ECCV 2026 Papers

Face images captured in real-world low light suffer multiple degradations—low illumination, blur, noise, and low visibility, etc. Existing cascaded solutions often suffer from severe error accumulation, while generic joint models lack explicit facial priors and struggle to resolve clear face structures. In this paper, we propose PASDiff, a Physics-Aware Semantic Diffusion in a training-free manner. To achieve a plausible illumination and color distribution, we leverage inverse intensity weighting and Retinex theory to introduce photometric constraints, thereby reliably recovering visibility and natural chromaticity. To faithfully reconstruct facial details, our Style-Agnostic Structural Injection (SASI) extracts structures from an off-the-shelf facial prior while filtering out its intrinsic photometric biases, seamlessly harmonizing identity features with physical constraints. Furthermore, we construct WildDark-Face, a real-world benchmark of 700 low-light facial images with complex degradations. Extensive experiments demonstrate that PASDiff significantly outperforms existing methods, achieving a superior balance among natural illumination, color recovery, and identity consistency. Code and dataset will be available at https://github.com/IVIPLab/PASDiff.

Accurate future video prediction requires both high visual fidelity and consistent scene semantics, particularly in complex dynamic environments such as autonomous driving. We present Re2Pix, a hierarchical video prediction framework that decomposes forecasting into two stages: semantic representation prediction and representation-guided visual synthesis. Instead of directly predicting future RGB frames, our approach first forecasts future scene structure in the feature space of a frozen vision foundation model, and then conditions a latent diffusion model on these predicted representations to render photorealistic frames. This decomposition enables the model to focus first on scene dynamics and then on appearance generation. A key challenge arises from the train-test mismatch between ground-truth representations available during training and predicted ones used at inference. To address this, we introduce two conditioning strategies, nested dropout and mixed supervision, that improve robustness to imperfect autoregressive predictions. Experiments on challenging driving benchmarks demonstrate that the proposed semantics-first design significantly improves temporal semantic consistency, perceptual quality, and training efficiency compared to strong diffusion baselines. We provide the implementation code at https://github.com/Sta8is/Re2Pix.

Video large language models have demonstrated strong video understanding capabilities, but suffer from high inference costs due to the massive number of tokens in long videos. Inspired by efficient biological vision systems, we propose EventSTU, a training-free framework guided by bio-inspired event cameras to significantly reduce redundant tokens, thereby enabling efficient spatio-temporal understanding. In the temporal domain, we design a coarse-to-fine keyframe sampling algorithm that exploits the change-triggered property of event cameras to eliminate redundant frames. In the spatial domain, we design an adaptive token pruning algorithm that leverages the visual saliency of events as a zero-cost prior to guide token reduction. From a holistic spatio-temporal perspective, we further integrate question relevance from keyframe sampling to adaptively allocate token retention budgets. To facilitate evaluation, we construct EventBench, the first event-inclusive, human-annotated multimodal benchmark that covers diverse real-world scenarios. Beyond physical event cameras, EventSTU also supports general video understanding through simulated events. Comprehensive experiments show that EventSTU achieves 2.87× FLOPs reduction and 3.10× prefilling speedup over the strongest baseline while still improving performance.

We introduce DiverseVAR, a test-time framework that enhances the output diversity of text-conditioned visual autoregressive models (VAR) without additional training or substantial computational overhead. While VAR models have recently emerged as strong competitors of diffusion and flow models for image and video generation, they suffer from a critical diversity limitation: even simple prompts often produce nearly identical outputs. This issue has largely gone unnoticed amid the field’s predominant focus on image quality. We address this limitation in two stages. First, we systematically explore diversity enhancement techniques by injecting noise into different components of VAR at test time, finding that noise injection into the text embedding yields the best diversity gains. However, this comes at the cost of sharp degradation in image quality. To recover quality, we propose scale-travel: a latent refinement technique that leverages a multi-scale autoencoder to extract coarsescale tokens to resume generation from intermediate stages. Extensive experiments demonstrate that combining text-embedding noise injection with scale-travel refinement substantially improves diversity while minimizing quality degradation, advancing the diversity-quality Pareto front. Project page: https://diverse-var.github.io/

Sign Languages (SLs) are the primary means of communication for millions of deaf1 individuals, yet existing evaluation metrics for generated SL remain simplistic and poorly aligned with human judgements. We introduce BackTranslation2.0, a linguistically grounded evaluation metric for text-to-sign translation that moves beyond naïve backtranslation. Our approach adopts an agentic framework in which a deterministic pipeline orchestrates a suite of specialised tools to assess four scoring dimensions—grammatical correctness, phonological accuracy, motion fluency, and generation fidelity—aligned with human rater assessments. Tool outputs are not treated independently: a set of large language model (LLM)-based cross-referential comparison modules evaluates consistency across tools and checks outputs against linguistic expectations, enabling structured reasoning over grammatical, phonological, and motion-level evidence. Final dimension scores are computed through deterministic weighted formulas over validated tool outputs. To validate BackTranslation2.0, we introduce and evaluate on a British Sign Language (BSL) dataset rated in a human rater study across the same quality dimensions, following a protocol developed in collaboration between linguists and deaf experts, benchmarking against six baseline metrics. Our method demonstrates strong correlation with human judgements across all dimensions, providing a more comprehensive, interpretable, and linguistically principled evaluation framework for sign language production systems.

Photographic style transfer aims to render a content imagewith the retouching style of a reference photo while preserving scenestructure and texture. Unlike artistic style, photographic style is dom-inated by subtle, continuous tone/color changes and smooth spatial ef-fects, which are difficult to represent with semantics-oriented or text-supervised image embeddings. We propose PETAL, a PhotographicEmbedding for Transfer with an Adaptive LUT. PETAL learns a dedi-cated photographic embedding through self-supervised photographic styleaugmentation and uses it to condition a lightweight adaptive 5D neu-ral LUT. The learned embedding outperforms prior representations inphotographic style retrieval, and the full framework enables reference-based style transfer without test-time optimization, achieving strongerquantitative and perceptual performance than state-of-the-art baselines.Project page: https://petal-pku.pages.dev/.

Referring Expression Counting (REC) requires distinguishing visually similar objects described by fine-grained text cues. Existing methods tackle this via image-text contrastive learning of visual features which aims to distinguish visual features corresponding to positive and negative referring expressions. However, this strategy relies heavily on accurate image–text alignment and is constrained by the limited number of available negative expressions. We argue that these limitations can be avoided by performing contrastive learning entirely in the visual embedding space. To this end, we introduce C-REX, a simple but effective supervised contrastive learning framework that learns fine-grained visual representations by contrasting visual tokens within the same image. By shifting the negative samples from incorrect Referring Expressions to visual tokens from the image, C-REX vastly increases the number of potential negatives, providing richer and more stable supervision. This design leads to stronger fine-grained visual discrimination and better generalization across complex open-world counting settings. C-REX serves as a plug-in module that can be seamlessly applied to any existing REC model without architectural changes. We apply it in three different REC architectures and achieve state-of-the-art results improving by up to 28% in MAE and 24.5% in RMSE. Moreover, we show that our framework is versatile and general, and can be applied to other counting tasks like class-agnostic counting, improving the performance of existing models.

Most PAN-sharpening networks are trained on synthetically degraded panchromatic (PAN) and multispectral (MS) pairs because real high-resolution MS (HRMS) labels are rarely available. However, real satellite acquisitions often suffer from cross-modal misregistration and sensor-dependent optical responses that are absent from synthetic training pipelines, leading to spectral distortions and double-edge artifacts at test time. This suggests that the key bottleneck lies not only in network design but also in the supervision paradigm. In this work, we show that natural videos can serve as a spatial teacher for PAN-sharpening: nearby HR frame pairs provide abundant high-frequency structures and exhibit frame-to-frame shifts that mimic realistic misregistration, enabling strong spatial supervision without HRMS labels. Based on this insight, we propose ViPS, a cross-domain training framework under the paradigm of “Video can teach PAN-Sharpening”. ViPS disentangles supervision sources by learning spatial detail restoration from video-derived pseudo PAN–MS pairs, while enforcing spectral fidelity from real satellite PAN–MS pairs. To align optical degradations across domains, we utilize point spread function (PSF) banks and randomly sample a kernel during training, enabling physically grounded cross-domain learning. Extensive experiments across multiple sensors show that ViPS outperforms very recent state-of-the-art methods under both reducedand full-resolution settings, while maintaining fast inference.

Given an input image, style transfer aims to recast it into the style of a reference image while preserving its content. While this problem is well addressed for natural images, we find that structured images such as charts, diagrams and tables demand capabilities fundamentally distinct from natural image stylization, i.e. visual elements encode data through geometry (bar heights, arc angles, point positions), and dense text must be preserved exactly. We term this structured style transfer as changing appearance while keeping strict element-level fidelity. Even frontier models like GPT-Image-1.5 and Nano-Banana-Pro often struggle here, producing inconsistent transfer, structural distortion, and content leakage. The obvious fix, training on data that restyles a content chart toward a reference, inherits these failures: it distorts the content and leaks the reference into the target. We introduce ChartForge, a data pipeline that instead builds triplets in reverse: starting from a style image, it produces a target of similar style but different content, then a matching content chart. Because the content is derived from the target, the two stay structurally are disentangled from the start, avoiding forwardpipeline leakage. With style-space resampling and multi-dimensional filtering, ChartForge yields ChartStyle-100K, over 100K triplets across 36 chart types and 3 visualization families. For evaluation, we further build ChartStyle-Bench, a benchmark of 300 content-style pairs with 6 metrics. Using ChartStyle-100K, Training ReChart by progressively fine-tuning Qwen-Image-Edit on ChartStyle-100K achieves the best overall performance across all baselines, including GPT-Image-1.5 and Nano-Banana-Pro.

Accurate 3D delineation of prostate cancer lesions in biparametric MRI (bpMRI) supports targeted biopsy and therapy planning, yet clinical reports typically provide only sparse 2D measurements or a single contour. We present SAPP (Spacing-Aware Prompt Propagation), a prompt-to-volume adaptation of SAM 2 that converts one clinical prompt on a single slice (diameter, box, circle, or contour) into a coherent 3D lesion mask from multimodal bpMRI. SAPP couples adaptive multimodal fusion for accurate image-level prompted segmentation with a spacing-aware volume-level prompt propagation module. The propagation uses spacing-decayed, confidence-gated streaming-memory attention and a learned stop rule to prevent leakage in anisotropic volumes. Trained on 3,256 scans from 7 institutions, SAPP generalizes to 14 held-out cohorts (3,504 fully annotated scans plus 516 routine-care weak-label studies), achieving 0.84/0.81 DSC on the prompted slice (box/diameter) and 0.86/0.78/0.76 DSC for volumetric masks (oracle-mask/box/diameter), consistently outperforming baselines while reducing over-propagation. Beyond segmentation, SAPP can accelerate annotation workflows and enable scalable generation of high-quality volumetric lesion masks, supporting the development of robust clinical AI models.

We propose a novel framework for modeling, synthesis, and rendering of mesoscale appearances based on 3D Gaussian Splatting (3DGS) for better quality and performance. The key problem is that synthesis of 3D Gaussians is difficult due to their explicit nature. Therefore, the core of our method is a multi-layer Gaussian texture representation to support high-quality and efficient synthesis of 3D Gaussians. Each layer of the texture contains Gaussians with similar heights to the object surface, which is compatible with 2D image texture synthesis approaches. Through by-example autocovariance-preserving synthesis, we achieve the run-time synthesis of mesoscale appearances with their structures preserved. Due to the deformation of mapping from texture to arbitrary objects and the baking of lighting, directly applying the Gaussian texture for rendering may cause visual artifacts. Thus, we further introduce two components: a deformation-aware mapping and a material-lighting decomposition. The mapping is based on barycentric UV coordinates with a scaling constraint of Gaussians to reduce artifacts. The decomposed materials are represented by latent features to maintain the consistency of the material attributes during synthesis. Consequently, our approach achieves a faithful representation of mesoscale appearances modeled from multi-view images, which supports synthesis and rendering on arbitrary objects in real time, achieving about 500× acceleration compared to the latest NeRF-Texture. Our code is available at: https://github.com/ChenXiang0810/multilayer-gaussian-texture.

Large language models (LLMs) can improve autonomous driving planning but are costly to query online, and existing fast–slow planners often rely on hand-designed triggering rules that either over-call the slow system or call it at the wrong times. We formulate slow-system invocation as a resource-aware sequential decision problem and propose the Adaptive Slow-System Control Gate (ASSCG), which makes framelevel Query/Cache/Drop decisions to refresh, reuse, or suppress slow guidance. ASSCG uses an RWKV backbone for efficient long-horizon gating and is trained with supervised fine-tuning followed by GRPOstyle compute-aware reinforcement fine-tuning. We apply ASSCG to two different fast–slow architectures: (i) AsyncDriver on nuPlan Hard20 closed-loop evaluation, where ASSCG improves score to 67.28 (+2.28) while reducing average end-to-end inference latency by ∼60%; and (ii) a RecogDrive-based dual system that we build by replacing its original VLM-2B module with a lightweight ViT-based fast planner and adding an LLM slow planner, evaluated on NAVSIM, where ASSCG achieves 91.4 PDMS (+0.6) and increases average speed by ∼25%. The project page, including video visualizations and additional results, is available at https://williamxuanyu.github.io/asscg/.

Accurate LiDAR-camera calibration is essential for robust multi-modal perception. Targetless approaches avoid manual setup but remain limited by the scarcity of discriminative cross-modal features. Recent methods address this by reconstructing the scene within a differentiable model, enabling extrinsic optimization through dense photometric supervision. Among these, 3D Gaussian Splatting (3DGS) has been widely adopted as a geometric proxy that bridges LiDAR and camera within a single differentiable framework. However, since 3DGS was originally designed for novel view synthesis, existing methods tend to prioritize rendering quality, causing the proxy geometry to drift from the true LiDAR structure. We propose a framework that preserves the metric geometry of the Gaussian proxy by aggregating multi-view LiDAR observations for dense depth supervision and blocking photometric gradients from updating the Gaussian spatial parameters. We validate our method on public driving datasets, where it consistently outperforms existing targetless methods in calibration accuracy. Code and dataset are available at https://github.com/AIRLABkhu/GeoP-Calib

Lineart colorization is a critical stage in professional content creation, yet achieving precise and flexible results under diverse user constraints remains a significant challenge. To address this, we propose OmniColor, a unified framework for multi-modal lineart colorization that supports arbitrary combinations of control signals. Specifically, we systematically categorize guidance signals into two types: spatially-aligned conditions and semantic-reference conditions. For spatially-aligned inputs, we employ a dual-path encoding strategy paired with a Dense Feature Alignment loss to ensure rigorous boundary preservation and precise color restoration. For semantic-reference inputs, we utilize a VLM-only encoding scheme integrated with a Temporal Redundancy Elimination mechanism to filter repetitive information and enhance inference efficiency. To resolve potential input conflicts, we introduce an Adaptive Spatial-Semantic Gating module that dynamically balances multi-modal constraints. Experimental results demonstrate that OmniColor achieves superior controllability, visual quality, and temporal stability, providing a robust and practical solution for lineart colorization. The source code is available at https://github.com/zhangxulu1996/OmniColor.

Vision-Language-Action (VLA) systems have shown strong potential for language-driven robotic manipulation. However, scaling them to long-horizon tasks remains challenging. Existing pipelines typically separate data collection, policy learning, and deployment, resulting in heavy reliance on manual environment resets and brittle multi-policy execution. We present RoboClaw, an agentic robotics framework that unifies data collection, policy learning, and task execution under a single VLM-driven controller. At the policy level, RoboClaw introduces Entangled Action Pairs (EAP), which couple forward manipulation behaviors with inverse recovery actions to form self-resetting loops for autonomous data collection. This mechanism enables continuous on-policy data acquisition and iterative policy refinement with minimal human intervention. During deployment, the same agent performs high-level reasoning and dynamically orchestrates learned policy primitives to accomplish long-horizon tasks. By maintaining consistent contextual semantics across collection and execution, RoboClaw reduces mismatch between the two phases and improves multi-policy robustness. Experiments in real-world manipulation tasks demonstrate improved stability and scalability compared to conventional open-loop pipelines, while significantly reducing human effort throughout the robot lifecycle, achieving a 25% improvement in success rate over baseline methods on long-horizon tasks and reducing human time investment by 53.7%. Code is available at: https://github.com/RoboClaw-Robotics/RoboClaw.

High temporal resolution is essential for capturing fine-grained dynamics in video understanding, yet current video large language models (VLLMs) and benchmarks predominantly rely on low-frame-rate sampling, discarding dense temporal information. This design compromise is driven by the high cost of frame-wise tokenization, which incurs redundant computation and linear growth in token count with video length. While effective for slowly changing content, such engineering trade-offs severely limit tasks where information is distributed across nearly every frame, such as educational or high-motion videos that require frameby-frame reasoning. To address this gap, we introduce the novel task of High-FPS Video Understanding, the first task specifically designed to evaluate video comprehension at high frame rates, where information is present in nearly every frame. We further propose the first benchmark, DIVE-Bench (Dense Information Video Evaluation Benchmark), to expose the limitations of existing benchmarks whose question– answer pairs are not sensitive to dense temporal variations. To enable efficient high-FPS processing, we propose Gated Residual Tokenization (GRT), a codec-style inter-tokenization framework that operates not only after but also during tokenization. First, Motion-Compensated Gated Inter-Tokenization applies pixel-level motion estimation to gate static regions during tokenization, achieving sub-linear growth in both tokenization time and token count. Second, Semantic-Scene Token Merging further reduces redundancy by merging semantically similar static tokens while preserving dynamic content. Extensive experiments on DIVEBench show that GRT improves high-FPS efficiency while preserving or improving answer quality under dense temporal perception tasks. These results highlight the importance of dense temporal modeling and establish codec-style inter-tokenization as a practical direction for scalable high-FPS video understanding.

Multi-view 3D surface reconstruction is a longstanding challenge in computer vision. Although recent large-scale reconstruction methods based on 3D Gaussian Splatting (3DGS) achieve impressive novelview synthesis, producing high-quality surfaces over large scenes remains difficult, due to complex geometry, long optimization, and limited memory. In this paper, we propose a novel yet simple partitioning method to efficiently and faithfully reconstruct large-scale scene surfaces. Our key insight lies in a scene partitioning method based on viewpoint orientation. This partitioning approach ensures that views with similar orientations are jointly involved for more accurate depth estimations, leading to precise surface reconstructions and balanced computation on multiple GPUs in parallel. In addition, we propose a strategy to detect and repair missing regions in the initial point cloud caused by sparse viewpoints or insufficient textures, thereby further improving the geometric quality. Extensive experiments on the GauU-Scene, MatrixCity, and UrbanScene3D datasets demonstrate that our method outperforms the state-of-the-art approaches in surface reconstruction for large-scale scenes. Project page: https://hanl2010.github.io/VOP-GS. 2 L. Han et al.

Speech-driven 3D facial animation aims to synthesize natural and synchronized facial motions from arbitrary speech audio. Deterministic models struggle with the complex many-to-many mapping, thus yielding over-smoothed motions and lacking expression nuances. In comparison, diffusion-based approaches excel in capturing expressive distributions but suffer from significant latency due to iterative sampling. To this end, we propose MeanTalker, a novel one-step speech-to-motion generative framework built upon Mean Flow. It utilizes a Dual-Time Denoising Transformer (DTDT) for training stability and Geometric-Aware Trajectory Learning (GATL) for precise manifold alignment. Specifically, DTDT introduces an interleaved dual-time encoding to support a progressive curriculum from modeling instantaneous to average velocities, effectively circumventing the instability of direct mean flow training. Furthermore, GATL rectifies the generation trajectory through synergistic dual-space supervision. By applying explicit vertex constraints and projecting latent velocity errors onto the surface manifold, it strictly regularizes the flow direction to guarantee precisely synchronized one-step synthesis. Extensive experiments demonstrate that MeanTalker establishes state-of-the-art lip-sync accuracy while accelerating inference by over 200× compared to iterative diffusion models. Achieving an exceptional Real-Time Factor (RTF) of 0.004, our method effectively bridges the gap between precise motion generation and real-time deployment.

Skeleton-based motion data leak personally identifiable information through both static skeletal structure and dynamic motion patterns, enabling re-identification even without facial features. We present DisentangledTMR, a Transformer Motion Retargeting (TMR) architecture that achieves privacy through explicit architectural disentanglement. Two encoders with complementary inductive biases, temporal convolutions for action and spatial graph convolutions for identity, feed a factorized decoder that fuses their representations through separate crossattention streams and adaptive gating. A three-stage training curriculum progressively establishes disentanglement, reconstruction, and endto-end refinement, and a tunable partial-retargeting ratio trades privacy for compatibility with pre-trained downstream models. On three benchmarks, DisentangledTMR substantially reduces re-identification while preserving action recognition, outperforming single-encoder baselines.