Commisions open

~no, i use self writen automatization for it(no sharing sorry), but manually delete bad frames after it. And this over 900 imgs is like 90% of good not dublicated frames of her from anime (maybe even more). And yes, all frames from anime, no arts. (in fact this process better than i realistically can made by hands, i rewrite code for make it ~perfect for my tasks like 2 months, manually make more than 100 images and don't make mistakes is hard)

@Dineko 1) because i make money on it. and i made it for increase quality, it not really for make faster/easier (for make faster you can just manually get 30 random good frames of character, it be faster than balance 900 frames, delete bad ones randomly leaks to folder, and clean/make tags);
2) because for make scripts/code public you need made more work than for do it privatelly. It just not writen for user friendly + hard to install. And i absolutelly don't want to spend time on it.

P.S. like 60% functional is this https://github.com/deepghs/imgutils perfect library from team who made bangumibase datasets. Just with custom setiings + some additional code for make things that i think necessary from my experience (like use/crop to resolutions that perfectly fits in buckets, for lora don't train on wrong aspect ratios, or heavy parallelism(many tasks like tagging, or even cropping easilly works in 64+threads in parallel, that much faster(i'm now literally analyze every frame from anime, cuz i can))).