ECCV 2026 Papers

When capturing images through glass surfaces or windshields on rainy days, raindrops and reflections frequently co-occur to significantly reduce the visibility of captured images. This practical problem lacks attention and needs to be resolved urgently. Prior de-raindrop, dereflection, and all-in-one models have failed to address this composite degradation. To this end, we first formally define the unified removal of raindrops and reflections (UR3) task for the first time and construct a real-shot dataset, namely RainDrop and ReFlection (RDRF), which provides a new benchmark with substantial, high-quality, diverse image pairs. Then, we propose an effective diffusion-based framework (i.e., DiffUR3) with several target designs to address this challenging task. By leveraging the powerful generative prior, DiffUR3 successfully removes both types of degradations. Extensive experiments demonstrate that our method achieves state-of-the-art performance on our benchmark and on challenging in-the-wild images.

Next-Best-View (NBV) planning is widely adopted for robotic grasping in cluttered and occluded environments. However, existing methods typically assume that a single observation at an arbitrary camera–object distance is sufficient for accurate grasp prediction. This overlooks a critical factor: grasp perception quality is highly sensitive to viewing distance, and the optimal distance varies significantly across object categories and scales. To address this limitation, we propose DiNBVGrasp, a real-time, distance-aware two-stage NBV framework. In Stage I, we distill a large vision-language model (VLM) into a lightweight planner for global semantic NBV exploration, resolving occlusions and generating an initial grasp hypothesis with low-latency inference. In Stage II, we introduce a local distance-aware refinement mechanism that explicitly models the relationship between viewing distance and grasp success probability. During manipulator approach, a lightweight scoring network evaluates intermediate observations and selects the most informative one for final grasp prediction. Extensive simulation and real-world experiments demonstrate that DiNBV-Grasp consistently improves grasp success rates across diverse objects, particularly for small and occluded targets, while preserving real-time performance, highlighting the importance of distance-aware perception in NBV-based grasping. Our code is available at https://github.com/XieZilongAI/DiNBV-Grasp.

Most sign language translation (SLT) methods focus on isolated native sign–spoken pairs (e.g., American Sign Language–English). Extending language-specific SLT models to multilingual translation would improve accessibility by enabling communication across diverse sign and spoken language communities. However, existing multilingual SLT approaches still struggle to learn a unified model that minimizes crosslingual conflicts while capturing shared cross-lingual semantics and preserving language-specific variations across different sign languages. Therefore, we propose Q-BridgeNet, a unified framework for multilingual SLT that jointly mitigates cross-lingual conflicts across both the sign language and spoken language sides. On the sign language side, QBridgeNet learns discrete Q-units via adaptive segmentation and residual vector quantization: a shared base codebook provides language-agnostic semantic primitives, while language-specific residual codebooks refine heterogeneous signing semantics. On the spoken language side, a multilingual LLM is fine-tuned to operate in the Q-unit space, leveraging cross-lingual priors to enable a unified SLT model. Experiments on PHOENIX14T, How2Sign, and CSL-Daily show that Q-BridgeNet effectively mitigates cross-lingual conflicts, achieving state-of-the-art performance on native sign–spoken pairs while also demonstrating strong generalization to non-native pairs. Our source code is publicly available at: https://github.com/FengLiQ/Q-BridgeNet

Category-level object pose estimation seeks to recover a similarity transform (R, t, s) for unseen instances without instance-specific CAD models. Most competitive methods are correspondence-based: prior-free variants regress canonical (NOCS) coordinates directly from local observations and implicitly memorize the canonical frame in the weights, which ties the parameters to category-typical orientations and hurts generalization under distribution shift; prior-based variants introduce a category prior but typically follow a serial deform-then-align pipeline, where underconstrained canonical completion can corrupt correspondences and induce error cascades in pose. We propose PriorPose, a reference-guided correspondence framework that keeps the category prior explicit and solves canonicalization and alignment jointly in a shared feature space. A reference-guided seeded transformer embeds the partial observation and the category prior as token sets and fuses them via geometry-aware seeds, from which the network jointly predicts a per-point NOCS field for visible points and a canonical deformation of the prior that reconstructs a full canonical instance, while a deep pose head regresses (R, t, s) from the induced correspondences. A two-part shape consistency objective, with canonical-space and camera-space consistency losses, couples correspondence, deformation, and pose, reducing reliance on memorized canonical orientations and avoiding deform-then-align error cascades. Experiments on standard and larger-category benchmarks demonstrate that PriorPose sets new state-of-the-art results on most evaluated metrics, especially under strict pose thresholds, while remaining competitive on relaxed pose and IoU metrics and showing improved robustness under shape variation and domain shift.

Generating high-resolution images with latent diffusion models is limited by the cost of high-resolution denoising, while post-hoc super-resolution introduces artifacts and additional latency by operating after decoding. We present the Latent Upscaling Adapter (LUA), a lightweight module that performs super-resolution directly on the generator’s latent code before the final VAE decoding step. LUA integrates as a drop-in component requiring no modifications to the base model or additional diffusion stages, enabling high-resolution synthesis through a single feed-forward pass in latent space. A shared Swin-style backbone with scale-specific pixel-shuffle heads supports ×2 and ×4 factors with nearly 3× lower decoding and upscaling time. Because LUA operates directly on latents, it transfers across diffusion systems–including SDXL, SD3, and FLUX–through architecture-level reuse, adapting only the input convolution with brief fine-tuning rather than retraining from scratch. Extensive experiments demonstrate that LUA closely matches the fidelity of native high-resolution generation at 2K and 4K while offering a practical and efficient path to scalable image synthesis.

Continual learning suffers from two forms of catastrophic forgetting: encoder drift, where visual representations shift as the model adapts to new tasks, and classifier overwriting, where gradient-based updates distort decision boundaries for previously learned classes. We present a unified exemplar-free framework that tackles both effects at the encoder and classifier levels. For the encoder, multi-anchor distillation maintains consistency with multiple historical encoder states and effectively bounds cumulative feature drift along the task sequence. For the classifier, a text-guided analytic classifier computes weights in closed form from incrementally accumulated statistics, completely bypassing gradient-based optimization. To further enhance the classifier, we incorporate a text prior that anchors classifier weights near CLIP’s text embeddings to inject cross-modal semantic knowledge, and propose text subspace regularization that encourages the classifier to learn directions orthogonal to the text embedding subspace, capturing discriminative visual patterns that text semantics alone cannot represent. Experiments on CIFAR-100, ImageNet-100, ImageNet-R, and CUB-200 show state-ofthe-art performance, including gains of +2.7% on CIFAR-100 and +3.5% on ImageNet-100 over prior methods. The code will be included in the supplementary material.

The recent surge in vision-language models (VLMs) has revolutionized open-world perception, yet transferring this success to 3D scene understanding remains impeded by the scarcity of large-scale point-text pairs and the computational overhead of existing frozen-VLM approaches. To bridge this gap, we present ORACLE-3D (Open-world Region-Aligned Cross-modal LEarning for Label-efficient 3D Scene Understanding), a unified framework that effectively co-embeds point clouds, images, and text into a shared latent space without relying on labor-intensive 3Dtext pair construction. ORACLE-3D leverages the image modality as a semantic bridge, employing a novel cross-modal distillation strategy that comprises logit distillation and feature distillation. These components transfer the rich, open-vocabulary capabilities of pre-aligned 2D VLMs directly to the 3D domain. Furthermore, to mitigate the semantic noise arising from inevitable calibration errors and motion artifacts during 3D-2D projection, we propose a vision-point matching module that dynamically rectifies point-pixel misalignments. To ensure robust convergence amidst the gradient bias inherent in heterogeneous multimodal learning, we introduce a two-stage optimization strategy coupled with four task-specific weighted losses. Extensive evaluations on the nuScenes, Waymo, SemanticKITTI, and ScanNet datasets demonstrate the superiority of our paradigm. Specifically, ORACLE-3D outperforms state-of-the-art methods by an average of 11.6 and 14.8 in mIoU on base-annotated and annotation-free tasks, respectively. Project Page: https://ocean-luna.github.io/ORACLE.github.io/.

As Vision-Language Models are increasingly deployed in safetycritical applications, the trustworthiness of their explanations becomes crucial. Explainable AI (XAI) methods for Vision-Language Models often suffer from explanation hallucination, where attribution maps highlight prominent image regions even when prompted with incorrect text descriptions (e.g., highlighting a dog when prompted “cat”). Although this problem is widespread, a formal mathematical analysis of XAI methods and CLIP embeddings is largely missing in the literature. We demonstrate that this phenomenon is not specific to a single architecture but is a fundamental consequence of Linear Semantic Leakage in high-dimensional embedding spaces. We propose a unified theoretical framework, Linear Semantic Attribution (LSA), which generalizes across discriminative methods. We introduce Orthogonal Semantic Projection (OSP), a geometric intervention that utilizes the residual property of OMP to disentangle unique semantic signals from shared concepts. We prove theoretically and demonstrate empirically that OSP minimizes hallucination by orthogonalizing the query vector against distractor concepts, rendering the attribution model blind to shared features while preserving fidelity for correct prompts. The project website, including the code and animations, is available at: https://emirhanbilgic.github.io/OrthogonalSemantic-Projection

Human image animation, which aims to generate a video of a reference subject following a provided action sequence, has received increasing research interest. With the development of diffusion-based/flowbased video foundation models, existing animation works have began to upgrade the guidance information from 2D skeleton/pose to 3D modeling conditions. Despite achieving reasonable results, these approaches face challenges in synthesizing trajectory-controllable human motion within natural scene under changed camera views. In this work, we present a scene-adaptive human image animation framework that controls both human motion and camera trajectories within a reconstructed 3D environment for video generation. To achieve this, we first develop a groundadaptive 3D motion retargeting approach to enable user-friendly motion trajectory control adapting to the changes of elevations of ground and orientations automatically. Then we design a viewpoint-adaptive latent fusion mechanism to inject point-cloud geometric priors through scenevisibility masking into the generative process, providing precise guidance of viewpoint changes under camera control. Experiments on two standard human image animation benchmark datasets demonstrate remarkable improvements of our method over the state of the arts in related video generation metics. Project page: https://robinhood256100.github.io/webdisp

Descriptor-free visual localization eliminates high-dimensional descriptor storage, preserves scene privacy, and simplifies map maintenance, yet its accuracy still lags far behind descriptor-based pipelines. We identify this gap to insufficient geometric discriminability in geometryonly matching. Without visual appearance, current methods underutilize local geometry cues, lack the global context among keypoints, and overfit to a single keypoint detector. We further observe that descriptorfree matching naturally enables multi-detector training, as heterogeneous keypoints can be optimized in a shared geometry-only space without aligning descriptor spaces. Building on these insights, we propose GeoMix, a descriptor-free 2D-3D matching framework that strengthens geometric discriminability at three levels. Locally, directional and distance-aware embeddings enrich neighborhood aggregation with finegrained spatial structure. Globally, learnable context nodes aggregate and redistribute scene-wide information via cross-attention to resolve ambiguities beyond local receptive fields. At the training level, MixTraining exploits this detector-agnostic geometry space to learn representations across multiple keypoint detectors. Extensive experiments on MegaDepth, Cambridge Landmarks, 7Scenes, and Aachen Day-Night show that GeoMix sets a new state of the art among descriptor-free methods, reducing 75th-percentile rotation error by 89% and translation error by up to 90% over the previous best, while generalizing zero-shot to unseen detectors and narrowing the gap to descriptor-based pipelines. Code is available at https://github.com/YejunZhang/Geomix.

We address the challenging problem of dense dynamic scenereconstruction and camera pose estimation from multiple freely movingcameras—a setting that arises naturally when multiple observers capturea shared event. Prior approaches either handle only single-camera input orrequire rigidly mounted, pre-calibrated camera rigs, limiting their practicalapplicability. We propose a two-stage optimization framework that decou-ples the task into robust camera tracking and dense depth refinement. Inthe first stage, we extend single-camera visual SLAM to the multi-camerasetting by constructing a spatiotemporal connection graph that exploitsboth intra-camera temporal continuity and inter-camera spatial overlap,enabling consistent scale and robust tracking. To ensure robustness underlimited overlap, we introduce a wide-baseline initialization strategy usingfeed-forward reconstruction models. In the second stage, we refine depthand camera poses by optimizing dense inter- and intra-camera consistencyusing wide-baseline optical flow. Additionally, we introduce MultiCam-Robolab, a new real-world dataset with ground-truth poses from a motioncapture system. Finally, we demonstrate that our method significantly out-performs state-of-the-art feed-forward models on both synthetic and real-world benchmarks, while requiring less memory. Our code is released at

Arbitrary-scale single image super-resolution (ASSR) has achieved tremendous success under synthetic degradations, but still faces performance deterioration in real-world scenarios where complex distortions are highly entangled with continuous zooming scales. While constructing paired real-world datasets with fractional scales attempts to alleviate this gap, such physical data is fundamentally constrained by discrete scale coverage and hardware-related biases. To overcome these limitations, we propose the Continuous Degradation Rectifier (CDR), a plug-and-play generative proxy that achieves zero-shot domain adaptation at inference time without requiring any real-world paired training data. Designed for any off-the-shelf ASSR model pre-trained on synthetic data, CDR leverages a conditional diffusion process to dynamically project scale-entangled real-world inputs into a canonical bicubic latent space, rectifying the continuous degradation manifold. Specifically, CDR achieves this through a progressive orthogonal decoupling mechanism: it first purifies scale-invariant content from real-world low-resolution images via a discrete codebook bottleneck, and subsequently extracts scaleindependent degradation and continuous scale modulation from unpaired bicubically downsampled reference image, which are orthogonally trained by scale-equivariant contrastive learning. Finally, these orthogonal representations are seamlessly integrated via scale-adaptive bandwidth modulation to guide the reverse diffusion process. Extensive experiments on real-world benchmarks demonstrate that equipping frozen, off-the-shelf ASSR models with CDR not only bypasses the need for domain-specific retraining but also establishes a new state-of-the-art.

Theearlydetectionandpreventionofcolorectalcancer(CRC)ishighly dependentontheaccuratepolypsegmentationfromendoscopicvideos.However, existing video polyp segmentation (VPS) methods are largely hampered by the scarcity of large-scale, high-quality pixel-level annotations. While some semisupervised VPS models have been proposed, their performance is limited by the unreliability of pseudo-labels produced in the training process. In addition, inherent challenges of endoscopic videos, such as low contrast, significant inter-frame variations, and consecutive low-quality frames, further exacerbate the above issues. To address these shortcomings, we propose a unified semi-supervised video polyp segmentation framework, dubbed synergistic confidence–dispersion learning (SCDL), which aims to simultaneously improve pseudo-label reliability, reduce regional supervision bias, and alleviate ambiguities caused by low-contrast features. Our framework has three innovative components: (1) a new spectral convex optimization separation (SCOS) algorithm, which employs convex optimization in a confidence–dispersion space for robust pseudo-label selection, (2) contextual consistency perturbation (CCP), a novel augmentation strategy that reduces regional supervision bias by forcing semantic reconstruction from unreliable areas; and (3) a prototype-guided adaptive refinement (PAR) module that leveragesprototype-drivenaggregationandobject-awarefusiontosharpenfeature discrimination, suppress background clutter, and strengthen temporal coherence. Extensive experiments demonstrate that our model consistently achieves SOTA performance on the SUN-SEG dataset across all labeled-data ratios, and exhibits zero-shot generalization capability on the unseen CVC-ClinicDB dataset. Code are available at https://github.com/Yuanqin-He/SCDL.

AI models are increasingly trained on personal images scraped from social media and public platforms, often without consent, leading to serious privacy violations, such as unauthorized facial recognition and targeted advertising. To counter this, researchers have developed unlearnable examples (UEs), images modified with imperceptible noise to prevent AI models from extracting meaningful information. However, existing UE methods primarily rely on pixel-space noise, which can be bypassed by relearning strategies such as adversarial training, image transformation, and compression. While some techniques improve robustness, they often come at the expense of significant degradation in image utility and perceptual quality. In this paper, we introduce DiffUE to overcome these limitations by injecting noise into the semantic space of images instead of the pixel space. Instead of corrupting pixel values, DiffUE modifies high-level semantic features of images, ensuring robust unlearnability while preserving visual quality and utility. By leveraging a diffusion-based autoencoder framework to manipulate semantic features, DiffUE generates purposeful, natural-looking modifications that effectively resist advanced relearning strategies. Extensive experiments on four datasets, CIFAR-10, CIFAR-100, CelebA-HQ, and ImageNet, as well as a subjective user study, demonstrate that DiffUE significantly enhances the trade-off between image quality and unlearnability, offering a more robust and effective solution for safeguarding personal data in an increasingly exploitative AI landscape.

World models have emerged as a promising paradigm for scaling autonomous driving (AD) data, yet existing video generative models fall short as interactive simulators. Layout-conditioned renderers rely on "oracle" future trajectories of all background agents, rendering them strictly non-reactive. Conversely, pure action-conditioned predictors lack semantic control over complex interactions and suffer from prohibitive diffusion latencies, hindering closed-loop policy learning. To bridge this gap, we present CausalDrive, a controllable, real-time foundation driving world renderer. CausalDrive operates solely on the initial front-view frame, the ego-vehicle’s trajectory, and a macroscopic text prompt. By excluding future NPC layouts, we compel the model to intrinsically predict causal interactions, enabling text-driven control over Driving Sociology—allowing users to dynamically orchestrate diverse counterfactual reactions to identical ego-actions. To overcome the efficiency bottleneck and address the covariate shift in autoregressive generation, we propose a novel Context-Forced DMD architecture. This combines continuous flow-matching with a self-correcting distillation objective, achieving interactive speeds of 12 FPS. This breakthrough transforms the passive video generator into a playable neural simulator. We demonstrate its versatility across three downstream applications: (1) generative closed-loop evaluation with significantly mitigated collision artifacts, (2) large-scale Reinforcement Learning (RL) post-training driven by a Video2Reward module, and (3) real-time human-in-the-loop simulation. Extensive experiments validate that policies trained within CausalDrive’s reactive scenarios exhibit superior interaction capabilities in the real world.

Generating functional neural-network weights from trained model collections is a central problem in weight-space learning. We introduce TiltDiff, a performance-tilted latent diffusion framework for neural-network generation. TiltDiff tokenizes network weights, encodes them into a compact latent space with a Transformer autoencoder, and uses a U-Net-based diffusion model to synthesize latent representations that decode into functional parameters. To favor stronger models, we weight the denoising loss by validation accuracy, biasing the learned distribution toward higher-performing regions of weight space. Experiments show that TiltDiff improves predictive performance, robustness to random parameter masking, and representational diversity over prior weight-generation methods. We further combine diffusion U-Net connectivity with attribution analysis to identify class-specific decision pathways. These pathways exhibit emergent correspondence across independently generated models despite differing raw parameters, and pathwaylevel masking verifies their importance for target-class prediction. Our results show that performance-tilted diffusion generates accurate, robust, diverse, and structurally interpretable neural-network weights.

Progress in tasks like 3D reconstruction and tracking of dynamic scenes from monocular video is constrained by the scarcity of high-quality datasets with dense, complete, and accurate geometric annotations. To address this limitation, we introduce Syn4D, a multiview synthetic dataset of dynamic scenes that includes ground-truth camera motion, depth maps, dense tracking, and parametric human pose annotations. A key feature of Syn4D is the ability to unproject any pixel into 3D to any time and to any camera. We conduct extensive evaluations across multiple downstream tasks to demonstrate the utility and e!ectiveness of the proposed dataset, including 4D scene reconstruction, 3D point tracking, geometry-aware camera retargeting, and human pose estimation. The experimental results highlight Syn4D’s potential to facilitate research in dynamic scene understanding and spatiotemporal modeling.

Vision Transformers (ViTs) achieve state-of-the-art performance but carry massive computational overhead that restricts edge deployment. Although structural pruning has emerged as a key strategy to reduce these costs, existing methods often suffer from severe accuracy degradation or require expensive retraining. Recently, Variance-Based Pruning (VBP) introduced a promising paradigm by selecting neurons based on activation variance; however, it remains limited by statistical noise in finite-sample activation covariance and reliance on bias-only updates that cannot fully account for structural reconstruction error. To address these limitations, we introduce Denoised Variance-Based Pruning with Optimal Brain Bias Compensation (DVBP + OB2C). We leverage random matrix theory to filter noise from the activation covariance spectrum for robust neuron selection and mathematically prove that integrating mean-shift compensation into the Optimal Brain Compression objective reduces the layer-wise Hessian exactly to the activation covariance matrix. This enables an optimal, closed-form update of the remaining weights using the same statistics gathered for selection. Extensive experiments on DeiT, Swin, and ConvNeXt architectures demonstrate that DVBP + OB2C achieves state-of-the-art training-free performance; at 50% MLP pruning, it retains over 90% of the original Top- 1 accuracy on Small and Base variants, outperforming VBP by up to 29.46% (ConvNeXt-T) and 7.33% (Swin-S). The code is available at: https://github.com/geontackee/DVBP_OB2C.

Designing urban spaces that provide pedestrian wind comfort and safety requires time-resolved Computational Fluid Dynamics (CFD) simulations, but their current computational cost makes extensive design exploration impractical. We introduce WinDiNet (Wind Diffusion Network), a pretrained video diffusion model that is repurposed as a fast, differentiable surrogate for this task. Starting from LTX-Video, a 2Bparameter latent video transformer, we fine-tune on a dataset of 13,000 2D incompressible CFD simulations over procedurally generated building layouts. A systematic study of training regimes, conditioning mechanisms, and VAE adaptation strategies, including a physics-informed decoder loss, identifies a configuration that outperforms purpose-built neural PDE solvers. The resulting model generates full 112-frame rollouts in under a second. As the surrogate is end-to-end differentiable, it doubles as a physics simulator for gradient-based inverse optimization: given an urban footprint layout, we optimize building positions directly through backpropagation to improve wind safety as well as pedestrian wind comfort. Experiments on singleand multi-inlet layouts show that the optimizer discovers effective layouts even under challenging multi-objective configurations, with all improvements confirmed by ground-truth CFD simulations.