opendiffusionai/laion2b-45ish-1120px

这是一个从LAION2B-en-aesthetic数据集中选取的子集，基于宽高比和更好的标题进行选择。这是一个通用案例数据集，大约只有30%的图片包含人类。此外，所有非真实图像都被过滤掉了，目的是创建一个“现实世界”的数据集。数据集中的图像数量大约为80k，磁盘上的大小约为45G。这个数据集不是由人工单独过滤的，只是批量筛选。数据集中包含了两种自动生成的标题风格：moondream和wd14。使用了这两种风格来过滤带水印的图片，并使用moondream模型为每张图片添加了AI生成的标题。

This is a subset of the https://huggingface.co/datasets/laion/laion2B-en-aesthetic dataset, selected based on aspect ratio and better captioning. It is a GENERAL CASE dataset with approximately only 30% of images containing humans. All non-realistic images have been filtered out, aiming to create a real world dataset. The dataset contains around 80k images, with an on-disk size of about 45G. The dataset is not individually human filtered but batch culled. It includes TWO auto-generated captioning styles: moondream and wd14. These styles were used to filter out watermarked images, and the moondream model was used to add AI-generated captioning for each image.