opendiffusionai/laion2b-squareish-1536px

这是一个从laion2b-en-aesthetic数据集子集中提取的数据集，用于一般情况下的（摄影）模型训练。所有图像的宽高比在1到5/4之间，可以安全地自动裁剪成方形，且所有图像的分辨率至少为1536像素高。数据集大约包含80,000张8k分辨率的图像。图像内容主要是现实世界的照片，大约有15,000张包含人类的图像。图像经过了筛选，去除了不清晰的图像、儿童图像、黑白图像以及大部分裁剪掉人头一半或全部的服装宣传照片。图像的标注是自动生成的，来源于AI moondream(2b)，也提供了一种由最新WD AI模型生成的标签风格的标注。

This dataset is a subset of the laion2b-en-aesthetic dataset, intended for general-case (photographic) model training. All images have an aspect ratio between 1 and 5/4, allowing for safe auto-cropping to square, and each image is at least 1536 pixels tall. There are approximately 80,000 8k resolution images. The images are primarily real-world photos, with about 15,000 containing humans. The images have been filtered, removing unclear images, images of children, black and white images, and most clothing promo shots that cut off half or all of the human head. The captions are auto-generated by the AI moondream(2b), and an alternative tagging-style set of captions is also provided by the latest WD AI model.