This is lora is a personal experiments, it's bad...

I recommend using hassakuXLIllustrious_v21fix with dpmpp 2m karras (~3.5 CFG, ~35 steps) with hires fix, if you want to try this lora.

notes:

• str = 0.75-1.25;

• trained with illustriousXL_v01;

• trigger = yuming li artstyle;

• trained with fp8 unet.

logs:

• v2 -> r=α=16, size=~22mb (~20x smaller than v1), target="^.*to.(q|v).*$", less img training data than v1;

• v1 -> r=α=256, size=~500mb, target="^.*(to.(q|v)|ff.net.0.proj).*$".