This model was trained using images from nonco for NoobAI-XL (0.5 epsilon). Trigger word nonco is optional, but it does make the style more pronounced.
I really like nonco's style and anatomy. I was going to release this for pony initially, but then illustrious noobai came out. This is the same dataset as the Illustrious-XL 0.1 DoRA.
Note: Gens on civit look borked, you might have to reduce strength. I use forge, and it doesn't have issues with the strength.
Changelog:
v9 uses 57 color images (manga & booru) and 26 black&white images (manga); labelled with wd1.4 and manually cleaned
Description
DoRA trained on 83 images with kohya_ss (NoobAI-XL 0.5 epsilon)
optimizer - AdamW
lr - 2e-4
unet lr - 1.5e-4
text lr - 4-e4
lr scheduler - cosine
gradient accumulation - 4