ECCV 2026 Papers

Recent studies have revealed critical vulnerabilities in Large Vision-Language Models (LVLMs) to sophisticated adversarial attacks that induce the model into generating adversary-specific content. However, existing LVLM-oriented detectors largely rely on external judging pipelines that incur substantial computational overhead, precluding their use in real-time scenarios. Towards this end, we propose PTI (PrefillTime Intervention), a method that probes an LVLM’s internal states to detect adversarial images before any response token is generated. PTI combines two complementary modules trained only on benign data: (i) a masked reconstruction module in the vision tower that identifies visual anomalies through token-wise reconstruction errors, and (ii) a dynamic modeling module in the language model backbone that tracks layer-wise trajectories of salient tokens to detect semantic inconsistencies. We establish a comprehensive evaluation framework covering both black-box and white-box attacks, including L -bounded, unrestricted, p and patch-based perturbations. Extensive experiments demonstrate that PTI achieves state-of-the-art detection performance and robust generalization across diverse threat models. Crucially, its minimal inference overhead paves the way for secure, real-time LVLM applications.

Monocular surface normal estimators trained on large-scale RGB-normal data often perform poorly in the edge cases of reflective, textureless, and dark surfaces. Polarization encodes surface orientation independently of texture and albedo, offering a physics-based complement for these cases. Existing polarization methods, however, require multi-view capture or specialized training data, limiting generalization. We introduce Poppy, a training-free framework that refines normals from any frozen RGB backbone using single-shot polarization measurements at test time. Keeping backbone weights frozen, Poppy optimizes perpixel offsets to the input RGB and output normal along with a learned reflectance decomposition. A differentiable rendering layer converts the refined normals into polarization predictions and penalizes mismatches with the observed signal. Across seven benchmarks and three backbone architectures (diffusion, flow, and feed-forward), Poppy reduces mean angular error by 23–26% on synthetic data and 6–16% on real data. These results show that guiding learned RGB-based normal estimators with polarization cues at test time refines normals on challenging surfaces without retraining. An interactive demo and code are available at https://irnkim.github.io/poppy/.

Visual autoregressive (VAR) models have recently emerged as a promising alternative to diffusion models, achieving competitive performance in text-to-image generation while offering substantially faster inference. Although conditional image generation has been widely studied, training-free prompt-guided image editing remains largely unexplored despite its importance for practical applications. In this paper, we present Visual AutoRegressive Inverse Noise (VARIN), the first noise inversion framework for training-free text-based image editing in visual autoregressive models. VARIN introduces Location-aware Argmax Inversion (LAI), a pseudo-inverse function for the argmax operator that reconstructs the inverse Gumbel noise used during discrete autoregressive sampling. The recovered inverse noise enables accurate reconstruction of the source image while providing controllable guidance for promptdriven image editing. Extensive experiments demonstrate that VARIN produces edits that closely follow the target prompts while effectively preserving the background and structural details of the source image. These results establish VARIN as a simple, efficient, and practical image editing framework for visual autoregressive models.

Weather synthesis aims to add weather effects to input videos while preserving scene identity, structure, and motion. The key limitation of existing methods is the lack of diversity in weather appearance and effective control over weather dynamics (e.g., temporal evolution and particle motion). Most approaches rely on text prompts, which are inherently underspecified and often fail to produce detailed weather characteristics. Additionally, general-purpose video editors optimized for clean and aesthetic outputs tend to suppress heavy weather phenomena, making dense particle effects difficult to generate. To address these, we propose a Semantic-Aware, Physics-Informed, and Geometry-Grounded framework that steers an off-the-shelf video editor to synthesize diverse global appearances and detailed particle dynamics. We factorize the synthesis into three conditional signals, so that each provides a distinct and stable source of guidance: semantics specifies what the weather should look like, dynamics governs how it evolves over time, and geometry determines where it should appear in the scene. Specifically, we introduce (1) semantic-aware appearance anchoring to establish the target appearance from scene semantics and user input; (2) physicsinformed dynamic simulation to generate particle effects by simulating a Gaussian-represented particle field under gravity, wind, and turbulence; and (3) geometry-grounded video synthesis to align the simulated particles with target scene geometry and synthesize the final video. Experiments demonstrate that our method produces diverse, physically and visually realistic weather effects. Furthermore, we show that our synthesized data significantly improves the robustness of autonomous driving semantic segmentation under adverse weather conditions. See our project page at https://jumponthemoon.github.io/w-crafter/.

While Vision-Language Models (VLMs) achieve near-perfect scores on digital document benchmarks like OmniDocBench, their performance in the unpredictable physical world remains largely unknown due to the lack of controlled yet realistic evaluations. We introduce Real5- OmniDocBench, the first benchmark to provide a full-scale, one-to-one physical reconstruction of the complete OmniDocBench v1.5 test set (1,355 images) across five critical real-world scenarios: Scanning, Warping, Screen-Photography, Illumination, and Skew. Unlike prior benchmarks that either lack digital correspondence or employ partial sampling, our complete ground-truth mapping enables, for the first time, rigorous factor-wise attribution of performance degradation, allowing us to pinpoint whether failures stem from geometric distortions, optical artifacts, or model limitations. Our benchmark establishes a challenging new standard for the community, demonstrating that the “reality gap” in document parsing is far from closed, and provides a diagnostic tool to guide the development of truly resilient document intelligence. The Real5-OmniDocBench dataset is publicly available at https:// huggingface.co/datasets/PaddlePaddle/Real5-OmniDocBench.

Open-vocabulary 3D Scene Graph (3DSG) can enhance various downstream tasks in robotics by leveraging structured semantic representations, yet current 3DSG construction methods suffer from semantic inconsistencies caused by noisy cross-image aggregation under occlusions and constrained viewpoints. To mitigate the impact of such inconsistency, we propose RAG-3DSG, which introduces re-shot guided uncertainty estimation. By measuring the semantic consistency between original limited viewpoints and re-shot optimal viewpoints, this method quantifies the underlying semantic ambiguity of each graph object. Based on this quantification, we devise an Object-level Retrieval-Augmented Generation (RAG) that leverages low-uncertainty objects as semantic anchors to retrieve more reliable contextual knowledge, enabling a VisionLanguage Model to rectify the predictions of uncertain objects and optimize the final 3DSG. Extensive evaluations across three challenging benchmarks and real-world robot trials demonstrate that RAG-3DSG achieves superior recall and precision, effectively mitigating semantic noise to provide more reliable scene representations for robotics tasks.

Leveraging the priors of 2D diffusion models for 3D editing has emerged as a promising paradigm. However, maintaining multiview consistency in edited results remains challenging, and the extreme scarcity of 3D-consistent editing paired data renders supervised finetuning (SFT), the most effective training strategy for editing tasks, infeasible. In this paper, we observe that, while generating multi-view consistent 3D content is highly challenging, verifying 3D consistency is tractable, naturally positioning reinforcement learning (RL) as a feasible solution. Motivated by this, we propose RL3DEdit, a single-pass framework driven by RL optimization with novel rewards derived from the 3D foundation model, VGGT. Specifically, we leverage VGGT’s robust priors learned from massive real-world data, feed the edited images, and utilize the output confidence maps and pose estimation errors as reward signals, effectively anchoring the 2D editing priors onto a 3D-consistent manifold via RL. Extensive experiments demonstrate that RL3DEdit achieves stable multi-view consistency and outperforms state-of-the-art methods in editing quality with high efficiency. To promote the development of 3D editing, we will release the code and model.

Chest X-ray report generation (CXR-RG) has the potential to substantially alleviate radiologists’ workload. However, conventional autoregressive vision–language models (VLMs) suffer from high inference latency due to sequential token decoding. Diffusion-based models offer a promising alternative through parallel generation, but they still require multiple denoising iterations. Compressing multi-step denoising to a single step could further reduce latency, but often degrades textual coherence due to the mean-field bias introduced by token-factorized denoisers. To address this challenge, we propose ECHO, an efficient diffusion-based VLM (dVLM) for chest X-ray report generation. ECHO enables stable one-step-per-block inference via a novel Direct Conditional Distillation (DCD) framework, which mitigates the mean-field limitation by constructing unfactorized supervision from on-policy diffusion trajectories to encode joint token dependencies. In addition, we introduce a Response-Asymmetric Diffusion (RAD) training strategy that further improves training efficiency while maintaining model effectiveness. Extensive experiments demonstrate that ECHO surpasses state-of-the-art autoregressive methods, improving RaTE and SemScore by 64.33% and 60.58%, respectively, while achieving up to an 8× inference speedup with negligible degradation in clinical accuracy. Code, data, and models are publicly available at https://echo-midea-airc.github.io/.

Current VLM evaluations often conflate language priors with genuine spatial reasoning. To address this, we introduce CRISP, a novel structural-diagnostic evaluation paradigm that assesses visual spatial intelligence through consistency, the alignment between implicit perception and explicit reasoning. Unlike traditional black-box QA, CRISP utilizes metric 3D Scene Graphs and an oracle intervention protocol to decouple latent reasoning capabilities from perceptual bottlenecks. This granular diagnosis uncovers a systematic perception-reasoning disconnect. Crucially, we reveal that while proprietary models possess robust latent reasoning engines, they suffer from inaccurate metric estimation and a critical failure to leverage their implicit structural representations. Conversely, open-source models remain fundamentally bottlenecked by their lack of multi-hop compositional reasoning. By shifting the focus from merely “guessing correctly” via language priors to genuinely “perceiving, verifying, and reasoning,” CRISP offers a rigorous roadmap for multimodal alignment beyond end-to-end post-training. The code and dataset are available at https://github.com/iiyamayuki/CRISP-Bench.

Vision–Language Encoders (VLEs) are widely adopted as the backbone of zero-shot referring image segmentation (RIS), enabling textguided localization without task-specific training. However, prior works underexplored the underlying biases within mid-layer representations that preserve positional and language-specific information. Through layerwise investigation, we reveal that the conventionally used final-layer multimodal embeddings prioritize global semantic alignment, leading to two coupled consequences. First, vision embeddings exhibit weak sensitivity to positional cues. Second, multilingual text embeddings form language-dependent geometric shifts within the shared space. Motivated by these findings, we identify an underexplored pathway within VLE mid-layers to construct a spatial map, applicable for improving zero-shot RIS by 1–7 mIoU on nine RefCOCO benchmarks. Furthermore, leveraging mixed-language mid-layer embeddings yields enhanced spatial grounding accuracy (+7–8 mIoU and IoU@50), albeit with increased inference cost, and also improves performance on the zero-shot text-to-image retrieval task. Our work opens up the discussion about the effects of effective representational bias probing of VLEs for enhanced spatial grounding. The code is available at https://github.com/kaist-cvml/Biased2Grounded.

State-of-the-art flow models generate stunning images from text or image prompts. However, they suffer from diversity collapse when generating multiple samples under the same conditioning. Existing methods address this issue via either latent guidance, which has limited effectiveness, or sample selection, which relies on external reward models that incur significant inference-time overhead. In this work, we introduce an efficient, training-free self-guidance mechanism to mitigate diversity collapse without requiring additional reward models. Specifically, we disperse the internal features of the flow model during batch generation with feature self-guidance. Further, to keep the features close to the manifold, we introduce a manifold regularization step that projects these dispersed features back onto the data manifold, ensuring diverse generation without sacrificing alignment with the input conditions. Our method integrates seamlessly as a plug-and-play module into pretrained flow models, adding only a marginal inference cost. Experiments demonstrate significant improvements in diversity while preserving fidelity across several conditional flow models, including multi-step and few-step text-to-image, depth-to-image, and reference image generation.

Forecasting longitudinal brain lesion evolution is critical for disease monitoring and treatment planning. Existing approaches typically learn a direct mapping from a baseline image to a future observation, without explicitly modeling the physical mechanisms underlying the lesion progression. Such an entangled modeling of structural deformation and image intensity variation limits physical plausibility, model generalization, and interpretability. To address this, we propose PDF, a Physics-grounded Disentangled Flow matching framework for longitudinal brain disease forecasting. We explicitly decompose the longitudinal modeling of lesion growth into two processes, each learned by a dedicated flow matching network: morphology evolution, which captures lesion growth and structural deformation; and intensity evolution, which models signal changes driven by variations in lesion concentration. To enforce physics-grounded constraints, we introduce a PDE-regularized loss based on lesion growth dynamics, that enforces a diffusion-reactionadvection formulation for morphological evolution. Experiments on three public longitudinal datasets spanning diverse brain diseases demonstrate state-of-the-art performance, validating the effectiveness of the disentangled modeling framework and physics-grounded learning design. Code is publicly available at https://github.com/jhuldr/PDF.

The Visual Geometry Grounded Transformer (VGGT) enables unified feed-forward 3D reconstruction from multi-view images. However, deploying such a high-performance model may expose critical security vulnerabilities. Traditional adversarial perturbations require costly per-scene optimization, while Universal Adversarial Perturbations (UAPs) rely on a single static pattern and fail to effectively attack VGGT. To address these limitations, we propose MVAP-G, a multiview adversarial perturbation generator that produces imperceptible consistent perturbations across multiple views in a single feed-forward pass. To ensure perturbation consistency across diverse scenes, we design a cross-view adversarial alignment mechanism to process multi-view images. Experiments demonstrate that MVAP-G significantly degrades VGGT performance without iterative optimization during inference. This work pioneers multi-view adversarial attacks on 3D foundation models, uncovering severe vulnerabilities and underscoring the urgent need for robust 3D vision systems. The code is available at https://github.com/ qsong2001/mvap-g.

Despite their output being ultimately consumed by human viewers, 3D Gaussian Splatting (3DGS) methods often rely on ad-hoc combinations of pixel-level losses, resulting in blurry renderings. To address this, we systematically explore perceptual optimization strategies for 3DGS by searching over a diverse set of distortion losses. We conduct the first-of-its-kind large-scale human subjective study on 3DGS, involving 39,320 pairwise ratings across several datasets and 3DGS frameworks. A regularized version of Wasserstein Distortion, which we call WD-R, emerges as the clear winner, excelling at recovering fine textures without incurring a higher splat count. WD-R is preferred by raters more than 2.3× over the original 3DGS loss, and 1.5× over the current best method Perceptual-GS. WD-R also consistently achieves state-of-the-art LPIPS, DISTS, and FID scores across various datasets, and generalizes across recent frameworks, such as Mip-Splatting and Scaffold-GS, where replacing the original loss with WD-R consistently enhances perceptual quality within a similar resource budget (number of splats for Mip-Splatting, model size for Scaffold-GS), and leads to reconstructions being preferred by human raters 1.8× and 3.6×, respectively. We also find that this carries over to the task of 3DGS scene compression, with ≈ 50% bitrate savings for comparable perceptual metric performance.

Deploying safety-critical agents requires anticipating the consequences of actions before they are executed. While world models offer a paradigm for this proactive foresight, current approaches relying on visual simulation incur prohibitive latencies, often exceeding several seconds per step. In this work, we challenge the assumption that visual processing is necessary for failure prevention. We show that a trained policy’s latent state, combined with its planned actions, already encodes sufficient information to anticipate action outcomes, making visual simulation redundant for failure prevention. To this end, we introduce DILLO (DIstiLLed Language-ActiOn World Model), a fast steering layer that shifts the paradigm from “simulate-then-act” to “describe-thenact.” DILLO is trained via cross-modal distillation, where a privileged Vision Language Model teacher annotates offline trajectories and a latentconditioned Large Language Model student learns to predict semantic outcomes. This creates a text-only inference path, bypassing heavy visual generation entirely, achieving a 14× speedup over baselines. Experiments on MetaWorld and LIBERO demonstrate that DILLO produces high-fidelity descriptions of the next state and is able to steer the policy, improving episode success rate by up to 15pp and 9.3pp on average across tasks. Code is available at github.com/MaxPappa/DILLO.

Knowing where and how fast objects are moving within a scene is important across various domains. Usually, cameras are used to capture the data necessary for this task, but adding cameras often raises privacy concerns, and the quality of captured frames is heavily influenced by lighting conditions. In this work, we explore using WiFi channel state information (CSI) instead of camera frames for optical flow estimation. We propose WiFlow, a CSI based flow estimator, a preprocessor evaluation for CSI, and three model architectures that offer different trade-offs between accuracy and complexity. Further, we create the first dataset for training and evaluating CSI-based optical flow estimators, and our experiments provide insights into key design elements for this task. Code and data are available at https://visinf.github.io/wiflow.

Real-world sRGB image denoising remains challenging due to the nonlinear characteristics of sensor noise and the difficulty of acquiring aligned clean-noisy image pairs. Supervised denoisers often overfit to limited paired datasets, while self-supervised methods still depend on sufficiently diverse noisy observations. These limitations motivate scalable noise synthesis methods that can model real-world noise without clean ground truth or camera metadata. We propose YeTI, a real-world sRGB noise generation framework that learns from only two noisy observations of the same scene. YeTI uses a Reconstruction Autoencoder to disentangle scene structure and noise characteristics, and models the latent noise distribution with a one-step Conditional Diffusion Transformer trained using consistency objectives. Given a single noisy input at inference time, YeTI generates realistic, signal-dependent noise while preserving the underlying scene content. Extensive experiments demonstrate the effectiveness of YeTI across real-world benchmarks. We evaluate noise generation on SIDD and further assess generalization on SIDD+, MAI2021, and SID, covering smartphone and diverse consumer-camera sensors. Downstream denoising results on DND further show that denoisers trained with YeTI-synthesized images achieve strong real-world performance, highlighting the practical value of clean-image-free and metadata-free noise generation. Code is available at: https://github.com/ByungWanLim/YeTI-You-Only-NeedTwo-Noisy-Images-for-Real-World-sRGB-Noise-Generation

In multi-object text-to-image (T2I) diffusion, ensuring semantic consistency between textual prompts and generated visual content is crucial for image synthesis. However, such consistency constraint is often underemphasized in the denoising process of diffusion models. Although token supervised diffusion models can mitigate this issue by learning object-wise consistency between the image content and object segmentation maps, it tends to suffer from the problems of segmentation map bias and semantic overlap conflict, especially when involving multiple objects. In this paper, we propose ELDiff, a new evidential learningsupervised T2I diffusion model, which leverages the advantages of uncertainty metric and conflict detection to enhance the fault tolerance of unreliable segmentation maps and suppress semantic conflicts, strengthening object-wise consistency learning. Specifically, a pixel evidence loss is proposed to restrain overconfidence in unreliable labels through evidential regularization, and a token conflict loss is designed to weaken the contradiction between semantics through optimizing a measured conflict factor. Extensive experiments show that our ELDiff outperforms existing training based and train-free based T2I diffusion models on SD v1.4, SD v2.1, SDXL, SD v3.5, and Qwen-Image, without requiring additional inference-time manipulations. Notably, ELDiff can be seamlessly extended to the existing training pipeline of T2I diffusion models. Code can be found at https://github.com/QingtaoPan/ELDiff.

Diffusion-based generators have made synthetic images ubiquitous, but detectors often fail under simultaneous shifts in generator, prompt/style, and source-domain. We study AI-generated image detection as a transfer system described by training prior, frozen representation, and decision rule, and ask when classifier head training adds value beyond what is already separable in modern features. As a controlled diagnostic, we fit a prior-conditioned Gaussian discriminant ladder: closed-form heads built from firstand second-order feature statistics under nested covariance assumptions. On Percept-Lens, a unified protocol over 39 public datasets (7.1 million images), the best rung is frequently competitive with, and sometimes exceeds, released detector heads when matched on both prior and encoder. We further quantify strong sensitivity to the training prior, data-efficiency of moment-based heads, and representation dependence of Gaussian shift metrics, motivating (prior, encoder, head)-level reporting and stronger analytical baselines for AIGI transfer.