Thank you!

You might also be interested in my Doki lora with 4 girls-

I'm used to doing multiple outfits by now, and humans are kind of like outfits, but made of flesh :D

I've also done 10 outfits/folders in 35 MB with my Rushia lora, I haven't been able to replicate it...

dim16 alpha8 can go a long way as you can see. It needs higher learning rate: 5e-4 unet, 1e-4 text encoder. For characters, cosine with 3 restarts is good. min_snr_gamma to 5 is also very good. I separate my outfits into folders with different activation tags, and it works the same for characters. For each character, get an amount of images that divides 100 (20, 25, 33, 50 or 100) and give that folder enough repeats to get to 100 (5, 4, 3, 2 or 1, respectively). Then finally, 10 epochs and 2 batch size should do the trick. More than 100 images per character is bad (200 tops), so go through them and pick the best ones. 50 great images is probably better than 100 mediocre images. Less than 20 images will start to hurt, don't try less than 10. Pick images that accurately represent the character's face and body features, and 30%-80% of the images should accurately represent their official outfit(s). Pick images with varying angles, poses, clothes, and artstyles; include nudes but probably not sex. The images should be tagged by the WD1.4 Tagger AI and the tags can be later mass-adjusted to get the best results. More on that another time.

You divide the characters into folders, each folder will have its own activation tag, and the repeats of each folder should be such that the total number of iterations is the same for all characters. The characters should have very little tags in common lest you confuse the ai, for example, if they wear different jackets don't tag them all as "jacket", do "jacket", "coat", "cardigan", etc. This applies to different outfits too.

You divide the characters into folders, each folder will have its own activation tag, and the repeats of each folder should be such that the total number of iterations is the same for all characters. The characters should have very little tags in common lest you confuse the ai, for example, if they wear different jackets don't tag them all as "jacket", do "jacket", "coat", "cardigan", etc. This applies to different outfits too.