license: openrail
tags:
- not-for-all-audiences
What is this?
This hosts a collaborative effort to clean up the mess that is the danbooru dataset. The Danbooru dataset is a wealth of (mostly) free anime style images, that are already individually tagged! The problem being that the images are indiscriminately included. Some pics are explicitly copyrighted and shouldnt be in it. Some have watermarks. Some have legally questionable subject matter. Some just frankly arent good. But the rest... are really good
So, this is a public effort to pull out just the clean, AI-training-usable images. Note that we do not remove a signature or watermark from an image. We just exclude the image entirely from the set.
We are working from https://huggingface.co/datasets/animelover/danbooru2022/
which has the advantage of containing images that have already been somewhat appropriately resized. So now we copy a block of them, and delete the "bad" images. One zip file at a time. (See the "STANDARDS.txt") file, for suggested criteria for deleting images.
There are literally hundreds of blocks of data, each block having perhaps 4000-5000 images.https://huggingface.co/datasets/animelover/danbooru2022 So this is an organized, crowd-sourced volunteer effort.
Training tips:
While he images should be "usable" for AI training, and they are all individually tagged with .txt files, if you just dumped them all into a trainer, the results would probably not be coherent. You probably will want to take an image directory browser, and select a group of them for either subject matter, or style, and pull those specific ones out for training.
You may also want to add in your own custom tags, on top of the ones that are already there.
Currently spoken-for segments:
- 0000-0010: ppbrown (data-0001 - data-0003 complete)
- 0010: complete
- 0040: complete