backbone via cross-attention. This design al-lows the model to not only e_x001B_ectively _x0010_attend_x0011_ to speci_x001C_c 3D structuralconstraints but also preserve its original generative priors. To bridgethe data gap, we contribute the Objaverse-TMS dataset, a large-scale⋆ „Equal contribution. Equal advising.dataset of 24k text-mesh-skeleton pairs. Extensive experiments con_x001C_rmthat our method achieves robust structural control while preserving thegeometry and texture quality of the foundation model, signi_x001C_cantly out-performing existing baselines. Furthermore, we extend this capability tolocal 3D editing, enabling region-speci_x001C_c editing of existing assets withskeletal guidance, which is unattainable by previous methods. Projectpage: https://sk-adapter.github.io/

Current sign language translation (SLT) systems attempt tolearn all aspects of signing—manual gestures, high-speed fingerspelling,and asynchronous non-manual facial cues—within a single end-to-endnetwork. Learning multiple tasks without detailed supervision leads topoor recognition of fingerspelled proper nouns and technical terms, andleaves rich disambiguating information from lip movements largely unex-ploited. We introduce SignBind-LLM, a modular framework that ad-dresses these limitations through three dedicated expert streams: one forcontinuous signing, one for fingerspelling, and one for lipreading. Eachexpert is pre-trained independently using CTC on approximately twomillion automatically generated pseudo-gloss sequences, removing theneed for manual gloss annotation. A lightweight transformer with learnedtemporal alignment fuses the expert outputs, and a pre-trained languagemodel translates the resulting pseudo-gloss sequences into fluent spokenEnglish. At matched decoder scale (250M parameters), our architecturealready surpasses all prior methods, confirming that the gains are archi-tectural rather than a consequence of scaling the language model. Scal-ing to a larger decoder sets a new state-of-the-art across How2Sign: 23.1,BOBSL: 7.0, and ChicagoFSWild+: 73.6%, while requiring significantlylower training cost than prior approaches.

Cross-View Geo-Localization (CVGL) is crucial for naviga-tion in GPS-restricted environments. However, bridging the severe per-spective gap between ground and aerial views remains challenging. Exist-ing approaches either rely on coarse image-level retrieval, suffer from un-stable optimization in pose regression, or require expensive fine-grainedannotations. These limitations suggest that current formulations of cross-view geo-localization are insufficient for reliable fine-grained localization.To address this limitation, we introduce a new task, Sector-Level Cross-View Geo-Localization (SLCVGL), which aims to perform fine-grainedsector-level geo-localization of a limited Field-of-View (FoV) ground im-age within an omnidirectional aerial image, while simultaneously infer-ring its ground-view orientation without manually annotated pose labels.To operationalize this task, we propose an Azimuthal Scanning frame-work. It decomposes aerial feature maps into multiple azimuthal sectors,instead of relying on holistic matching. By aligning ground features withthese sectors, the model explicitly retrieves a sector-level sub-region andimplicitly estimates orientation, eliminating the need for continuous poseregression supervision. This design also reduces interference from unob-served overhead regions. To benchmark the proposed task, we repur-pose standard panoramic datasets into a limited-FoV sector-level proto-col with random FoV cropping and soft labeling. Extensive experimentson CVUSA and CVACT demonstrate improvements over state-of-the-art methods, with a 30.17% R@1 improvement on CVACT under the 70°FoV setting, while maintaining strong robustness across varying FoVs.

Spatio-Temporal Video Grounding (STVG) aims to retrievethe visual trajectory of a specific object from a video stream as describedby a natural language expression. However, most advanced methodsstruggle to balance global context modeling with precise boundary lo-calization. Due to the prohibitive computational costs of processing longvideos, these approaches typically resort to low-rate temporal downsam-pling and implicit motion modeling. This inevitably suppresses high-frequency boundary cues and neglects the explicit inter-frame depen-dencies required for precise boundary delineation. To address these lim-itations, we present ScanFocus, a novel coarse-to-fine framework thatdecouples the STVG task into a global spatio-temporal scan and a localboundary focus. Specifically, we utilize a unified vision-language fusionencoder combined with a lightweight Deformable Semantic-Motion Fu-sion module to efficiently align multimodal features and generate coarseproposals. To recover the suppressed fine-grained details, we introducethe Semantic-Guided Temporal Aggregator (SGTA) in the refinementstage. By densely sampling around coarse boundaries, SGTA explicitlymodels short-term temporal interactions under semantic guidance, cap-turing rapid motion changes for precise timestamp regression. Extensiveexperiments on three widely used benchmarks demonstrate the perfor-mance superiority of our proposed method over previous approaches.Code will be released at https://github.com/TenMinutes209/ScanFocus.

The global rose cultivation industry has experienced continued expansion in recent years. There has been a set of value evaluation criteria covering multiple dimensions for roses. However, the evaluations are mainly conducted by human experts after harvesting, which is time-consuming and labor-intensive. 3D computer vision technologies are promising to automate the process before harvesting, but the lack of large-scale datasets with localization annotations that capture plant architecture hinders the progress. To bridge this gap, we propose Rosetum3D, which is a large-scale 3D vision dataset for preharvest roses. The dataset is constructed via occlusion-robust multi-view RGB-D capture protocols in commercial greenhouses. We provide fine-grained 2D localization labels using bounding boxes and botanically defined keypoints, and then obtain the 3D structures recovered through depth backprojection. Rosetum3D contains 21,114 images and 46,848 annotated rose objects. Models trained on Rosetum3D have achieved 2D/3D rose localization, which is a crucial step for automated preharvest quality grading and growth monitoring. Beyond localization, Rosetum3D serves as a benchmark for agricultural vision tasks, including 2D rose object detection, local feature matching, depth estimation, and instance reidentification. By enabling data-driven precision agriculture, Rosetum3D paves the way for robotic harvesting systems and AI-driven yield prediction in protected cultivation. The dataset are available at https: //huggingface.co/datasets/WaterMelon2333/Rosetum3D/tree/main.

Recent advances in text-to-image (T2I) models have revolu-tionized the field of image generation and editing. However, identifyingsemantics that a T2I model can successfully edit in an image contin-ues to be a challenging task. Most existing approaches require usersto manually specify semantics to modify a particular image, a time-consuming process that often involves extensive trial and error. In thispaper, we present RankT2I, a novel, training-free, and model-agnosticframework that automates the discovery of editable semantics in diffu-sion and FLUX-based models. Given a visual domain, we first utilizea multimodal vision-language model to gather a broad set of candidatesemantics. We then frame semantic discovery as a set selection problemand use a submodular objective to identify semantics that are relevant,editable, and diverse. Our method helps users efficiently identify a widerange of semantics for text-to-image editing models across several do-mains while outperforming existing methods.

Recent approaches in controllable novel view video generationoften rely on fine-tuning pre-trained Video Di!usion Models (VDMs).This dominant paradigm is computationally expensive and frequentlysu!ers from catastrophic forgetting of the model’s original generative pri-ors. To address this challenge, here we propose InverseCrafter, a VDMtraining-free framework that reformulates novel view video generationas an inpainting-based inverse problem in the latent space, eliminatingthe need for any annotated 4D training data. The core of our methodis to establish operator equivalence by employing a lightweight latentmask encoder to define a latent-domain masking operation via a con-tinuous, multi-channel representation. This principled representationfaithfully models the forward process in the latent domain, enablinge"cient, backpropagation-free solvers while bypassing the costly bottle-neck of repeated VAE operations. InverseCrafter achieves high-fidelity,spatio-temporally coherent novel view synthesis with near-zero additionalinference overhead and excels at general-purpose video inpainting andediting by fully preserving the pre-trained VDM’s generative capabilities.

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.