PublicButt LoRA – Specialized in ultra-realistic shots of women from behind or side-rear view in everyday public and semi-public settings (malls, supermarkets, offices, parks, streets, stairs, escalators, etc.). Strong emphasis on perfectly round butts of all types – from slim and toned to thick and curvy – highlighted by tight, figure-hugging clothing such as low-rise/skinny/ripped jeans, glossy or seamless leggings, yoga pants, shorts, mini skirts, bodycon dresses, and similar outfits, paired with casual or sexy tops (crop tops, sports bras, hoodies, sweaters, blouses). Supports fullbody views, close-ups, and shots from farther away. Includes a variety of natural poses such as walking, standing, bending over, squatting, leaning, and sitting. You can also create whale tail/G-string or butt cracks in dynamic poses.
For good quality, use a resolution of 1280x1920, which will give you very good detail. You can check my sample photos to see how best to write the prompts, but you can also test everything yourself to see what works best for you.
Settings : CFG1 Euler/normal or Beta/Simple (Euler/Normal has worked best for me so far) 8-10 Steps Strength 0.8-0.9 is use 0.9
For more colors, slightly better quality, and a clearer image, you can try increasing the CFG to 1.1-1.5
Description
FAQ
Comments (3)
Will this lora make the butt small as that one?
You can choose how big you want the butt to be. If you want a big butt, enter “bubble butt” or “curvy”—the usual terms. It actually works very well with the prompts. Best regards
Big butts are actually all the vocabulary tokens with more than one a in a row. so ./comfy/text_encoders/qwen25_tokenizer/vocab.json is the file for the text embedding model. Do something lik $ cp vocab.json vocab.bak to be sure you keep a backup. Then do $ cat vocab.json | jq > /tmp/vocab.json. That will get the file decompressed and in a temporary RAM directory. Then do $ grep -v \"aa.*\" /tmp/vocab.json | jq -cM > ./comfy/text_encoders/qwen25_tokenizer/vocab.json. The jq package is for unpacking and repackaging json files. The grep command is searching for lines that contain (an escaped) quote followed by two aa's in a row and any number of additional characters before the ending quote. Just skip the pipe to jq to see what it is removing if you want. The -v option for grep is to print non matching lines only. You have a copy that saved that you can swap back at any time with the inverse command, or just use `git restore ./vocab.json from within the directory to reset it with git. Those tokens are the main cause of the "fertility is arousing" bias. The same is true with CLIP if you edit the vocab.json in the sd1 tokenizer file.