ECCV 2026 Papers

We propose HeadsUp, a scalable feed-forward method forreconstructing high-quality 3D Gaussian heads from large-scale multi-camera setups. Our method employs an efficient encoder-decoder archi-tecture that compresses input views into a compact latent representation.This latent representation is then decoded into a set of UV-parameterized3D Gaussians anchored to a neutral head template. This UV represen-tation decouples the number of 3D Gaussians from the number andresolution of input images, enabling training with many high-resolutioninput views. We train and evaluate our model on an internal dataset withmore than 10 000 subjects, which is an order of magnitude larger than ex-isting multi-view human head datasets. HeadsUp achieves state-of-the-artreconstruction quality and generalizes to novel identities without test-timeoptimization. We extensively analyze the scaling behavior of our modelacross identities, views, and model capacity, revealing practical insightsfor quality-compute trade-offs. Finally, we highlight the strength of ourlatent space by showcasing two downstream applications: generating novel3D identities and animating the 3D heads with expression blendshapes.

Executing long-term tasks in dynamic environments requiresembodied agents to maintain robust and adaptive 3D scene represen-tations. However, most existing 3D memory frameworks rely on staticworld assumptions. When objects are displaced by human activities orunobserved events, agents encounter memory-observation conflicts andoften require costly geometric recomputations or inefficient global re-exploration. To address this, we propose HitMem, a hierarchical tempo-ral 3D memory framework with a multi-modal context-aware retrievalmechanism. Through continuous perception, HitMem unifies semanticand spatial information into a lightweight topological graph that cap-tures support relationships, while a temporal decay mechanism dynami-cally regulates memory activeness to mitigate the impact of stale repre-sentations. In addition, the multi-modal context-aware retrieval mecha-nism defaults to filtering candidates using integrated semantic, spatial,and temporal memory features, and activates a specialized two-stage re-trieval process when object displacement is detected. This process com-bines spatial constraints inferred from external agent trajectories withsemantic common sense grounded in class affinities, efficiently identi-fying high-probability candidate regions. Extensive evaluations on ourconstructed Dyna-THOR benchmark demonstrate that HitMem signif-icantly improves object relocation accuracy, reduces exploration costs,and enhances task execution performance in dynamic environments.

We propose a societal bias evaluation method for large visionlanguage models (LVLMs) in the era of strong safety guardrails. Existing benchmarks rely on prompts that ask models to infer attributes of people in images (e.g., “Is this person a CEO or a secretary?”). However, we find that LVLMs with strong guardrails, such as GPT and Claude, often refuse these prompts, making evaluations unreliable. To address this, we change the prior evaluation paradigm by decoupling the task from the depicted person: instead of inferring person’s attributes, we use prompts that do not ask about the person (e.g., “Write a fictional story about an imaginary person.”) and attach the image as provisional user information to implicitly provide demographic cues, then compare outputs across user demographics. Instantiated across three tasks — story generation, term explanation, and exam-style QA — our method avoids refusals even in guardrailed LVLMs, enabling reliable bias measurement. Applying it to 20 recent LVLMs, both open-source and proprietary, we find that all models undesirably use user demographic information in person-irrelevant tasks; for instance, characters in stories are often portrayed as mechanic for male users and nurse for female users. Although still biased, proprietary models like GPT-5 show lower bias than opensource ones. We analyze potential factors behind this gap, discussing continuous model monitoring and improvement as a possible contributor for reducing bias.

Previous works that leverage video models for image-to-3Dscene generation often suffer from geometric distortions and blurry con-tent. Using video generation models to implicitly maintain geometricconsistency according to a single-frame input is ineffective. In this paper,we present a two-stage method, named GeoWorld, that renovates theimage-to-3D scene generation pipeline by providing full-frame geometryfeatures. The first-stage video generation model, followed by a multi-viewgeometry model, produces full-frame geometry features, which are thenused as a mental draft of geometric conditions to aid the second-stagevideo-generation model. A geometric loss is proposed to impose real-worldgeometric constraints, and a geometry adaptation module is introducedto ensure the effective utilization of geometry features. Thanks to full-frame geometric modeling, the two smaller video models in our two-stagemethod can generate higher-fidelity 3D scenes than SOTA methods, whilebeing even faster, e.g. 7.5× faster than Hunyuan-Voyager. Project page:https://peaes.github.io/GeoWorld.

Time awareness is a fundamental ability of omni-modal largelanguage models, especially for understanding long videos and answeringcomplex questions. Previous approaches mainly target vision-languagescenarios and focus on the explicit temporal grounding questions, suchas identifying when a visual event occurs or determining what event hap-pens at a specific time. However, they often make insufficient use of theaudio modality, and overlook implicit temporal grounding across modal-ities (for example, identifying what is visually present when a characterspeaks, or determining what is said when a visual event occurs), de-spite such cross-modal temporal relations being prevalent in real-worldscenarios. In this paper, we first formally define the audiovisual tem-poral grounding task to systematically encompass both explicit andimplicit cross-modal temporal relations. To tackle this task, we proposeChronusOmni, an omni-modal large language model designed to en-hance temporal awareness for both explicit and implicit audiovisual tem-poral grounding. Specifically, we interleave text-based timestamp tokenswith visual and audio representations at each time unit, enabling uni-fied temporal modeling across modalities. Furthermore, to enforce correcttemporal ordering and strengthen temporal reasoning, we incorporate re-inforcement learning with specially designed reward functions. Moreover,we construct ChronusAV, a temporally-accurate, modality-complete, andcross-modal-aligned dataset to support the training and evaluation of theaudiovisual temporal grounding task. Experimental results demonstratethat ChronusOmni achieves state-of-the-art performance on audiovisualtemporal grounding datasets such as ChronusAV and LongVALE, out-performing the second-best results by 17.3% and 28.8%, respectively, anddelivers top-tier performance on visual-only temporal grounding datasetsincluding Charades-STA and ActivityNet. These achievements highlightthe strong temporal awareness of our model across modalities, while pre-serving its robust capabilities in general video and audio understanding.

We perceive our surrounding environments with an active fo-cus, paying more attention to regions of interest, such as the shelf labelsin a grocery store or a family photo on the wall. When it comes to scenereconstruction, this human perception trait calls for spatially varyingdegrees of detail ready for closer inspection in critical regions, prefer-ably reconstructed on demand as users shift their focus. While recentapproaches in 3D Gaussian Splatting (3DGS) can achieve fast, general-izable scene reconstruction from sparse views, their uniform resolutionoutput leads to high computational costs, making them unscalable tohigh-resolution training. As a result, they cannot leverage available imagecaptures at their original high resolution for detail reconstruction. Per-scene optimization methods reconstruct finer details with heuristic-basedadaptive density control, yet require dense observations and lengthy of-fline optimization. To bridge the gap between the prohibitive cost of high-resolution holistic reconstructions and the user needs for localized finedetails, we propose the problem of localized high-resolution reconstruc-tion through on-demand generalizable Gaussian densification. Given aninitial low-resolution 3DGS reconstruction, the goal is to learn a gener-alizable network that densifies the reconstruction to capture fine detailsin a user-specified local region of interest (RoI), based on sparse high-resolution observations of the RoI. This formulation avoids the high costand redundancy of uniformly high-resolution reconstructions and enablesthe full leverage of high-resolution observations in critical regions. To ad-dress the problem, we propose GaussianLens, a feed-forward densifica-tion framework that fuses multi-modal information from the initial 3DGSand multi-view images. We further propose a pixel-guided densificationmechanism that effectively captures details under significant resolutionincreases. Experiments demonstrate our method’s superior performancein local high-fidelity detail reconstruction and strong scalability to im-ages of up to 1024 × 1024 resolution.

Large Vision-Language Models (VLMs) have achieved re-markable multimodal performance yet remain prone to factual halluci-nations, particularly in long-tail or specialized domains. Moreover, cur-rent models exhibit a weak capacity to refuse queries that exceed theirparametric knowledge. In this paper, we propose a systematic frame-work to enhance the refusal capability of VLMs when facing such un-known questions. We first curate a model-specific "Visual-Idk" (Visual-Idon’t know) dataset, leveraging multi-sample consistency probing to dis-tinguish between known and unknown facts. We then align the modelusing supervised fine-tuning followed by preference-aware optimization(e.g., DPO, ORPO) to effectively delineate its knowledge boundaries. Re-sults on the Visual-Idk dataset show our method improves the TruthfulRate from 57.9% to 67.3%. Additionally, internal probing also demon-strates that the model genuinely recognizes its boundaries instead of justmemorizing refusal patterns. Our framework further generalizes to out-of-distribution medical and perceptual domains, providing a robust pathtoward more trustworthy and prudent visual assistants.

Group Activity Detection (GAD) aims to jointly infer groupmemberships and collective activities from videos. Existing approachestypically rely on actor-centric features to model group activities, assum-ing that collective semantics can be inferred solely from inter-actor re-lations. This overlooks critical context-interactive information, such asexplicit spatial constraints and actor–scene dependencies, which are es-sential in crowded and multi-view environments. To address these lim-itations, we propose a Context-Interactive Reasoning framework thatjointly models spatial topology and group semantic context. Specifically,a Short-Term Interaction Regularizer (STIR) softly constrains inter-actor spatial relations with learnable frame-level topology priors, sup-pressing spurious connections and promoting spatially coherent group-ing. Complementarily, a Long-Range Context Conditioner (LRCC) se-lectively incorporates global scene semantics via data-dependent gating,enabling activity-aware context utilization while avoiding unnecessarynoise. Extensive experiments on challenging benchmarks demonstratethat our method achieves state-of-the-art performance in both groupmembership identification and collective activity recognition.

Multimodal large language models (MLLMs) have made rapidprogress, yet they still exhibit object hallucination, generating plausiblebut incorrect descriptions that are inconsistent with the visual input. Di-rect Preference Optimization (DPO) mitigates this by training modelsto prefer non-hallucinated responses over hallucinated ones, and recentefforts further enrich the preference data with relevant context. How-ever, it remains unclear whether DPO actually leverages such context.To investigate this, we propose Contextual Preference Gain (CPG), asimple metric that measures how much a model’s preference strengthenswhen relevant context is provided. We find that higher CPG consistentlycorresponds to lower hallucination, yet standard DPO and its variantsexhibit only limited CPG, indicating that they underutilize contextualinformation and thus remain prone to hallucination. To address this, wepropose Context-Calibrated DPO (C2 -DPO), which directly maximizesCPG while preserving the original preference ordering. Across multiplebenchmarks, C2 -DPO substantially reduces hallucination without com-promising general reasoning, relatively reducing the Object HalBenchhallucination rate of Qwen2-VL-Instruct-2B by 36%. Code is availableat https://github.com/mlvlab/C2-DPO

Composition is a high-level visual intent that governs wheresubjects are placed and how a scene is organized, yet current unifiedmultimodal models remain unreliable at fine-grained composition recog-nition and struggle to turn such intent into controllable generation. Wepresent COMPASS, the first unified multimodal framework that groundscomposition-intent control in a single system spanning both compositionperception and composition-guided generation, with a shared expert to-ken τc as the central intent anchor. On the perception side, COMPASSinjects composition expertise into an MoE backbone in a minimally in-vasive manner and distills the inferred intent into τc . On the generationside, COMPASS reuses τc as a global conditioning signal that steers thedenoising trajectory, effectively converting passive composition analysisinto explicit layout control. To support systematic instruction-followingcomposition learning and evaluation at scale, we construct Comp-11, alarge-scale dataset with an 11-class taxonomy and reasoning-augmentedannotations. Extensive experiments show that COMPASS substantiallyimproves category-level composition understanding and delivers morecomposition-consistent, prompt-faithful generation than strong baselines.The code and dataset for this work will be released here.

Current token-sequence-based Large Language Models (LLMs)struggle to directly process 3D Boundary Representation (B-rep) mod-els that contain complex geometric and topological information. To thisend, we propose BrepLLM, the first multimodal framework that enablesLLMs to directly parse and reason over raw B-rep data. BrepLLM adoptsa two-stage training pipeline: cross-modal alignment pre-training andtwo-stage LLM fine-tuning. In the first stage, we design an adaptiveUV sampling strategy to convert B-reps into graph representations thatintegrate geometric and topological information. Subsequently, we con-struct a hierarchical BrepEncoder to extract features from geometricelements (faces and edges) and topology, generating a global token anda sequence of node tokens. Then, via contrastive learning, we conduct aninitial alignment between this global token and the text embeddings of afrozen CLIP text encoder (ViT-L/14). In the second stage, we integratethe pre-trained BrepEncoder into the LLM and employ a two-stage pro-gressive strategy to align the sequence of node tokens: (1) training anMLP-based semantic mapping network that utilizes the prior knowledgeof a 2D-VLM to align the B-rep representation to the 2D visual semanticspace; (2) utilizing LoRA for parameter-efficient fine-tuning of the Q-Former and the LLM backbone network to achieve the final 3D-languagegeneration capability. Furthermore, we construct the Brep2Text dataset,which contains 269,444 B-rep and text question-answer pairs. Experi-ments demonstrate that BrepLLM achieves SOTA performance on 3Dobject classification and captioning tasks.The project page is availableat https://user-deng.github.io/BrepLLM/.

Recent advances in vision–language models (VLMs) such asCLIP have demonstrated strong generalization across natural-image do-mains. However, adapting these models to biomedical imaging is non-trivial: full-model fine-tuning is computationally expensive, while medi-cal data are often scarce and exhibit subtle, fine-grained inter-class differ-ences, making parameter-efficient adaptation particularly critical. VisualReprogramming (VR) offers a parameter-efficient alternative by inject-ing learnable perturbations into the input space, but existing VR ap-proaches for VLMs mainly focus on positive class prompts and overlookconfusing negatives, leading to miscalibrated predictions in fine-grainedmedical scenarios. We present BioMedVR, the first VR-based frame-work for biomedical imaging, enabling few-shot adaptation of pretrainedVLMs through compact learnable VR modules. To mitigate class confu-sion, we introduce a Confusion Minimization Mechanism that leveragesLLM-generated confusion-aware attributes together with a Confusion-Suppression Loss to explicitly reduce false-positive alignment. More-over, the designed Mixture-of-Prompt Experts combines a positive ex-pert for main-class discrimination and a negative expert for confusionsuppression, balanced via adaptive gating. Extensive experiments on 18datasets—including 11 biomedical datasets and 7 natural image bench-marks—demonstrate that BioMedVR achieves superior accuracy andgeneralization, effectively bridging VR and VLMs in biomedical domains.Project page: https://jxliu-ai.github.io/biomedvr-page/.

Even modern AI models often remain vulnerable to multimodal queries in which harmful intent is embedded in images. A widely used approach for safety alignment is training with extensive multimodal safety datasets, but the costs of data curation and training are often prohibitive. To mitigate these costs, inference-time alignment has recently been explored, but they often lack generalizability across diverse multimodal jailbreaks and still incur notable overhead due to extra forward passes for response refinement or heavy pre-deployment calibration procedures. Here, we identify insufficient visual attention to safety-critical image regions as one of the key causes of multimodal safety failures. Building on this insight, we propose Multimodal Risk-Adaptive Steering (MoRAS), which enhances safety-critical visual attention via concise visual contexts for accurate multimodal risk assessment. This risk signal enables risk-adaptive steering for direct refusals, reducing inference overhead while remaining generalizable across diverse multimodal jailbreaks. Notably, MoRAS requires only a small calibration set to estimate multimodal risk, substantially reducing pre-deployment overhead. We conduct various empirical validations across multiple benchmarks and MLLM backbones, and observe that the proposed MoRAS consistently mitigates jailbreaks, preserves utility, and reduces computational overhead compared to state-of-the-art inference-time defenses. Code is available at https://github.com/snumprlab/moras.

Multi-teacher distillation has emerged as a powerful paradigmfor consolidating diverse visual knowledge from multiple Vision Founda-tion Models (VFMs). However, existing methods typically rely on si-multaneous feature alignment, which forces the student to match mul-tiple teacher distributions directly. This rigid approach often inducesrepresentational dissonance and semantic conflicts, where teachers withdisparate architectures and divergent optimization goals lead to sub-optimal performance and restricted architectural flexibility. In this pa-per, we propose a novel framework termed ACTok that recasts multi-teacher distillation as a token-mediated aggregation process. By treat-ing visual features from both CNNs and Transformers as sets of tokens,we introduce learnable tokens that serve as flexible neural proxies tobridge diverse representational spaces. To avoid specialized modifica-tions on student backbones, we further design an alternating interactionmechanism that enables these learnable tokens to dynamically recon-cile complementary knowledge across teachers. Our ACTok is inherentlyarchitecture-agnostic, facilitating effective knowledge transfer to bothCNN and ViT students. Furthermore, by resolving teacher conflicts atthe token level, our method achieves superior performance using onlythe ImageNet-1K dataset, rendering it orders of magnitude more data-efficient than existing billion-scale distillation baselines. Extensive exper-iments on classification, segmentation, and depth estimation demonstratethat our token-based aggregation consistently outperforms state-of-the-art multi-teacher distillation methods, providing a robust and scalablesolution for VFM knowledge consolidation. Code is publicly availableon https://github.com/VISION-SJTU/ACTok.

We introduce 4DGS360, a di!usion-free framework for 360→dynamic object reconstruction from casual monocular video. Existingmethods often fail to reconstruct consistent 360→ geometry, as their heavyreliance on 2D-native priors causes initial points to overfit to visible sur-face in each training view. 4DGS360 addresses this challenge through anadvanced 3D-native initialization that mitigates the geometric ambigu-ity of occluded regions. Our proposed 3D tracker, AnchorTAP3D, pro-duces reinforced 3D point trajectories by leveraging confident 2D trackpoints as anchors, suppressing drift and providing reliable initializationthat preserves geometry in occluded regions. This initialization, com-bined with optimization, yields coherent 360→ 4D reconstructions. Wefurther present iPhone360, a new benchmark where test cameras areplaced up to 135→ apart from training views, enabling 360→ evaluationthat existing datasets cannot provide. Experiments show that 4DGS360achieves state-of-the-art performance on the iPhone360, iPhone, andDAVIS datasets, both qualitatively and quantitatively. Project websiteat https://jaewon040.github.io/4dgs360/

3D Gaussian Splatting (3DGS) delivers real-time and high-fidelity rendering but remains challenged by unconstrained in-the-wildscenes, where drastic appearance variations and transient objects violatemulti-view consistency. Existing methods are fundamentally limited byindependent and discrete embeddings that struggle to capture continuousenvironmental changes or model spatially-varying local illumination. Toaddress these limitations, we propose WilLaGS, a unified frameworkfor robust 3D scene reconstruction and generative appearance synthe-sis under unconstrained settings. Specifically, we introduce a generativeappearance model where a β-VAE learns a structured and continuousmanifold of global appearance. Conditioned on the latent code, we con-struct a 3D neural appearance field that generates dynamic Tri-Plane fea-tures to encode spatially-varying local illumination effects. Furthermore,to suppress transient artifacts, we present a self-supervised perceptualmasking mechanism that leverages a Teacher-Student (EMA) architec-ture to derive a stable scene consensus, robustly identifying inconsistentregions via perceptual discrepancies. Extensive experiments on multipledatasets demonstrate that WilLaGS achieves state-of-the-art perfor-mance in reconstruction quality and novel view appearance synthesis,while maintaining real-time rendering efficiency.

While feedforward 3D reconstruction excels at efficient novelview synthesis, it typically falters when faced with scenes under varyingillumination. To this end, we introduce WildSplat, the first feedforward3D Gaussian Splatting framework capable of appearance-conditionednovel-view synthesis for unposed in-the-wild images. To handle inconsis-tent photometric conditions, we propose a dual-branch architecture thatexplicitly decouples geometry from appearance. The geometry branchextracts an appearance-invariant 3D structure and jointly predicts cam-era poses. To govern the rendering appearance, the appearance branchinjects target appearance cues into the content features via a globallypre-modulated cross-attention mechanism. To further prevent featureentanglement, we introduce a joint multi-reference training strategy thatstabilizes the training process. Extensive experiments show that Wild-Splat surpasses existing optimization-based and feedforward methods,achieving state-of-the-art performance in in-the-wild novel view synthesisand appearance editing from sparse inputs in a single forward pass.

Vision-language models (VLMs) achieve strong performanceon spatial reasoning benchmarks, yet it remains unclear whether this re-flects structured 3D understanding or reliance on statistical shortcuts innatural images. We introduce a representation-level analysis frameworkthat constructs minimal contrastive pairs to measure how spatial axes areorganized and disentangled within VLM embeddings. Our analysis acrossmultiple model families reveals a consistent vertical-distance entangle-ment: models conflate vertical image position with distance, mirroringthe perspective bias of natural photographs. This bias produces a signif-icant accuracy gap between perspective-consistent and counter-heuristicexamples, and intensifies under data scaling even as overall benchmarkaccuracy improves. We further show that models with similar benchmarkscores can exhibit different internal representations, and that these dif-ferences predict accuracy and robustness across diverse spatial reasoningbenchmarks. To isolate this bias from evaluation-set skew, we introduceSpatialTunnel, a synthetic benchmark designed to expose spatial short-cut biases by removing common correlations present in natural images.Experiments suggest that the entanglement is model-intrinsic, and thatmodels with well-separated spatial axes exhibit greater robustness, indi-cating that well-structured spatial representations lead to more reliablespatial reasoning across diverse benchmarks. Code and benchmark areavailable on the project website.

We propose Vecchia Kernel Surface Reconstruction (VKSR),an accurate implicit surface reconstruction method that efficiently scalesrecent kernel-based techniques to large point clouds with millions ofpoints. While existing (global) kernel methods work well in a sparsesetting, due to low-rank approximations, performance degrades quicklywhen presented with dense point clouds sampled from surfaces with highgeometric complexity or large-scale inputs with millions of points. Toovercome this limitation and inspired by the Gaussian Process litera-ture, VKSR uses Vecchia’s approximation instead of low-rank approx-imations, which naturally shifts computation from a global to a locallevel and allows reconstructing 14M+ points in minutes. VKSR achievesstate-of-the-art results on several challenging datasets while retainingkernel methods’ favorable properties when reconstructing sparse inputs.