>>82
>Alright, seems good enough for /e/ material but very bad for /h/ smut (all of this kind of expected given the type of dataset you have for the style)
I would like to have some example boxes to investigate those /h/ fails
>I would like to mention that more than often I couldn't get a pure simple background out of it even if prompted correctly (I got random things on it, text and signs more than anything) but I'm not sure if it was because the model fault or some issues with the lora
This is just the main problem with training vpred lora using kohya sd-scripts, no matter what settings I set, pure black/white just breaks, I just monkey patched it, adding main colors as a dataset part and it works somewhat stable, guess not with all styles, especially bad with those, which doesn't have black/white bg sovl, but would like to see boxes of your fails too
>Btw, I would like to ask you a favor, I would like to know if you can bake some loras for me
Sure can do