laion2b-45ish-1120px
收藏资源简介:
这是一个从LAION2B-en-aesthetic数据集中选取的子集,基于长宽比进行了筛选,并且具有更好的标题注释。这是一个通用型数据集,其中大约只有30%的图片包含人类。所有非现实风格的图片都被过滤掉了,旨在创建一个“现实世界”的数据集。这个数据集大约有80,000张图片,磁盘占用大约为45G。这个数据集没有经过人工单独筛选,只是批量剔除。图片的长宽比为4:5的肖像画风格,或者稍微宽一些(因为稍微太宽的图片可以被安全地裁剪得更窄)。作者对于选择1120像素的最低高度可能有所偏差,建议可能需要选择1152像素(因为64的倍数)。数据集中包含了两种自动生成的标题风格:“moondream”和“wd14”,并使用这些风格来过滤掉带水印的图片。使用“moondream”模型为每张图片添加了AI生成的标题,并使用该模型过滤掉带水印的图片。
This is a subset selected from the LAION2B-en-aesthetic dataset, filtered based on aspect ratio and featuring improved caption annotations. It is a general-purpose dataset, with only approximately 30% of its images containing human subjects. All non-photorealistic images have been filtered out, with the goal of creating a "real-world" dataset. This dataset contains roughly 80,000 images and occupies approximately 45 GB of disk space. This dataset has not undergone individual manual screening, only batch filtering. The images have an aspect ratio of 4:5 (portrait orientation), or slightly wider, as excessively wide images can be safely cropped to a narrower aspect ratio. The authors may have a bias in choosing a minimum height of 1120 pixels, and it is recommended that 1152 pixels (a multiple of 64) be selected instead. The dataset includes two automatically generated caption styles: "moondream" and "wd14", which are used to filter out watermarked images. AI-generated captions were added to each image using the "moondream" model, and this model was also used to filter out watermarked images.




