ECCV 2026 Papers

Semi-supervised cross-domain few-shot object detection (SemiCDFSOD) is a practical yet underexplored task that requires adapting detectors to novel domains using scarce support examples alongside abundant unlabeled data. While pseudo-labeling via open-set detectors offers a direct pathway to exploit unlabeled data, it is fundamentally impaired by two intrinsic defects: semantic misclassification, where localized objects are assigned incorrect categories or suppressed as background, and low recall, where the majority of objects are missed entirely. To systematically address these limitations, we propose a training-free pseudo-label calibration framework for Semi-CDFSOD. First, we construct robust class prototypes from the scarce labeled supports and high-confidence pseudo-labels to rectify semantic misclassifications, dynamically recalibrating noisy category assignments while preserving welllocalized detections. Second, we introduce a visual self-prompting mechanism that harnesses high-confidence pseudo-labels as prompts for the Segment Anything Model (SAM), generating supplementary proposals to recover undetected objects and dramatically enhance recall. The harmonized pseudo-labels, together with the support samples, are then employed to adapt the final detector. We establish the first comprehensive benchmark for Semi-CDFSOD, spanning six diverse domains. Extensive experiments demonstrate that our method consistently outperforms existing approaches. The code will be released at https://github.com/ xSheep123/Semi-CDFSOD.git.

Despite the remarkable capabilities of Multimodal Large Lan-guage Models (MLLMs), they still suffer from visual fading in long-context scenarios. Specifically, the attention to visual tokens diminishesas the text sequence lengthens, leading to text generation detached fromvisual constraints. We attribute this degradation to the inherent induc-tive bias of Multimodal RoPE, which penalizes inter-modal attention asthe distance between visual and text tokens increases. To address this, wepropose inter-modal Distance Invariant Position Encoding (DIPE), asimple but effective mechanism that disentangles position encoding basedon modality interactions. DIPE retains the natural relative positioningfor intra-modal interactions to preserve local structure, while enforc-ing an anchored perceptual proximity for inter-modal interactions. Thisstrategy effectively mitigates the inter-modal distance-based penalty, en-suring that visual signals remain perceptually consistent regardless ofthe context length. Experimental results demonstrate that by integrat-ing DIPE with Multimodal RoPE, the model maintains stable visualgrounding in long-context scenarios, significantly alleviating visual fad-ing while preserving performance on standard short-context benchmarks.Code is available at https://github.com/lchen1019/DIPE.

We present ActionPlan, a unified motion diffusion frame-work that bridges real-time streaming with high-quality offline genera-tion within a single model. The core idea is to introduce a per-frameaction plan: the model predicts frame-level text latents that act as densesemantic anchors throughout denoising, and uses them to denoise thefull motion sequence with combined semantic and motion cues. To sup-port this structured workflow, we design latent-specific diffusion steps,allowing each motion latent to be denoised independently and sampledin flexible orders at inference. As a result, ActionPlan can run in ahistory-conditioned, future-aware mode for real-time streaming, whilealso supporting high-quality offline generation. The same mechanism fur-ther enables zero-shot motion editing and in-betweening without addi-tional models. Experiments demonstrate that our real-time streaming is5.25× faster while also achieving 18% motion quality improvement overthe best previous method in terms of FID.∗ †Equal contribution. Corresponding author

In fluorescence microscopy, spectral unmixing aims to re-cover individual fluorophore concentrations from spectral images thatcapture mixed fluorophore emissions. Since classical methods operatepixel-wise and rely on least-squares fitting, their performance degradeswith increasingly overlapping emission spectra and higher levels of noise,suggesting that a data-driven approach that can learn and utilize a struc-tural prior might lead to improved results. Learning-based approachesfor spectral imaging do exist, but they are either not optimized for mi-croscopy data or are developed for very specific cases that are not ap-plicable to fluorescence microscopy settings. To address this, we pro-pose λSplit, a physics-informed deep generative model that learns aconditional distribution over concentration maps using a hierarchicalVariational Autoencoder. A fully differentiable Spectral Mixer enforcesconsistency with the image formation process, while the learned struc-tural priors enable state-of-the-art unmixing and implicit noise removal.We demonstrate λSplit on 3 real-world datasets that we syntheticallycast into a total of 58 challenging spectral unmixing benchmarks. Wecompare our results against a total of 10 baseline methods, includingclassical methods and a range of learning-based methods. Our resultsconsistently show competitive performance and improved robustness inhigh noise regimes, when spectra overlap considerably, or when the spec-tral dimensionality is lowered, making λSplit a new state-of-the-art forspectral unmixing of fluorescence microscopy data. Importantly, λSplitis compatible with spectral data produced by standard confocal micro-scopes, enabling immediate adoption without specialized hardware mod-ifications. The code and all datasets used in the experiments are availableat https://github.com/juglab/lambdaSplit.

A deep understanding of the physical world is essential forrobotic manipulation and physically realistic simulation. While currentmethods, including VLM-based and other learning-based approaches,have shown promise in physical property inference, their evaluation isoften hindered by the lack of physically grounded reference data. Toaddress this gap, we introduce XDen-1K, the first large-scale multi-modal dataset that provides physically grounded density field for real-world objects. XDen-1K comprises 1,000 real-world objects spanning137 categories, with comprehensive data for each object, including ahigh-resolution, carefully curated 3D geometric model with part-levelannotations and paired real-world biplanar X-ray scans. In addition,XDen-1K includes high-fidelity volumetric density field reconstructedfrom sparse biplanar X-ray views via a novel optimization framework.XDen-1K also provides a benchmark for density estimation and enablesX-ray-conditioned volumetric segmentation. Experiments further demon-strate that the center-of-mass prior derived by XDen-1K can improverobotic manipulation performance. By providing real-world X-ray scansand physics-consistent density field, XDen-1K establishes a foundationfor advancing physical property inference and embodied AI. Additionaldetails and dataset resources are available at https://xden-1k.github.io/ and https://huggingface.co/datasets/zhangjxx/XDen-1K.

The effectiveness of multimodal instruction tuning dependsnot only on dataset scale, but also critically on whether training samplesgenuinely require visual reasoning. However, existing instruction datasetsoften contain a substantial portion of visually redundant samples (solv-able from text alone), as well as multimodally misaligned supervisionthat can degrade learning. To address this, we propose VisNec (VisualNecessity Score), a principled data selection framework that measuresthe marginal contribution of visual input during instruction tuning. Bycomparing predictive loss with and without visual context, VisNec iden-tifies whether a training instance is vision-critical, redundant, or mis-aligned. To preserve task diversity, we combine VisNec with semanticclustering and select high-necessity samples within each cluster. Across10 downstream benchmarks, training on only 15% of the LLaVA-665Kdataset selected by VisNec achieves 100.2% of full-data performance. Onthe smaller Vision-Flan-186K dataset, our selection not only further re-duces data size but surpasses full-data training by 15.8%. These resultsdemonstrate that measuring and leveraging visual necessity provides aneffective solution for both efficient and robust multimodal instructiontuning. Project Page: https://dmk041218.github.io/VisNec/.

Multimodal large language models often generate reasoningchains containing subtle errors that lead to incorrect answers. Currentverification approaches have notable limitations. Existing approaches ei-ther require expensive labelled supervision with inconsistent cross-taskperformance or aggregate scores from multiple sources by simple ag-gregations, missing a key insight: when these scores disagree, that dis-agreement itself carries important information about whether a reasoningstep is truly valid or not. We formalise this as a coupled scoring prob-lem among disparate, frozen verifiers, interpretable as a coordinationgame with a unique closed-form equilibrium where agreement signalsvalid steps while disagreement reveals instability. Towards this end, wepropose a training-free domain-agnostic step-wise verification approachwe call VERDICT: VERification via Disagreement-Informed CoupledThresholding. To our knowledge, VERDICT is the first training-freeverifier that makes the structure of cross-modal disagreement explicit andactionable. It computes consensus scores through a closed-form solution,enabling both disagreement-aware filtering and stability-conscious rank-ing of reasoning steps. Evaluated across six benchmarks, VERDICTconsistently improves over the base model by up to +5.95%, and per-forms competitively with domain-specific critics that demand extensivesupervision, demonstrating that cross-modal agreement provides robustverification signals without task-specific adaptation.

In embodied intelligence, the embodiment gap between roboticand human hands brings significant challenges for learning from humandemonstrations. Although some studies have attempted to bridge thisgap using reinforcement learning, they remain confined to merely re-producing human manipulation, resulting in limited task performance.Moreover, current methods struggle to support diverse robotic hand con-figurations. In this paper, we propose UniBYD, a unified framework thatuses a dynamic reinforcement learning algorithm to discover manipula-tion policies aligned with the robot’s physical characteristics. To enableconsistent modeling across diverse robotic hand morphologies, UniBYDincorporates a unified morphological representation (UMR). Building onUMR, we design a dynamic PPO with an annealed reward schedule, en-abling reinforcement learning to transition from offline-informed imita-tion of human demonstrations to online-adaptive exploration of policiesbetter adapted to diverse robotic morphologies, thereby going beyondmere imitation of human hands. To address the severe state drift causedby the incapacity of early-stage policies, we design a hybrid Markov-based shadow engine that provides fine-grained guidance to anchor theimitation within the expert’s manifold. To evaluate UniBYD, we proposeUniManip, the first benchmark for cross-embodiment manipulation span-ning diverse robotic morphologies. Experiments demonstrate a 44.08%average improvement in success rate over the current state-of-the-art.Our project page is https://zhanheng-creator.github.io/UniBYD/.Mocap Data2-fingered3-fingeredUMRReinforcement learningShadow Engine5-fingeredDynamic PPOUniBYD Our results Imitation-based reinforcement learningFig. 1: Leveraging human demonstrations, UniBYD learns manipulation strategiesthat transcend mere imitation and are tailored to a broad spectrum of robotic handmorphologies.

Optimizing the noise samples of diffusion and flow models isan increasingly popular approach to align these models to target rewardsat inference time. However, we observe that these approaches are usuallyrestricted to differentiable or cheap reward models, the formulation ofthe underlying pre-trained generative model, or are memory/computeinefficient. We instead propose a simple trust-region based search algorithm(TRS) which treats the pre-trained generative and reward models as ablack-box and only optimizes the source noise. Our approach achieves agood balance between global exploration and local exploitation, and isversatile and easily adaptable to various generative settings and rewardmodels with minimal hyperparameter tuning. We evaluate TRS acrosstext-to-image, molecule and protein design tasks, and obtain significantlyimproved output samples over the base generative models and otherinference-time alignment approaches which optimize the source noisesample, or even the entire reverse-time sampling noise trajectories in thecase of diffusion models. Our source code is publicly available⋆ .

Multi-label classification models are commonly trained byassuming conditional independence among labels, typically through theuse of binary cross-entropy objectives. While this assumption simpli-fies optimization, it introduces a structural bias that ignores dependen-cies among labels and can adversely affect the calibration of predictedprobabilities. In this work, we theoretically analyze the impact of thisassumption and show that neglecting label dependencies leads to a mis-calibration effect that is linked to the conditional dependency of thelabels. Motivated by this observation, we propose Pairwise CorrelationDifference (PCD), a novel auxiliary loss designed to incorporate labeldependency information during training. PCD aligns the pairwise cor-relations of model logits with those of the ground-truth labels, therebyencouraging the network to capture conditional dependencies among la-bels. We combine PCD with the standard binary cross-entropy loss toform the Correlated Multi-Label Loss (CMLL), which serves as a objec-tive for dependency-aware training. We provide theoretical justificationfor the functional form of the proposed loss and its connection to thedependency structure of the label distribution. Extensive experimentson three benchmark multi-label datasets demonstrate that CMLL con-sistently produces better-calibrated predictions while maintaining com-petitive classification accuracy compared to existing multi-label learningobjectives. Additionally, we show that the proposed approach exhibitsrobustness to the choice of binning schemes used in calibration evalua-tion.

Medical Large Vision-Language Models (Med-LVLMs) haveemerged as powerful tools for medical image understanding and au-tomated reporting. However, existing models often suffer from modal-ity bias, where correlations between background context and diagnos-tic labels cause over-reliance on spurious priors rather than lesion evi-dence, leading to clinically irrelevant or hallucinated outputs. To addressthis, we propose a Counterfactual Medical Preference Optimization(CoMedPO) framework that learns an unbiased policy from a biasedreference model. Unlike conventional direct preference optimization (DPO)methods that inherit dataset biases, CoMedPO mitigates spurious cor-relations while preserving the valuable indirect effect between lesionand background dependencies in medical images. This causal formula-tion yields a new preference-based loss with theoretical guarantees forbias mitigation and robust policy learning. Our framework is model-agnostic and seamlessly integrates with existing Med-LVLM alignmentpipelines. Extensive experiments on Med-VQA and medical report gen-eration demonstrate that CoMedPO consistently improves factual accu-racy and clinical relevance, outperforming state-of-the-art DPO variants.These results suggest that CoMedPO provides a principled route towardtrustworthy and clinically grounded multimodal medical reasoning. Ourcode is available at https://github.com/zxgapollo/CoMedPO.

Multimodal large language models (MLLMs) can inadver-tently memorize privacy-sensitive information during training. Whileexisting unlearning methods can remove such content, they often severelydegrade the model’s foundational capabilities, such as general imageunderstanding. This critical shortfall motivates our investigation into be-nign memory forgetting, the precise removal of targeted, privacy-sensitiveknowledge while rigorously preserving unrelated capabilities. To pioneerand evaluate progress toward this objective, we introduce S-MLLMUnBench, the first benchmark designed to jointly and quantitatively assessan unlearning method’s efficacy in knowledge erasure and the preservationof image understanding. Furthermore, we propose the Sculpted MemoryForgetting Adapter (SMFA), a new framework that enables benign mem-ory forgetting. SMFA confines forgetting to designated memory regions,maintaining overall model performance. By initially fine-tuning the modelto replace sensitive outputs with refusals, SMFA generates a memoryforgetting adapter, followed by a retaining anchor-guided masking mech-anism that safeguards unrelated knowledge. Extensive experiments onS-MLLMUn Bench demonstrate that existing methods fail to achievebenign forgetting, whereas our proposed SMFA serves as an effectivebaseline, successfully achieving targeted knowledge erasure without com-promising the model’s foundational visual capabilities. Code and dataare available at https://github.com/zeng-zhen/S-MLLMUn.

Deploying AI-generated video detectors in real-world ser-vices demands an ultra-low false positive rate (FPR) on real videosto avoid falsely rejecting authentic content, a regime where standardmetrics such as AUROC fail to reflect actual operating behavior. Weintroduce Spatial Patch-Level Incoherence and Temporal Roughness(SPLIT), a training-free detector that operates on patch tokens froma frozen vision encoder to detect both fully generated and partiallyedited videos. SPLIT computes two complementary signals: Two-stepTemporal Roughness (TTR), capturing non-smooth patch trajectoriesvia one-step and two-step feature variation contrast, and Local SpatialMotion Incoherence (LSMI), measuring spatially inconsistent temporalchanges through gradients of a feature-space motion field. The two arefused multiplicatively with gamma correction to sharpen real–fake sepa-ration at strict thresholds. We further propose a service-aligned evalua-tion protocol based on Fake Recall at fixed FPR with real-only thresholdcalibration and cross-real threshold transfer. Across three benchmarks—FakeParts, GenVideo, and ViF-Bench—SPLIT achieves the highest FakeRecall at FPR = 0.1%, substantially outperforming supervised and training-free baselines while remaining robust to post-processing with negligi-ble overhead. The code is publicly available at https://github.com/mldljyh/SPLIT.

Scalable generation of outdoor driving scenes requires 3Drepresentations that remain consistent across multiple viewpoints andscale to large areas. Existing solutions either rely on image or video gen-erative models distilled to 3D space, harming the geometric coherenceand restricting the rendering to training views, or are limited to small-scale 3D scene or object-centric generation. In this work, we propose a3D generative framework based on Σ-Voxfield grid, a discrete representa-tion where each occupied voxel stores a fixed number of colorized surfacesamples. To generate this representation, we train a semantic-conditioneddiffusion model that operates on local voxel neighborhoods and uses 3Dpositional encodings to capture spatial structure. We scale to large scenesvia progressive spatial outpainting over overlapping regions. Finally, werender the generated Σ-Voxfield grid with a deferred rendering moduleto obtain photorealistic images, enabling large-scale multiview-consistent3D scene generation without per-scene optimization. Extensive experi-ments show that our approach can generate diverse large-scale urbanoutdoor scenes, renderable into photorealistic images with various sen-sor configurations and camera trajectories while maintaining moderatecomputation cost compared to existing approaches.

Semantic segmentation for uncrewed aerial vehicles (UAVs) isfundamental for aerial scene understanding, yet existing RGB and RGB-Tdatasets remain limited in scale, diversity, and annotation efficiencydue to the high cost of manual labeling and the difficulties of accurateRGB-T alignment on off-the-shelf UAVs. To address these challenges, wepropose a scalable geometry-driven 2D-3D-2D paradigm that leveragesmulti-view redundancy in high-overlap aerial imagery to automaticallypropagate labels from a small subset of manually annotated RGB imagesto both RGB and thermal modalities within a unified framework. Bylifting less than 3% of RGB images into a semantic 3D point cloud andrendering it into all views, our approach enables dense pseudo ground-truth generation across large image collections, automatically producing97% of RGB labels and 100% of thermal labels while achieving 91% and88% annotation accuracy without any 2D manual refinement. We furtherextend this 2D–3D–2D paradigm to cross-modal image registration, using3D geometry as an intermediate alignment space to obtain fully automatic,strong pixel-level RGB-T alignment with 87% registration accuracy and nohardware-level synchronization. Applying our framework to existing geo-referenced aerial imagery, we construct SegFly, a large-scale benchmarkwith over 20,000 high-resolution RGB images and more than 15,000geometrically aligned RGB-T pairs spanning diverse urban, industrial, andrural environments across multiple altitudes and seasons. On SegFly, weestablish the Firefly baseline for RGB and thermal semantic segmentationand show that both conventional architectures and vision foundationmodels benefit substantially from SegFly supervision, highlighting thepotential of geometry-driven 2D-3D-2D pipelines for scalable multi-modalaerial scene understanding. The SegFly dataset and our Firefly baselineare available at https://github.com/markus-42/SegFly.

Recent agentic frameworks for 3D scene synthesis have ad-vanced realism and diversity by integrating heterogeneous generationand editing tools. These tools are organized into workflows orchestratedby an off-the-shelf LLM. Current approaches typically adopt an exe-cute–review–reflect loop: at each step, the orchestrator executes a tool,renders intermediate results for review, and then decides on the tooland its parameters for the next step. However, this design has two keylimitations. First, next-step tool selection and parameter configurationare driven by heuristic rules, which can lead to suboptimal executionflows, unnecessary tool invocations, degraded output quality, and in-creased runtime. Second, rendering and reviewing intermediate resultsafter each step introduces additional latency. To address these issues, wepropose SceneOrchestra, a trainable orchestration framework that opti-mizes the tool-call execution flow and eliminates the step-by-step reviewloop, improving both efficiency and output quality. SceneOrchestra con-sists of an orchestrator and a discriminator, which we fine-tune witha two-phase training strategy. In the first phase, the orchestrator learnscontext-aware tool selection and complete tool-call trajectory generation,while the discriminator is trained to assess the quality of full trajectories,enabling it to select the best trajectory from multiple candidates. In thesecond phase, we perform interleaved training, where the discriminatoradapts to the orchestrator’s evolving trajectory distribution and distillsits discriminative capability back into the orchestrator. At inference, weonly use the orchestrator to generate and execute full tool-call trajec-tories from instructions, without requiring the discriminator. Extensiveexperiments show that our method achieves state-of-the-art scene qualitycompared to previous work.

Spatial reasoning is a core aspect of human intelligence forperception and inference in 3D environments. However, current vision-language models (VLMs) struggle to maintain geometric coherence andcross-view consistency for spatial reasoning in multi-view settings. Inthis paper, we aim to understand the successes and failure of VLMs inmulti-view spatial reasoning, via structured analysis of how VLMs en-code perceptual evidence, integrate relational information, construct in-termediate spatial representations and perform perspective transforma-tion. Guided by cognitive science theories of human spatial reasoning,our analysis examines VLM’s consistency and representational persis-tence across reasoning phases that align with human’s spatial cognition,in explicit and implicit manners. Using a self-constructed spatial bench-mark, namely ReMindView-Bench, which consists of >50,000 VQA sam-ples with complementary views and controlled viewpoint configurations,spatial patterns and query types, our explicit analysis using an LLM-as-graph-parser workflow and self-consistency prompting shows that VLMsremain reliable for in-frame perceptual encoding but degrade sharplywhen integrating information across views. Our implicit analysis of linearprobing and entropy dynamics further reveals progressive loss of task-relevant spatial information and increasing uncertainty over reasoningphases. These findings provide important insights of how multi-view spa-tial mental models are formed, destabilized, and degraded across reason-ing phases in VLMs. The ReMindView-Bench benchmark is available athttps://huggingface.co/datasets/Xue0823/ReMindView-Bench, andthe source codes of benchmark construction and VLM reasoning analy-sis are available at https://github.com/pittisl/ReMindView-Bench.

Recent advances in 3D content generation from text or im-ages have achieved impressive results, yet view inconsistency from 2Dgenerators and the scarcity of high-quality 3D data remain significantbottlenecks. Existing solutions [21, 28] typically adapt large-scale pre-trained text-to-image latent diffusion models to generate 3D GaussianSplats (3DGS). However, these approaches often rely on training complexcascade pipelines that are computationally expensive and scalability-limited. Most critically, the quality of generated 3D assets is inherentlyconstrained by each component capacity and compressed latent space,leading to decoding artifacts and accumulated errors. To address theselimitations, we propose PixGS, a single-stage pipeline for direct high-quality 3DGS generation, which leverages recent advances in pixel-spacediffusion to bypass lossy latent compression while still benefiting from thevast 2D generative priors. By directly denoising 3D Gaussian attributesat each timestep, our method enables precise, splat-level regularizationof both appearance and geometry. Furthermore, we introduce a compre-hensive supervision strategy that incorporates surface normals, depth,and high-frequency structural information, which is often overlooked inprior works. Experiments demonstrate that PixGS outperforms currentstate-of-the-art methods while maintaining a fast inference speed (≈ 1son a single A100 GPU), offering a robust and efficient alternative tomulti-stage generation pipelines.

Large vision-language models (LVLMs) have demonstratedstrong general multimodal capability and are increasingly deployed indownstream systems. This trend has driven growing interest in LVLMpersonalization, which aims to enable models to quickly and effectivelylearn out-of-distribution multimodal concepts to meet user-specific needs.However, many existing methods rely on inference-time training, whichreduces efficiency. They also struggle to maintain accuracy in complexmulti-image, multi-concept settings. These limitations restrict the broaderdeployment of LVLM-based systems. Therefore, this paper proposes in-context prompt tuning (ICPT). Specifically, ICPT employs a lightweightprojection module capable of operating in complex scenarios to extractfine-grained visual semantics from multiple reference images, seamlesslytransforming these features alongside identity-label mappings into con-tinuous prompts. To maximize computational efficiency, this moduleadaptively determines the prompt length based on the intrinsic visualcomplexity of each concept. Crucially, to overcome the environmental bi-ases and cross-concept interference prevalent in real-world applications,we introduce two novel geometric regularizations. These constraints re-fine prompt representations by decoupling key identities from transientenvironmental states and separating concepts to avoid semantic confu-sion. Extensive experiments show that ICPT achieves state-of-the-artpersonalization accuracy across diverse tasks and LVLM backbones.