ECCV 2026 论文列表

High-fidelity removal of eyeglasses from video is a major challenge in facial attribute editing, as the underlying facial geometry is often obscured by complex refractive distortions and view-dependent specular reflections. While large-scale generative priors have shown promise in eyeglasses removal via static image inpainting, they often lack the structural constraints necessary to maintain identity, expression, and pose, leading to visible “identity drift” in both static images and dynamic sequences. In this paper, we propose a novel transfer framework that addresses the stochastic nature of generative priors. Our pipeline first extracts highfidelity synthetic face images from a commercial-grade generative model (Nano Banana, Gemini 3 Pro Image), regularizes them via a three-stage structural filtering process to preserve identity, expression, and pose, and finally applies physically-based simulation of lens optics during training to provide diverse, paired data. This process transfers Nano Banana’s photo-realistic, multi-view knowledge into a specialized restoration architecture, JFSnet (Joint Feature-Spatial network). JFSnet integrates DINOv2-based semantic features with a convolutional decoder for spatial reconstruction, leveraging translation equivariance constraints to improve temporal consistency and high-frequency detail preservation. Evaluations on the curated Flickr-Faces-HQ (FFHQ) subset (12,163 images) show that our approach achieves high fidelity and structural accuracy, while maintaining inference speed of 27.68 FPS. In perceptual studies on CelebV-Text video sequences, our results are consistently preferred over diffusion and GAN-based baselines for ocular consistency, temporal stability, and overall restoration quality.

Multimodal deep neural networks deployed in realistic environments must contend with runtime variations: changes in modality quality, overall input complexity, and available platform resources. Current networks struggle with such fluctuations – adaptive networks cannot adhere to a strict compute budget, controller-based networks neglect to consider input complexity, and statically provisioned networks fail at all the above. Consequently, they do not extract maximum utility from the expended computational resources. We present SWAN (Sample and World-Aware Multimodal Network), the first adaptive multimodal network that accomplishes all three goals. SWAN employs a quality-aware controller to assign resources among modalities according to a variable user-specified maximum budget. Within this budget, an adaptive gating module further optimizes e!ciency by scaling layer utilization according to sample complexity. For further gains, SWAN also employs a token dropping module that masks semantically irrelevant multimodal features before performing detections. We evaluate SWAN in the domain of autonomous driving with complex multi-object 3D detection, reducing FLOPs by up to 49% with minimal degradation.

We propose an image generation pipeline that preserves geometric consistency for 6D object pose estimation. Most pose estimation methods rely on training with large-scale annotated real datasets, but collecting such data is often difficult and expensive in practice. To address this limitation, we introduce GenerationPose. First, we create a largescale synthetic dataset by rendering objects from multiple viewpoints using their 3D meshes. We then convert these synthetic images into pseudoreal images using a geometry-conditioned appearance generation model trained on large-scale synthetic–real image pairs. Our experiments show that the model generalizes well to unseen objects and can generate realistic images from synthetic views only. Furthermore, we show that pose estimation models trained on the generated data achieve significantly better performance compared to baseline methods. The code is available at https://github.com/JiafengZhang-1117/GenerationPose.

Recent vision-language models (VLMs) like CLIP have shown impressive anomaly detection performance under significant distribution shift by utilizing high-level semantic information through text prompts. However, these models often overlook fine-grained defect cues, e.g., hole, cut, or scratch, that are essential for understanding the anomaly’s nature. Moreover, the modality gap between images and text can lead to subtle visual evidence being poorly captured in textual descriptions. To address the gap, we enhance the representation of “abnormal” with structured semantics, bridging coarse anomaly signals and fine-grained defect categories. We propose a hybrid prompting mechanism that combines human-readable descriptions of defect types with learnable token embeddings. Building on these ideas, we introduce DAPO, a Defect-aware Prompt Optimization framework for zero-shot multi-type and binary anomaly detection and segmentation under distribution shift. DAPO aligns anomaly-relevant visual features with their corresponding textual semantics by learning hybrid defect-aware prompts that combine fixed textual anchors with trainable token embeddings. We conducted experiments on public benchmarks (MPDD, VisA, MVTec-AD, MAD, and Real-IAD) and an internal dataset. The results suggest that compared to the baseline models, DAPO achieves a 3.6% average improvement in AUROC and average precision metrics at the image level under distribution shift, and a 5.2% average improvement in AUROC and F1 when localizing novel anomaly types under zero-shot settings.

Most image colorization systems operate in Lab space by predicting chroma (ab) while preserving an input-derived luminance channel (L). While effective on standard benchmarks, this fixed-luminance design restricts brightness changes and becomes unreliable when grayscale formation deviates from natural-image luminance, as in historical orthochromatic photography. We propose a luminance-agnostic colorization framework that formulates colorization as full-RGB image editing using a foundation image-editing model. To bridge modern panchromatic and historical orthochromatic conditions, we introduce a mixed grayscale objective that trains the model under both standard luminance grayscale and a red-insensitive grayscale formation. Experiments on COCO, ImageNet, and a multi-instance benchmark show that our method is competitive on standard grayscale inputs and substantially more robust under orthochromatic inputs, with qualitative comparisons and a human study indicating fewer visible color artifacts.

Recent studies demonstrate that Federated Learning (FL) is vulnerable to gradient leakage attacks (GLAs). Revisiting key insights in GLAs under FL reveals that (1) Defenses should focus on protecting semantic and fine-grained details of data; (2) GLAs are effective mainly in early rounds; and (3) To avoid semantic leakage, defenses shouldn’t infer the true labels of private images for obfuscation. Building upon these insights, we present a simple, time-adaptive defense strategy that obfuscates the private gradient by employing the gradient from a synthesized image. To this end, a client trains a diffusion model on its own private dataset to generate synthesized images that fit the distribution of private dataset but are distinct in fine-grained details. In addition, a synthesized image is generated conditioned on a non-identical label from the private image to resist semantic leakage. Defense analysis and empirical evaluations demonstrate that our time-adaptive and label agnostic method can better maintain the trade-off between privacy preservation and model utility against GLAs. Our implementation is publicly available at https://github.com/lalakitchen/Diff2Obs.

Reconstructing articulated objects with multiple movable parts is essential for understanding object structure and enabling physical interaction. However, this reconstruction task poses significant challenges due to the entanglement of geometry, appearance, and motion parameters during optimization. Existing methods rely primarily on photometric supervision, which commonly fails to disentangle these interdependent components, resulting in poor part decomposition with blurred boundaries and geometric artifacts. To address this limitation, we introduce StructureGS, a reconstruction framework for articulated objects that integrates structure-aware guidance into 3D Gaussian Splatting. Our approach leverages oriented bounding boxes of object parts to enforce two key structural properties: spatial coherence, which constrains each part’s geometry to remain compact and spatially coherent within its designated region, and structural connectivity, which enforces physically plausible contact relationships between adjacent parts. These properties are realized through structure-aware losses that inject explicit structural constraints into the optimization process. Extensive experiments demonstrate that our method achieves state-of-the-art performance in articulated object reconstruction, producing high-quality results with welldefined part geometries.

Multi-oriented text is ubiquitous in real-world scenes and remains a major challenge for scene text recognition (STR). Existing rotation-aware methods explicitly estimate text orientation. However, due to the lack of theoretical guarantees, they are prone to error accumulation, increased computational cost, and strong reliance on data. In this work, we incorporate rotation invariance into the STR framework to address these limitations. Specifically, we adopt an encoder– decoder architecture, embedding rotation equivariance in the encoder and rotation invariance in the decoder to construct a fully rotationinvariant network. On the decoder side, we first identify and prove the rotation-invariant property of the cross-attention mechanism and use it to formulate a rotation-invariant text decoder that maps visual features to output text in a rotation-invariant manner. On the encoder side, we propose a rotation-equivariant local–global extraction network that integrates deep equivariant convolutions with self-attention, enabling rotation-equivariant feature extraction while modeling inter-character dependencies and preserving fine-grained visual details. By integrating the encoder and decoder, we obtain an end-to-end Rotation-Invariant Scene TExt Recognition network (RISTER). RISTER provides rotation invariance with theoretical guarantees, enhancing robustness on multioriented samples without introducing additional inference computation or relying on data-driven orientation correction. Experiments show that RISTER achieves state-of-the-art performance on both standard and multi-oriented benchmarks, surpassing the second-best model by 4.0% in accuracy on the general multi-oriented dataset.

Understanding human attention is fundamental for scene interpretation, yet existing approaches often rely on heavily trained models that lack interpretability. Prior methods struggle to jointly reason about gaze targets, attended objects, and visual grounding without extensive supervision. To the best of our knowledge, this work introduces the first training-free Gaze Target Agent (GTA) for gaze-guided reasoning across tasks such as gaze target prediction, attention localization, and object identification. This is achieved by leveraging pretrained vision–language models, augmenting them with visually guided prompts, and employing a memory-based retrieval strategy for high-uncertainty samples to improve performance without additional training. We evaluate our approach using both quantitative metrics and qualitative results. Quantitatively, our method achieves state of the art performance on the GazeFollow and GazeHOI benchmarks. Qualitatively, our agent provides detailed semantic predictions, predicts the correct targets even when ground truth labels are wrong, and remains flexible without vocabulary constraints.

Video vision-language models (VLMs) are increasingly used in long-horizon and streaming settings, yet most video encoders still rely on spatiotemporal self-attention, causing compute and latency to grow quadratically with the number of frames. Existing efficiency methods improve scalability but often lose accuracy relative to full self-attention, for example through aggressive frame/token dropping or coarse attention approximations. We introduce StateKV, an inference-time method that adapts pretrained long-video VLMs to linear-time video prefill by carrying cross-frame context in a fixed-capacity, importance-based recurrent state, paired with a second full per-frame cache used for decoding. Across three long-video benchmarks and seven models spanning three families and multiple scales, StateKV remains close to full self-attention and consistently outperforms dominant sliding-window / recency-based streaming approximations, without fine-tuning or architectural changes. StateKV also reduces video-prefill cost measured FLOPs, enabling stronger accuracy at a fixed compute budget by running larger models. These results suggest a practical step toward scalable long-video understanding.

Choreographic motion generation poses unique challenges for AI, demanding precise semantic control over complex, temporally structured, and expressive full-body dynamics. While existing models can synthesize motion from music, they remain largely black boxes. Conversely, attempting to condition generation on both text and music frequently leads to modality collapse, where dense acoustic rhythms overwhelm sparse semantic text prompts, destroying user controllability. To resolve this spatial-temporal conflict, we propose STREAM (Structural-Temporal Rhythmic Energy-based Attention for Motion), a modality-decoupled diffusion transformer. STREAM strictly separates conditioning pathways: global text semantics dictate the kinematic structure via Adaptive Layer Normalization (AdaLN), while a novel Bimodal Energy-Based Attention Module (BEAM) routes these features to the musical beat without overwriting the semantics. We further introduce Motorica++, a newly curated dataset enriched with domain-specific dance vocabulary and frame-level semantic annotations from existing Motorica dataset. Additionally, to rigorously quantify zero-shot editability, we propose the Exchange Evaluation Protocol and Editable Dance Score (EDS). Through extensive experiments, STREAM achieves state-of-the-art alignment between motion and music while preserving choreographic semantics, positioning AI not merely as a reactive synthesizer, but as a controllable, collaborative partner for artistic direction. The source code and datasets are available at https://github.com/SeongJong-Yoo/STREAM.

Rotary positional embeddings (RoPE) are widely used in diffusion transformers (DiTs) to encode spatial relationships, yet their behavior with mixed-resolution tokens remains underexplored. A natural approach is to rescale token positions from different resolutions into a unified coordinate system before attention, but we show this fails. Our analysis shows that with RoPE, the attention similarity score is a highly structured and periodic function of token distance, so rescaling distances across resolutions moves token pairs to different regions of this periodic function, leading to incorrect attention scores. Motivated by this, we introduce Phase-Aligned Mixed-Resolution Attention (PMA), a trainingfree mechanism that stabilizes mixed-resolution attention. PMA modifies the RoPE position mapping to enforce a consistent positional scale for every query–key pair, ensuring that relative distances are evaluated under a single reference scale. To further improve local coherence near resolution transitions, we incorporate a lightweight boundary refinement module that softly exchanges features across adjacent scales. Experiments on image and video diffusion models validate our analysis and demonstrate consistent improvements in visual fidelity and computational efficiency. Project page: https://hao-yu-wu.github.io/mixed_res/.

Action Quality Assessment (AQA) aims to objectively evaluate performance quality from action videos. Most existing methods follow a “one-by-one” paradigm, training a separate model for each action type. This setting limits real-world deployment, as it requires prior actiontype knowledge to select the corresponding model and suffers from poor generalization across diverse actions. To address these limitations, we study the challenging task of all-in-one AQA, which aims to assess heterogeneous actions within a single unified model. We propose a novel Mixture of Action Knowledge Experts (MoAKE) framework, designed to mitigate negative knowledge transfer caused by large semantic discrepancies among actions. MoAKE learns complementary experts that capture diverse action patterns within a shared semantic space and dynamically aggregates their knowledge to adapt the assessment to the input action. Each expert is tailored with segment-aware prototypes to handle varying temporal lengths, together with an Adaptive Intraand Inter-Segment Relationship Modeling (AIISRM) module to model multi-granularity temporal dynamics. Furthermore, we establish comprehensive benchmarks for all-in-one as well as zero/fewshot AQA. Extensive experiments on three long-term datasets demonstrate that MoAKE significantly outperforms existing methods in the all-in-one setting, while also achieving consistent generalization on three short-term datasets under zero/few-shot evaluation. Code is available at https://github.com/XuHuangbiao/MoAKE.

Video generation models based on Diffusion Transformers (DiTs) have achieved remarkable performance in video synthesis, yet they suffer from high inference latency and computational costs due to the quadratic complexity of 3D attention. Existing acceleration methods primarily reduce computational complexity within each individual denoising steps through techniques such as sparse attention and KVcaching. However, they rigidly adhere to the inherent constraint of the standard diffusion pipeline: every frame in the target video sequence must be subjected to a complete, dense denoising process across all diffusion timesteps. We observe that due to the corresponding contents and motions among adjacent frames, when keyframes with critical semantic transitions are anchored, the intermediate states of others often follow more predictable trajectories, which indicates that such uniform, dense denoising process is inherently redundant for natural video data. To this end, we introduce RhymeFlow, a training-free framework that decouples the denoising trajectories of different frames. Specifically, we first identify a sparse set of pivotal key frames that dominate the latent semantic evolution. Then, only these keyframes undergo dense, step-by-step denoising to ensure structural integrity, while non-keyframes progressively skip denoising steps to minimize computational cost. Since skipped intermediate states of non-keyframes break the temporal coherence in keyframe denoising steps, leading to visual degradation, we further introduce a latent trajectory projection module, which enables keyframes to interact with a complete and temporally consistent sequence representation. Extensive experiments on current DiT-based video generation models demonstrate our method outperforms existing acceleration baselines with higher inference speed and better visual quality.

We present Seen2Scene, the first flow matching-based approach that trains directly on incomplete, real-world 3D scans for scene completion and generation. Unlike prior methods that rely on complete and hence synthetic 3D data, our approach introduces visibility-guided flow matching, which explicitly masks out unknown regions in real scans, enabling effective learning from real-world, partial observations. We represent 3D scenes using truncated signed distance field (TSDF) volumes encoded in sparse grids and employ a sparse transformer to efficiently model complex scene structures while masking unknown regions. We employ 3D layout boxes as an input conditioning signal, and our approach is flexibly adapted to various other inputs such as text or partial scans. By learning directly from real-world, incomplete 3D scans, Seen2Scene enables realistic 3D scene completion for complex, cluttered real environments. Experiments demonstrate that our model produces coherent, complete, and realistic 3D scenes, outperforming baselines in completion accuracy and generation quality.

As generative models increasingly produce samples that are indistinguishable from human-created content, it becomes difficult to determine whether a given data point was part of a model’s natural training set or was generated by the model itself, especially when models memorize and reproduce training data. We formalize this challenge as Member vs Generated Inference (MGI): given a sample and a target generative model, infer whether the sample is a true training member or a generated output of that model. Focusing on image generation, we show that existing membership inference methods systematically misclassify generated samples as training members, while attributionbased methods often misclassify true members as generated. This failure arises because both approaches rely on likelihood-related signals that are similarly elevated for training examples and for the model’s own outputs. To address MGI, we propose Data Circuit Breaker (DCB), a three-stage method that combines complementary signals from a generative model’s autoencoder and latent generator to distinguish training members from generated samples. Across multiple generative models, including image autoregressive and diffusion models, DCB consistently addresses the shortcomings of membership inference and attribution methods, remains effective even when models reproduce near-duplicates of training samples, and generalizes to challenging model derivative settings in which new models are trained on generated data. Our code is available at https: //github.com/sprintml/MemberGeneratedInference.

Faithfully capturing the intricate relationship between motion and geometry in dynamic scenes is essential for 4D reconstruction. Recent state-of-the-art methods rely on monolithic deformation networks or global time-basis factorizations that struggle to represent complex, non-rigid topological changes. We propose MoBa-GS, a framework that resolves this entanglement by introducing a structural inversion: a spatially-factorized motion field coupled with adaptive geometric optimization. First, our model learns a low-frequency dynamic canonical space to represent coarse scene motion. Next, it decomposes complex, non-rigid motion into a Spatially-Varying Motion Basis of local kinematics, predicted from the canonical geometry, which is then linearly combined using dynamic blending weights. This formulation directly acts as an implicit neural scaffold, recovering both geometry and motion from random initialization, thereby removing the dependency on SfM point cloud priors. This design is further augmented with motion-guided densification and positional annealing to reduce geometry overfitting. Extensive experiments show that our framework surpasses prior state-of-theart methods in reconstruction fidelity. Enabled by time-invariant caching, MoBa-GS requires an order-of-magnitude shorter training time (<18 minutes), a compact storage (∼ 11 MB), and achieves real-time rendering speeds (>163 FPS). Our work establishes a new foundation for highfidelity, efficient 4D representations without relying on explicit geometric priors. Code is available at https://github.com/tgy1221/MoBa-GS.

We propose CAT, a trajectory-level continuous action representation framework for robotic manipulation. Existing visuomotor systems often entangle action representation with control frequency or rely on fixed temporal parameterizations. This leads to representational redundancy at high sampling rates and limits the modeling of critical motion. CAT instead encodes action trajectories within a fixed realtime interval into a set of continuous latent tokens. To ensure temporal consistency across varying control frequencies, we further incorporate a frequency-aware positional encoding that establishs a shared temporal coordinate system. Trajectory-level regularization further stabilizes the latent representation. This approach prevents representation growth with timestep density and avoids reliance on predefined temporal parameterizations. Extensive system-level evaluations on LIBERO, MimicGen, and real-world long-horizon manipulation tasks demonstrate that CATbased policies consistently outperform both competitive VQ-based and continuous visuomotor baselines under matched training settings. Across various model backbones and control frequencies, CAT consistently improves success rates. These results highlight the advantages of trajectorylevel continuous action modeling for scalable robotic manipulation across varying control rates.

Diffusion-based decoding has recently emerged as an appealing alternative to autoregressive (AR) generation, offering the potential to update multiple tokens in parallel and reduce latency. However, diffusion vision language models (dVLMs) still lag significantly behind mainstream autoregressive vision language models. This is due to the scarcity and weaker performance of base diffusion language models (dLLMs) compared with their autoregressive counterparts. This raises a natural question: Can we build high-performing dVLMs directly from existing powerful AR models, without relying on dLLMs? We propose DiffusionVL, a family of dVLMs obtained by translating pretrained AR models into the diffusion paradigm via an efficient diffusion finetuning procedure that changes the training objective and decoding process while keeping the backbone architecture intact. Through an efficient diffusion finetuning strategy, we successfully adapt AR pretrained models into the diffusion paradigm. This approach yields two key observations: (1) The paradigm shift from AR-based multimodal models to diffusion is remarkably effective. (2) Direct conversion of an AR language model to a dVLM is also feasible, achieving performance comparable to that of the same AR model finetuned with standard autoregressive visual instruction tuning. To enable practical open-ended generation, we further integrate block decoding, which supports arbitrarylength outputs and KV-cache reuse for faster inference. Our experiments demonstrate that despite training with less than 5% of the data required by prior methods, DiffusionVL achieves a comprehensive performance improvement, with a 34.4% gain on the MMMU-Pro (vision) benchmark and 37.5% gain on the MME (Cog.) benchmark, alongside a 2× inference speedup. The model and code are released at https://github.com/hustvl/DiffusionVL.