WACV 2026 论文列表

Benefiting from the powerful data learning and representation capabilities of neural networks, Learned Image Compression (LIC) methods have demonstrated better Rate-Distortion (RD) performance than traditional image compression frameworks. In this paper, we analyze the role of latent variables in image compression, both qualitatively and quantitatively. We then propose a latent variable compensation method to mitigate the loss introduced by quantization. We also introduce a regularization term for the latent variable mean square error into the loss function, providing more explicit guidance for the compression and reconstruction of the model's internal latent variables. Additionally, we propose a noise compensation method that acts as a plug-and-play component to enhance reconstruction in lossy image compression without significant additional encoding or decoding time. We also present a data augmentation technique involving image inversion, which helps the training set conform to the symmetry inherent in probabilistic modeling for image compression tasks. Extensive experiments demonstrate that the proposed method enhances rate-distortion metrics and visual quality.

Large foundation models have emerged in the last years and are pushing performance boundaries for a variety of tasks. Training or even finetuning such models demands vast datasets and computational resources, which are often scarce and costly. Adaptation methods provide a computationally efficient solution to address these limitations by allowing such models to be finetuned on small amounts of data and computing power. This is achieved by appending new trainable modules to frozen backbones with only a fraction of the trainable parameters and fitting only these modules on novel tasks. Recently, the VeRA adapter was shown to excel in parameter-efficient adaptations by utilizing a pair of frozen random low-rank matrices shared across all layers. In this paper, we propose PVeRA, a probabilistic version of the VeRA adapter, which modifies the low-rank matrices of VeRA in a probabilistic manner. This modification naturally allows handling inherent ambiguities in the input and allows for different sampling configurations during training and testing. A comprehensive evaluation was performed on the VTAB-1k benchmark and seven adapters, with PVeRA outperforming VeRA and other adapters.

Reconstructing animatable 3D animals from videos traditionally depends on sparse semantic keypoints to fit parametric models. Acquiring these keypoints is labor-intensive, and detectors trained on limited animal datasets are often unreliable. We propose 4D-Animal, a keypoint-free framework that reconstructs animatable 3D animals directly from videos. Our method employs a dense feature network to map 2D image representations to SMAL parameters, improving both efficiency and stability. Additionally, we introduce a hierarchical alignment strategy that leverages silhouette, part-level, pixel-level, and temporal cues from pretrained 2D models, ensuring accurate and temporally coherent reconstructions. Extensive experiments demonstrate that 4D-Animal outperforms both model-based and model-free baselines on dog dataset. Moreover, the high-quality 3D assets generated by our method can benefit other 3D tasks, underscoring its potential for large-scale applications. The code is released at \href https://github.com/zhongshsh/4D-Animal \texttt https://github.com/zhongshsh/4D-Animal .

Foundational models are trained on extensive datasets to capture the general trends of a domain. However, in medical imaging, the scarcity of data makes pre-training for every domain, modality, or task challenging. Instead of building separate models, we propose MAFM3 (Modular Adaptation of Foundation Models for Multi-Modal Medical AI), a framework that enables a single foundation model to expand into diverse domains, tasks, and modalities through lightweight modular components. These components serve as specialized skill sets that allow the system to flexibly activate the appropriate capability at the inference time, depending on the input type or clinical objective. Unlike conventional adaptation methods that treat each new task or modality in isolation, MAFM3 provides a unified and expandable framework for efficient multitask and multimodality adaptation. Empirically, we validate our approach by adapting a chest CT foundation model initially trained for classification into prognosis and segmentation modules. Our results show improved performance on both tasks. Furthermore, by incorporating PET scans, MAFM3 achieved an improvement in the Dice score 5% compared to the respective baselines. These findings establish that foundation models, when equipped with modular components, are not inherently constrained to their initial training scope but can evolve into multitask, multimodality systems for medical imaging. The code implementation of this work will be made available upon acceptance.

Despite recent advances in text-conditioned 3D indoor scene generation, there remain gaps in the evaluation of these methods. Existing metrics often measure realism by comparing generated scenes to a set of ground-truth scenes, but they overlook how well scenes follow the input text and capture implicit expectations of plausibility. We present SceneEval, an evaluation framework designed to address these limitations. SceneEval introduces fine-grained metrics for explicit user requirements--including object counts, attributes, and spatial relationships--and complementary metrics for implicit expectations such as support, collisions, and navigability. Together, these provide interpretable and comprehensive assessments of scene quality. To ground evaluation, we curate SceneEval-500, a benchmark of 500 text descriptions with detailed annotations of expected scene properties. This dataset establishes a common reference for reproducible and systematic comparison across scene generation methods. We evaluate six recent scene generation approaches using SceneEval and demonstrate its ability to provide detailed assessments of the generated scenes, highlighting strengths and areas for improvement across multiple dimensions. Our results identify significant gaps in current methods, underscoring the need for further research toward practical and controllable scene synthesis.

In this work, we address the problem of grounding abnormalities in medical images, where the goal is to localize clinical findings based on textual descriptions. While generalist Vision-Language Models (VLMs) excel in natural grounding tasks, they often struggle in the medical domain due to rare, compositional, and domain-specific terms that are poorly aligned with visual patterns. Specialized medical VLMs address this challenge via large-scale domain pretraining, but at the cost of substantial annotation and computational resources.To overcome these limitations, we propose Knowledge to Sight (K2Sight), a framework that introduces structured semantic supervision by decomposing clinical concepts into interpretable visual attributes, such as shape, density, and anatomical location. These attributes are distilled from domain ontologies and encoded into concise instruction-style prompts, which guide region-text alignment during training. Unlike conventional report-level supervision, our approach explicitly bridges domain knowledge and spatial structure, enabling data-efficient training of compact models.We train compact models with 0.23B and 2B parameters using only 1.5% of the data required by state-of-the-art medical VLMs. Despite their small size and limited training data, these models achieve performance on par with or better than 7B+ medical VLMs, with up to 9.82% improvement in mAP_ 50 . Code and models: https://lijunrio.github.io/K2Sight/

Domain adaptation is a key strategy for enhancing the generalizability of deep learning models in real-world scenarios, where test distributions often diverge significantly from the training domain. However, conventional approaches typically rely on prior knowledge of the target domain or require model retraining, limiting their practicality in dynamic or resource-constrained environments. Recent test-time adaptation methods based on batch normalization statistic updates allow for unsupervised adaptation, but they often fail to capture complex activation distributions and are constrained to specific normalization layers. We propose Adaptive Quantile Recalibration (AQR), a test-time adaptation technique that modifies pre-activation distributions by aligning quantiles on a channel-wise basis. AQR captures the full shape of activation distributions and generalizes across architectures employing BatchNorm, GroupNorm, or LayerNorm. To address the challenge of estimating distribution tails under varying batch sizes, AQR incorporates a robust tail calibration strategy that improves stability and precision. Our method leverages source-domain statistics computed at training time, enabling unsupervised adaptation without retraining models. Experiments on CIFAR-10-C, CIFAR-100-C, and ImageNet-C across multiple architectures demonstrate that AQR achieves robust adaptation across diverse settings, outperforming existing test-time adaptation baselines. These results highlight AQR's potential for deployment in real-world scenarios with dynamic and unpredictable data distributions.

Video Instance Segmentation (VIS) faces significant annotation challenges due to its dual requirements of pixel-level masks and temporal consistency labels. While recent unsupervised methods like VideoCutLER eliminate optical flow dependencies through synthetic data, they remain constrained by the synthetic-to-real domain gap. We present AutoQ-VIS, a novel unsupervised framework that bridges this gap through quality-guided self-training. Our approach establishes a closed-loop system between pseudo-label generation and automatic quality assessment, enabling progressive adaptation from synthetic to real videos. Experiments demonstrate state-of-the-art performance with 52.6 \text AP _ 50 on YouTubeVIS-2019 \texttt val set, surpassing the previous state-of-the-art VideoCutLER by 4.4%, while requiring no human annotations. This demonstrates the viability of quality-aware self-training for unsupervised VIS. We will release the code at https://github.com/wcbup/AutoQ-VIS.

We introduce Memoire, a fully automatic, on-device system for personalized photo selection that learns a user's persona directly from gallery tags--people & relations, locations, and events--and ranks images by personal impact rather than generic aesthetics. Memoire constructs a per user tag graph and trains PERSONA-GAT to produce tag importance scores summarizing user preferences across the gallery images. These scores are projected to pixels via PAT, a deterministic grounding module that fuses tags and their importance into personal attention maps. To obtain scalable supervision without collecting user labels, we synthesize virtual-user galleries (diverse identities, events, and locations) and use a vision-language model to annotate image pairs as High vs. Low personal impact conditioned on their personal attention maps. An impact predictor is then trained with a pairwise ranking loss and coupled with a diversity-aware selector to deliver non-redundant top-k image selection. To maintain user privacy, persona learning and inference run entirely on device. Synthetic data and VLM supervision are used only for training of impact predictor. On a real 100-user gallery study, Memoire outperforms strong aesthetics, memorability and multimodal baselines.

Existing color constancy methods based on deep learning primarily rely on the RGB domain and often struggle with accurate illuminant estimation in scenes with minimal spatial information, such as monochromatic environments, leading to suboptimal performance. To address this issue, this paper introduces an approach that utilizes MS (multi-spectral) images estimated by a pretrained RGB-to-MS model, enabling more accurate illuminant estimation. Additionally, we propose a graph-based spectral attention mechanism designed to effectively extract spectral features within the multi-spectral domain, further enhancing the robustness and accuracy of color constancy. This approach demonstrates outstanding effectiveness on our custom dataset, significantly outperforming existing methods. Additionally, when evaluated in the widely recognized NUS-8 and Cube+ datasets, the proposed method shows a substantial relative improvement of 21.5% in NUS-8 and 9.9% in Cube+ compared to previous state-of-the-art methods.

Segment Anything Models (SAM) have demonstrated strong generalization in object segmentation across diverse datasets. However, their training on large-scale semantic segmentation data induces a shape bias, leading to over-segmentation in texture-dominant scenes and severely limiting performance. This limitation is particularly pronounced in domains such as remote sensing and metallographic imaging, where meaningful boundaries are defined by texture variations rather than semantic structure. In this study, we investigate SAM's shape bias and show that a simple fine-tuning strategy -- based on incremental texture augmentations of semantically labeled data -- can effectively calibrate this bias and guide the model toward texture-aware segmentation. By interpolating and replacing textures within semantically labeled regions, we generate texture-diverse instances of the same semantic category, enabling effective fine-tuning without requiring additional manual annotations. We release both the texture-oriented variant of SAM ("TextureSAM") and the texture-augmented dataset used in our experiments to support reproducibility and facilitate further research on shape-texture bias in foundation models. We show that this fine-tuning approach mitigates SAM-2's shape bias, improving segmentation performance on both real-world (RWTD, +0.20 mIoU) and synthetic (STMD, +0.18 mIoU) texture segmentation benchmarks. Our method also demonstrates improved zero-shot generalization on in-the-wild datasets, spanning aerial imagery, satellite scenes, and metallographic microstructures, nearly doubling SAM-2's mIoU in domains where annotations align with texture boundaries. Our code and data are available at https://github.com/Scientific-Computing-Lab/TextureSAM

Portrait stylization aims to cast vivid artistic style drawn from style examples to portrait photos. This task has recently been extensively studied with machine learning algorithms, but it is still difficult for existing methods to stylize portraits from a single style reference, severely limiting these methods for real-world applications. In this paper, we propose a portrait stylization method that learns style reference from a single artistic portrait image. Unlike previous StyleGAN based methods that heavily rely on the quality of GAN inversion or diffusion based methods that introduce computational expensive operations and fall short of precise control, our method achieves high-quality stylization with small computation and parameter budget. Specifically, we employ geometric alignment to build spatial correlation between content images and style reference. A content LoRA and a style LoRA are then jointly optimized based on a pre-trained diffusion backbone respectively, with orthogonal adaptation used to disentangle the content and style information. During inference, the style LoRA is integrated into the diffusion backbone and ControlNet is further combined to facilitate better spatial and identity control. We illustrate abundant stylized portraits with multiple styles. Qualitative comparison, quantitative validation and user study all prove that our method outperforms existing methods, and ablation study demonstrates the effectiveness of each components. Code and pre-trained model will be made publicly available upon paper acceptance.

Accurate polyp segmentation during colonoscopy is critical for the early detection of colorectal cancer and still remains challenging due to significant size, shape, and color variations, and the camouflaged nature of polyps. While lightweight baseline models such as U-Net, U-Net++, and PraNet offer advantages in terms of easy deployment and low computational cost, they struggle to deal with the above issues, leading to limited segmentation performance. In contrast, large-scale vision foundation models such as SAM, DINOv2, OneFormer, and Mask2Former have exhibited impressive generalization performance across natural image domains. However, their direct transfer to medical imaging tasks (e.g., colonoscopic polyp segmentation) is not straightforward, primarily due to the scarcity of large-scale datasets and lack of domain-specific knowledge. To bridge this gap, we propose a novel distillation framework, Polyp-DiFoM, that transfers the rich representations of foundation models into lightweight segmentation baselines, allowing efficient and yet accurate deployment in clinical settings. In particular, we infuse semantic priors from the foundation models into canonical architectures such as U-Net and U-Net++ and further perform frequency domain encoding for enhanced distillation, corroborating their generalization capability. Extensive experiments are performed across five benchmark datasets, such as Kvasir-SEG, CVC-ClinicDB, ETIS, ColonDB, and CVC-300. Notably, Polyp-DiFoM consistently outperforms respective baseline models significantly, as well as the state-of-the-art model, with nearly 9x reduced computation overhead.

Autonomous driving demands robust perception modules trained on diverse scenarios, yet collecting and annotating real-world datasets is both expensive and often lacks sufficient coverage of all possible driving conditions. Semantic Image Synthesis (SIS)---the process of generating realistic images from semantic label maps---has proven effective for producing large-scale labeled data. However, extending SIS to the video domain as Semantic Video Synthesis (SVS), where entire sequences are generated from semantic maps, remains underexplored. We introduce SVS-GAN, a framework specifically tailored for SVS that generates high-quality, temporally coherent videos at a resolution of 1024x512 in real-time (45 FPS). Our approach leverages a deformable motion triple-pyramid generator and a segmentation-aware discriminator to ensure strong semantic alignment and visual fidelity. Through this combination of tailored architecture and loss design, we bridge the gap between SIS and SVS, outperforming state-of-the-art GAN- and diffusion-based baselines on both Cityscapes and KITTI-360. When combined with a semantic-map generator, SVS-GAN enables controllable generation of diverse driving scenarios, providing a scalable source of labeled video for data augmentation and closed-loop testing.

Controllable generative models for images and videos have seen significant success, yet 3D scene generation, especially in unbounded scenarios like autonomous driving, remains underdeveloped. Existing methods lack flexible controllability and often rely on dense view data collection in controlled environments, limiting their generalizability across common datasets (e.g., nuScenes). In this paper, we introduce MagicDrive3D, a novel framework for controllable 3D street scene generation that combines video-based view synthesis with 3D representation (3DGS) generation. It supports multi-condition control, including road maps, 3D objects, and text descriptions. Unlike previous approaches that require 3D representation before training, MagicDrive3D first trains a multi-view video generation model to synthesize diverse street views. This method utilizes routinely collected autonomous driving data, reducing data acquisition challenges and enriching 3D scene generation. In the 3DGS generation step, we introduce Fault-Tolerant Gaussian Splatting to address minor errors and use monocular depth for better initialization, alongside appearance modeling to manage exposure discrepancies across viewpoints. Experiments show that MagicDrive3D generates diverse, high-quality 3D driving scenes, supports any-view rendering, and enhances downstream tasks like BEV segmentation, demonstrating its potential for autonomous driving simulation and beyond.

Split learning emerges as a promising paradigm for collaborative distributed model training, akin to federated learning, by partitioning neural networks between clients and a server without raw data exchange. However, sequential split learning suffers from poor scalability, while parallel variants like parallel split learning and split federated learning often incur high server resource overhead due to model duplication and aggregation, and generally exhibit reduced model performance and convergence owing to factors like client drift and lag. To address these limitations, we introduce CycleSL, a novel aggregation-free split learning framework that enhances scalability and performance and can be seamlessly integrated with existing methods. Inspired by alternating block coordinate descent, CycleSL treats server-side training as an independent higher-level machine learning task, resampling client-extracted features (smashed data) to mitigate heterogeneity and drift. It then performs cyclical updates, namely optimizing the server model first, followed by client updates using the updated server for gradient computation. We integrate CycleSL into previous algorithms and benchmark them on five publicly available datasets with non-iid data distribution and partial client attendance. Our empirical findings highlight the effectiveness of CycleSL in enhancing model performance. Our source code is available at https://gitlab.lrz.de/hctl/CycleSL.

In this paper we introduce ISALux, a novel transformer-based approach for Low-Light Image Enhancement (LLIE) that seamlessly integrates illumination and semantic priors. Our architecture includes an original self-attention block, Hybrid Illumination and Semantics-Aware Multi-Headed Self-Attention (HISA-MSA), which integrates illumination and semantic segmentation maps for enhanced feature extraction. ISALux employs two self-attention modules to independently process illumination and semantic features, selectively enriching each other to regulate luminance and highlight structural variations in real-world scenarios. A Mixture of Experts (MoE)-based Feed-Forward Network (FFN) enhances contextual learning, with a gating mechanism conditionally activating the top K experts for specialized processing. To address overfitting in LLIE methods caused by distinct light patterns in benchmarking datasets, we enhance the HISA-MSA module with low-rank matrix adaptations (LoRA). Extensive qualitative and quantitative evaluations across multiple specialized datasets demonstrate that ISALux is competitive with state-of-the-art (SOTA) methods. Additionally, an ablation study highlights the contribution of each component in the proposed model.

We present mean-shift distillation, a novel diffusion distillation technique that provides a provably good proxy for the gradient of the diffusion output distribution. This is derived directly from mean-shift mode seeking on the distribution, and we show that its extrema are aligned with the modes. We further derive an efficient product distribution sampling procedure to evaluate the gradient. Our method is formulated as a drop-in replacement for score distillation sampling (SDS), requiring neither model retraining nor extensive modification of the sampling procedure. We show that it exhibits superior mode alignment as well as improved convergence in both synthetic and practical setups, yielding higher-fidelity results when applied to both text-to-image and text-to-3D applications with Stable Diffusion.

The generalization of deepfake detectors to unseen manipulation techniques remains a challenge for practical deployment. Although many approaches adapt foundation models by introducing significant architectural complexity, this work demonstrates that robust generalization is achievable through a parameter-efficient adaptation of one of the foundational pre-trained vision encoders. The proposed method, GenD, fine-tunes only the Layer Normalization parameters (0.03 %of the total) and enhances generalization by enforcing a hyperspherical feature manifold using L2 normalization and metric learning on it. We conducted an extensive evaluation on 14 benchmark datasets spanning from 2019 to 2025. The proposed method achieves state-of-the-art performance, outperforming more complex, recent approaches in average cross-dataset AUROC. Our analysis yields two primary findings for the field: 1) training on paired real-fake data from the same source video is essential for mitigating shortcut learning and improving generalization, and 2) detection difficulty on academic datasets has not strictly increased over time, with models trained on older, diverse datasets showing strong generalization capabilities. This work delivers a computationally efficient and reproducible method, proving that state-of-the-art generalization is attainable by making targeted, minimal changes to a pre-trained foundational image encoder model. The code is at: https://github.com/yermandy/GenD