All sample imgs were generated by yesmix
It doesn't seem to work very well in the realistic model........
no need to prompt "blonde_hair"
and train more picture
swordmaidenv2.safetensors