updates for v2 on my lora im trying to train, i basically x3 my image training data (100+ now, basically all from transparent bg high resolution vsts screenshots vtuber dumped on twitter for memes, i also improved my tagging big time since the first training attempt, tried to describe more accurately her outfit / hairstyle / etc., first attempt i just basically rolled with whatever hw tagger rec'd for me
what got better:
* outfit accuracy drastically improved, i suspect my much better tagging this time around likely played more of a role in this than my training data though
* hair ornaments/braid placement now consistent
* turning off flip augmentation fixed left/right confusion
problems:
* still cant produce the eye, tried some more unique tagging (gear-shaped pupils for the eyes, but it still doesnt come close to genning it right, even for eye-focused gens, despite me feeding it much better quality eye images directly from the vtuber's model) heres the eyes i want:
https://files.catbox.moe/a3u94p.png and this is the eyes i get on an eye-focused gen:
https://files.catbox.moe/zh7pil.png* cant generate the pendent looking thing on her chest, guessing probably cause i only gave it a generic "pendant" tag in the trianing images
* think it's overtrained / overfit now, generations look like literal screenshots of the vtuber, cant really get it to do any different styles, not super suprised, 75% of my training is now the vsts screenshots
* it doesn't allow me to change the gens at all unless i reduce the weights down by like 60%, different checkpoints/style tags dont seem to help
think i might try reducing my training images by half, only keeping the best ones or something, and refine my tagging for more unique tags...
does anybody know if i should be tagging things unrelated to the vtuber in my training images? e.g, i still kept like 50% of the ai-generated tags that hw tagger made for me, e.g., lots of "simple background" tags
