Archived: This model was an experiment on a 1 byte precision model and do to lack of feedback and interest has been archived.

Epic Porn HD 1GB (SD1.5)

• This is a Full EMA fine tune at 512

• A second pass non EMA at 768

• Example images are the first or second click on a prompt without complex negatives.

• USE Natural language - Not Danbooru (Danbooru can be used to enhance)

• Remember SD is limited to 75 words

Why make a FP8 model with only 2bits of precision vs 10bits on FP16 you may ask. To prove how much the ability to finetune at highest level effects quality of outcome.

Codeformer/GFPGAN are off by default now and stayed off for these images, but the faces would be crisper with it on. Image to image up scaling will also clean up the image.