data-archetype/ffhq_captioned_1024

ffhq_captioned_1024是一个包含70,000张方形人脸和肖像图像的数据集，每张图像都有相应的文本描述。这些图像来自FFHQ数据集，经过解码、RGB转换和高质量JPEG重新编码。描述由Gemini 2.5 Flash Lite和Mistral Medium 3.1生成。数据集主要用于文本到图像训练、图像文本表示学习等场景。技术细节包括bucketed-shards存储格式、图像预处理流程（如EXIF转置、RGB转换、调整大小和裁剪）、桶分布和描述选择策略。数据集还提供了高效加载方法，支持Python的webdataset和tarfile库。

ffhq_captioned_1024 is a dataset containing 70,000 square face and portrait images, each accompanied by a text caption. The images are sourced from FFHQ, decoded from the original dataset, deterministically converted to RGB, and re-encoded as high-quality JPEGs. Captions were generated using Gemini 2.5 Flash Lite and Mistral Medium 3.1. The dataset is intended for text-to-image training, image-text representation learning, and other similar applications. Technical details include the bucketed-shards storage format, image preprocessing steps (such as EXIF transpose, RGB conversion, resizing, and cropping), bucket distribution, and caption selection strategies. The dataset also provides efficient loading methods, supporting Pythons webdataset and tarfile libraries.