language:
- en
size_categories:
- n<1K
Anime Dataset 02
My second anime dataset. What you will find here is, the original images unmodified, and square cropped images that are NOT downscaled. Depending on which folder you go into, the captions will either be empty, all a single word / phrase, or the output of JoyTag. Note that since many images are very large in resolution, 2 cropping passes were done, pass 1 results in the majority of the dataset and usually captures the desired features, like faces and bodies. Pass 2 is done on the resulting parts that were cropped out from the images originally, only if they are large enough and worth it. This might be a detriment to the dataset in the JoyTag situation because it makes it harder for JoyTag to be correct, but we'll see. Almost all of these images were pulled from Gelbooru, with an exception of 1 or 2 that I downloaded during testing. They were all manually reviewed by myself, that is to say I only included images downloaded from Gelbooru that I thought were good, don't worry most of the ones downloaded were diabolical and deleted. If you are curious these images were pulled from the following id range on Gelbooru, 9,000,000 - 9,001,000 Why 9 million, well when I was around 1 million everything was just trash, so I moved it up to more recent posts.
Tools Used
To download, crop, caption, and rename these images I used a set of scripts I made, along with the JoyTag model. Here was the overall workflow. Download -> manual review by myself, deleting bad pictures -> download more and review until I have enough good ones -> crop -> review crops -> second pass crop -> review again -> rename images -> caption -> upload to HGF
Usage
I mean, the art is not mine, and I do not claim ownership of it by any means, so do with it as you will, I just processed it and maybe made your life easier. If you do use it though on some training, and you get good results, let me know, I would be interested.