CounterSushi - Stable Cascade版本CounterSushi v0.3 (ID: 365424)

this is a finetune for anime style with stable cascade (stage_c lite)

my goal is to make similar style to counterfeit and darksushi for stable cascade

for training i followed this from github

0.1:

i traned for 5k steps with ~300 images from counterfeit then i trained another 5k steps with ~350 images from custom merge model that i made a while ago

v0.2:

trained with more images ~750

recommended for better quality: use compression=32

negative:

low quality, realistic - sometimes it makes the image worse, sometimes it helps

bad anatomy, deformed - generally helps

positive: i've posted the prompt template used when training in trigger words, but you can try whatever prompt you want (duh)

nsfw: please don't try to generate nsfw, i trained with ~100 nsfw images but it looks bad, i'll add more in the future

v0.3:

retrained from scratch

my experimental settings so far:

cfg: 2-2.5 seems good

5 steps for stage_b, 10 if you want a bit more detail

negative:

cartoon, closeup, lowres, simple background, artifacts, (realism, blurred:1.2)

(should be better than nothing, but feel free to test with other negatives)

cartoon in negative seems to add more detail to the image

4k, masterpiece, high quality doesn't make the image better

compression=42 is good

if you want the character to look more manly, put 1girl, girl in negative, it could help

retrained:

>1k 1024x1024 for 15k steps

650 1024x1792 for 7k steps

328 1536x1024 for 5k steps

129 1024x1280 for 2.5k steps (which are 768x when training)

so the model can generate images with different sizes,

my idea with the last one is that when characters occupy a relatively small portion of the overall image or appear at a distance the model will know how to generate them (cause it's seen how to make characters in small portions of pixels), dunno if it actually works

描述:

训练词语: