test training on limited colors dataset, most of the image were manga panels

upd:

now require a trigger word ("yozo_stanky")

v2.0 should support coloring now, it is the artist's old style, and require careful prompting to get the correct color (highres and absurdres from base model could outstrength the lora, for example, you could also get his new coloring style through prompting, though i haven't find a way to consistently produce it)

(old: the manga style came out fine (though a little overfitted)

i don't really like the coloring style yet, sometimes it produce the correct style, but

most of the time it just the coloring style of the base model)

use with strength ~0.8