opendiffusionai/cc12m-4mp-realistic

这是CC12M数据集的一个子集，专注于高质量、大尺寸（至少4百万像素）的真实世界图像。这些图像附有长或短两种风格的描述。数据集目前包含约21,000张图像，经过人工筛选，不良图像比例约为0.1%。数据集排除了杂志封面、海报、黑白图像、绘图、对焦不良、水印、颗粒感、经过Photoshop处理的图像以及视频游戏图像等。此外，数据集还提供了一些特殊子集，如仅包含单独女性的图像和仅包含单独男性的图像，以及Parquet格式的数据文件。

This is a subset of the CC12M dataset, focusing on high-quality real-world images with a resolution of at least 4 megapixels. The images come with captions in long or short style. The current subset includes only images described as A man or A woman, totaling around 21,000 images. The images are human-filtered for high quality, excluding magazine covers, posters, black-and-white images, drawings, blurry images, watermarked images, grainy images, Photoshopped images, and video game images. The dataset provides a download script and specific subset files, such as woman_jsonlgz for images of women only and man_jsonlgz for images of men only. Additionally, a converted version in parquet format is available.