ECCV 2026 Papers

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.

Recent work explores new opportunities at the intersectionof vision-language-action models (VLAs) and geometric foundation mod-els (GFMs) for 3D reconstruction, such as VGGT. While the resultinggeometric VLAs often show improved performance, it remains unclear(i) if modern VLAs already have sufficient geometric understanding tostart with, (ii) what is the best architecture to inject geometric under-standing into a VLA, and (iii) what is the effect of other design choicesthat affect geometric VLAs. In this paper we provide a rigorous exper-imental analysis to shed light on these questions, for a specific choiceof VLA (GR00T-N1.5) and GFM (VGGT). Our first contribution is toformalize prior work’s intuition that current VLAs lack geometric un-derstanding, by providing a rigorous analysis based on linear probing.The analysis quantifies, for the first time, the “geometric gap” betweenVLAs and GFMs. Our second contribution is to identify and comparedifferent strategies to bridge GFMs with VLAs. We implement threedifferent architectures, which differ in the way they inject geometry inthe VLA, while keeping low-level implementation details as similar aspossible, to ensure a fair comparison. Finally, we analyze the impact ofnon-architectural choices (e.g., training data, number of cameras, recon-struction quality) on the performance of the geometric VLAs.

Diffusion models (DMs), despite their impressive capabilitiesacross a wide range of generative tasks, have been shown to be vulnera-ble to backdoor attacks. However, existing backdoor methods face criticaltrade-offs among key factors: attack performance, stealthiness, time com-plexity, and required poison rates. For example, achieving high attackperformance typically demands a high poison rate and prolonged train-ing, which undermines stealthiness, making the attack more detectableby backdoor defenses. This paper proposes TooBad (trigger optimizationfor backdoor diffusion models), a backdoor framework which introducesa novel DM-tailored trigger optimization technique to dramatically en-hance the performance of backdoor attacks on DMs. Experiments on rep-resentative benchmarks such as CIFAR-10 show that TooBad can achievehigh ASRs (> 85%) at only 0.5% poison rate, significantly lower thanthe 10% typically required by prior work on the same datasets. At 5%poison rate, TooBad reaches nearly 100% ASR within just 3-5 backdoorinjection epochs1 , whereas existing methods need at least 30-50 epochsat double the poison rate for comparable results. Despite its potency,TooBad easily evades SOTA defenses and maintains high utility. Theseresults reveal a critical threat on DMs and highlight the need for morerobust defenses against such stealthy yet efficient attacks. Our code isavailable at https://github.com/tuanvu171/TooBad.

Text-conditioned diffusion models have advanced image andvideo super-resolution by using prompts as semantic priors, and modernsuper-resolution pipelines typically rely on latent tiling to scale to highresolutions. In practice, a single global caption is used with the latenttiling, often causing prompt misguidance. Specifically, a coarse globalprompt often misses localized details (errors of omission) and provideslocally irrelevant guidance (errors of commission) which leads to substan-dard results at the tile level. To solve this, we propose Tiled Prompts,a unified framework for image and video super-resolution that generatesa tile-specific prompt for each latent tile and performs super-resolutionunder locally text-conditioned posteriors to resolve prompt misguidancewith minimal overhead. Our experiments on high resolution real-worldimages and videos show that tiled prompts bring consistent gains in per-ceptual quality and fidelity, while reducing hallucinations and tile-levelartifacts that can be found in global-prompt baselines. Project Page:https://bryanswkim.github.io/tiled-prompts/.

We propose TacFiLM, a lightweight modality-fusion ap-proach that integrates visual-tactile signals into vision-language-action(VLA) models. While advances in VLAs have introduced robot policiesthat are both generalizable and semantically grounded, these modelsmainly rely on vision-based perception. Vision alone, however, cannotcapture the complex interaction dynamics that occur during contact-richmanipulation, including contact forces, surface friction, compliance, andshear. While recent attempts to integrate tactile signals into VLA modelsoften increase complexity through token concatenation or large-scale pre-training, the heavy computational demands of behaviour models necessi-tate lightweight fusion strategies. To address these challenges, TacFiLMoutlines a post-training finetuning approach that conditions intermedi-ate visual features on pretrained tactile representations using feature-wise linear modulation (FiLM). Experimental results on insertion anddrawer opening tasks demonstrate consistent improvements in successrate, direct task performance, completion time, and force stability acrossboth in-distribution and out-of-distribution tasks. Together, these resultssupport our method as an effective approach to integrating tactile sig-nals into VLA models, improving contact-rich manipulation behaviours.Project page: https://charliem7.github.io/projects/TacFilm/

Stylized video generation requires synthesizing dynamic con-tent that reflects the artistic characteristics of a reference image whilemaintaining semantic control and temporal coherence. Existing fine-tuning-based approaches suffer from two recurring limitations: motion degrada-tion caused by training on static image stylization datasets, which lacktemporal supervision and erode the temporal priors of pre-trained back-bones; and a representation bottleneck from encoder-based style injec-tion, which discards fine-grained stylistic details and bounds stylizationquality by the encoder’s representational capacity. We address both limi-tations through a unified pipeline that covers dataset construction, modelarchitecture, and training strategy. Specifically, we construct MoStyle-5K, a stylized video dataset of 5,000 videos spanning 27 artistic stylesand 4,000 preference pairs, built via a decoupled content-motion syn-thesis pipeline that explicitly preserves temporal priors in the supervi-sion signal. Building on this, we propose Implicit Gated Style To-ken Injection (IGST), which directly concatenates style patches as to-kens to bypass the encoder bottleneck, disables RoPE for style tokensto remove spatial alignment bias, and applies head-wise gated atten-tion to suppress residual structural correlations. The framework is opti-mized with a hybrid image-video objective and refined via Diffusion-DPO post-training using the structured preference pairs in MoStyle-5K. Experiments show that our method achieves superior style fidelityand motion dynamics compared with existing baselines. Project page:https://sixi111.github.io/MoStyle/.

Long-form audio description (AD) requires more than de-scribing visible actions: it must preserve characters, events, relationships,and story context across scenes so that blind and low-vision (BLV) audi-ences can follow a film. Modern video–language models (VLMs) are effec-tive on short clips, but they often treat each moment independently, pro-ducing descriptions that miss who characters are, why events matter, andhow the current scene connects to earlier narrative context. We proposeStoryTeller, a training-free framework for story-aware long-form AD. In-stead of relying only on local visual cues, StoryTeller maintains a verifiednarrative memory that carries forward story-relevant information acrossscenes, enabling later descriptions to remain coherent, grounded, andcontextually informative. Given only raw video and a movie title, Sto-ryTeller can optionally retrieve public movie metadata to resolve namesand story context, while accepting only facts that are supported by thevideo through semantic filtering and VLM verification. The method re-quires no subtitles, scripts, AD transcripts, aligned captions, characterbanks, precomputed face identities, or task-specific fine-tuning. To evalu-ate whether generated AD preserves narrative information, we introduceStoryAD-QA1 , a question-answering benchmark that tests whether alanguage model can answer story-context questions using only the gener-ated descriptions. Experiments on standard AD benchmarks and diverselong-form videos show that StoryTeller consistently improves narrativecoherence, factual grounding, and story comprehension over strong base-lines in automatic, QA-based, and human evaluations.

3D Gaussian Splatting (3DGS) achieves real-time novel-viewsynthesis by optimizing millions of anisotropic Gaussians, yet its trainingremains expensive, with the backward pass dominating runtime in thepost-densification refinement phase. We observe substantial update re-dundancy in this phase: many sampled views have near-plateaued lossesand provide diminishing gradient benefits, but standard training stillruns full backpropagation. We propose SkipGS with a novel view-adaptivebackward gating mechanism for efficient post-densification training. SkipGSalways performs the forward pass to update per-view loss statistics, andselectively skips backward passes when the sampled view’s loss is consis-tent with its recent per-view baseline, while enforcing a minimum back-ward budget for stable optimization. On Mip-NeRF 360, compared to3DGS, SkipGS reduces end-to-end training time by 23.1%, driven by a42.0% reduction in post-densification time, with comparable reconstruc-tion quality. Because it only changes when to backpropagate withoutmodifying the renderer, representation, or loss, SkipGS is plug-and-playand compatible with other complementary efficiency strategies, enablingadditive speedups. Code is available at https://github.com/ASU-ESIC-FAN-Lab/SkipGS.

Text-to-motion generation aims to synthesize realistic andsemantically aligned human motions from natural language descriptions.Recent diffusion-based approaches predominantly adopt Transformer orCNN-based denoisers; However, the former suffers from quadratic com-plexity with respect to sequence length, while the latter is fundamen-tally constrained by limited receptive fields. To bridge this gap, we pro-pose SFM, a state-space-enhanced diffusion framework that introducesa structured and efficient long-range modeling scheme tailored for mo-tion generation. At its core, SFM integrates two novel components into aUNet backbone: a Frequency State-Space Module (F-SSM), which trans-forms motion features into the frequency domain with text-gated multi-scale modulation, and a Spatial State-Space Module (S-SSM), whichcaptures bidirectional spatial dependencies through parallel state tran-sitions. This dual-axis decomposition enables SFM to model both theglobal structure and fine-grained variations of human motion while pre-serving alignment with linguistic prompts. Extensive experiments onHumanML3D and KIT-ML demonstrate that SFM outperforms strongTransformer- and CNN-based baselines in both fidelity and text-motionalignment, achieving superior R-Precision and FID with optimized speed.

Vision-language dataset distillation (VLDD) compresses alarge image-text paired dataset into a small set of synthetic pairs thatcan efficiently train contrastive vision-language models under strict dataand compute budgets. Most existing methods match expert trajectoriesor cross-modal statistics, yet still enforce full-dimensional alignment in aEuclidean embedding space. This is often overly restrictive due to rank-deficient image–text correlation, with shared semantics concentrated in alow-dimensional range and remaining variation spread across a weaklycorrelated residual subspace. LoRS relaxes alignment at the similaritylevel by low-rank factorization, but does not explicitly control domi-nant alignment capacity and structure in the representation space. Wethus propose a rank-aware hyperbolic alignment (RAHA) that combineshierarchical geometry with explicit alignment-capacity control. RAHAlifts multimodal representations to hyperbolic space and optimizes dis-tilled pairs with asymmetric objectives that enforce geodesic alignmentin the shared range while regularizing the residual subspace to preservemodality-private diversity and improve transfer robustness. Experimentson benchmarks show that RAHA demonstrates competitive cross-modalretrieval and improved transfer indicators under fixed budgets.

Manipulating fast and dynamically moving targets in un-structured 3D environments remains challenging for embodied AI. Ex-isting visual-language-action models and world models struggle with ac-curate 3D geometry and physically meaningful forecasting. We proposePhysMani, a framework that couples a physics-principled 3D Gaussianworld model with a future-aware action policy model. The world modellearns a divergence-free Gaussian velocity field via online optimizationfor fast and physically grounded future dynamics prediction. The pol-icy model integrates the predicted 3D scene future dynamics through alearnable token based cross-attention module. We introduce PhysMani-Bench, a dynamic manipulation benchmark with 16 tasks, and demon-strate a superior success rate over strong baselines in both simulationand real-world robot experiments. Our code and data are available athttps://github.com/vLAR-group/PhysMani

Continual learning enables vision systems to adapt to ever-changing data distributions. Despite significant advances, existing ap-proaches fail to capture continuous and concurrent shifts in classes anddomains, a critical capability for real-world deployment. This work in-troduces Online VIL (Online Versatile Incremental Learning),a novel scenario where class concepts and visual domains evolve simul-taneously online without explicit boundaries. To better adapt to thechallenges of such dynamic environments that more closely resemblereal-world conditions, we propose a novel framework TopFlow, Topologypreservation with Flow matching representation that contains two com-plementary mechanisms: Domain-agnostic Flow Matching (DFM)and Global Topology Preservation (GTP). DFM guides the modelto have domain-agnostic representations by integrating the geodesic flowkernel into contrastive learning. In contrast, GTP maintains the globalstructure of the feature space without explicitly storing past examples. Ourextensive experiments demonstrate that TopFlow e!ectively addresses thelimitations of existing methods within the Online VIL scenario, achievingstate-of-the-art performance in challenging Online VIL. The proposedmethods suggest potential directions for building continual learning sys-tems in realistic dynamic environments. Our implementation code isavailable at https://github.com/KU-VGI/Online-VIL.