DiffDoctor
收藏资源简介:
DiffDoctor数据集由香港大学、同义实验室、蚂蚁金融服务集团和浙江大学的研究团队创建,旨在解决图像扩散模型生成图像时出现的伪影问题。该数据集包含超过100万条有缺陷的合成图像,涵盖了形状扭曲、不合理内容和水印等多种伪影类型。数据集的创建过程采用了人机协作的标注策略,确保标注的准确性和类别的平衡。该数据集的应用领域主要集中在图像生成模型的优化,通过像素级的伪影检测和反馈,帮助模型减少生成图像中的伪影,提高生成图像的质量和稳定性。
The DiffDoctor Dataset was created by research teams from The University of Hong Kong, Tongyi Laboratory, Ant Financial Services Group, and Zhejiang University, aimed at resolving artifact issues in images generated by image diffusion models. This dataset contains over 1 million defective synthetic images, encompassing multiple artifact categories including shape distortion, nonsensical content, watermarks, and others. A human-machine collaborative annotation approach was employed during the dataset's construction to ensure annotation accuracy and category balance. Its primary application lies in the optimization of image generation models: through pixel-level artifact detection and feedback, it assists models in reducing artifacts in generated images, thereby improving the quality and stability of the output images.

- 1DiffDoctor: Diagnosing Image Diffusion Models Before Treating香港大学, 同义实验室, 蚂蚁金融服务集团, 浙江大学 · 2025年



