I have mostly been happy with the way tattoos generate on Krea 2 models. If you give detailed prompt they will often produce something close to what you want.

The idea of this LoRA is to improve prompt adherence and hopefully realism.

 I wanted to say thanks to @lonecatone23 who gave me a nudge to get this going!

 This LoRA is generated from 165 AI generated images. Some of tattooed women, some effectively “flash” of close-ups. The images have been generated on a mix of SDXL, Flux 1 and Krea 2 models over a few years.

Images and styles in the dataset include Irezumi/Japanese, “traditional American” and blackwork.

 The LoRA works with the others I have tried it with.

Usage, for this Krea 2 model a weight of 0.8 - 1.0 seems to work well I am mostly using a weight 0.8 as 0.8 softens the tattoos a little (mostly) for a more realistic feel.

I use er_sde for a sampler and simple for the scheduler.

Detailed prompting usually gives better results. Just saying “tattoos” often generates generic flower tattoos.

Anyway I hope this is of some use, and as always have fun!

Any resemblance to real persons, living or dead, is purely coincidental and unintentional.