not perfect, but works well enough, i'll hopefully update this as soon as i'm a bit more certain how the hell flux works.

you seem to have to write a whole short story for a prompt...

also even though the dataset isn't even containing that many bare breasts, flux likes to put it's signature blob-nipples everywhere it pleases, prompted or not. you might have to write a whole second short story to get flux to cover up those things.

also partly to remind myself:

dataset 300 images

tagging with cogvlm, one sentence

learning rate: 0.0001

steps around 2000