chinoll/Danbooru2024-Webp-4MPixel-NL

Danbooru2024-Webp-4MPixel-NL数据集是deepghs/danbooru2024-webp-4Mpixel数据集的扩展，专门为大约780万高质量图像提供了自然语言描述，这些图像来源于官方Danbooru平台。每个图像都搭配有一个使用fancyfeast/llama-joycaption-alpha-two-hf-llava模型生成的详细文本描述。该数据集采用了过滤规则，确保只包含长宽比在0.5到2.5之间的图像，以保持描述的一致性和相关性。数据集以joy_caption.tar.gz格式打包，可以通过WebDataset格式访问。

The Danbooru2024-Webp-4MPixel-NL dataset is an extension of the deepghs/danbooru2024-webp-4Mpixel collection, specifically curated to provide natural language descriptions for approximately 7.8 million high-quality images sourced from the official Danbooru platform. Each image is paired with a detailed textual description generated using the fancyfeast/llama-joycaption-alpha-two-hf-llava model. The dataset applies a filtering rule to ensure that only images with an aspect ratio between 0.5 and 2.5 are included, promoting consistency and relevance in the descriptions. The dataset is packaged as joy_caption.tar.gz and can be accessed using the WebDataset format.