Drop some buzz, leave ideas, leave your images
This lora is meant to create body painted imitation clothing that resembles real clothing and patters painted onto the skin. It is not 100% foolproof, but it does work overall. It took about 5 training attempts to get it working, so I hope you enjoy it (It was a really frustrating result).
Work the various trigger words into captions to describe the imitation clothing and to get it to work. See image prompts from image showcase for additional examples.
There are four trigger phrases:
"Imitation clothing"
"body paint”
"2-D cell-shaded clothing"
"flat clothing"
Penis/Testicle States:
Can do flaccid, or erect penis, you can leave the penis and testicles unpainted using "unpainted penis" or "unpainted testicles". The color of genitals can be different than the other imitation clothing.
Nipples:
It does not always like to paint the nips. add "(color) painted nipples" to get it to cover them, match the color to the color around it.
Women:
There are feminine bodies in the dataset, but they are much less focused on, they should work (see example image). Just don't tag the Vagina color, it isn't specialized in that as it is peens, just use "painted vagina".
For some checkpoints they try and make the design messy for some reason, sadly you might get the same results, try a new checkpoint. It works really well with the FASCIUM KREA2 checkpoint:
https://civarchive.com/models/2745073/fascium-krea2?modelVersionId=3248159
Struggles:
Collared outfits usually still try and make a collar (see example image), you can try tweaking the prompt, or avoiding that outfit, sometimes it doesn't make that bit 3-D, the same with long sleeve jackets, (shrug, I can't keep fighting it, its already been like 6000 buzz to get it working).
Description
FAQ
Comments (5)
Fantastic concept & execution!
Thank you! I literally got so mad trying to get it to work. I hope you enjoy it!
@DerekG2 I can imagine. When I saw it, I thought to myself that the caption strategy must have been a doozy with a large image dataset to get it done right or it was just over trained and brittle to technically achieve the effect. Then after I did one of my regular LoRA stacking tests (usually that will show how brittle or over trained it is), I was impressed that it worked like a champ. I thought about doing something like this but honestly I thought the concept would be so much of a headache to get it right - I'm glad you did!
@jrewingwannabe947 honestly the activation words hold it together, has to tweak the captioning four times. realized it just REALLY hated the checkpoint I was using...which felt like super betrayal.
Fascinating, thanks!


