A follow-up to my most popular textual inversion - Desigirl-NSFW. I created that, like I mentioned, because current checkpoints have a hard time calibrating for Indian looks and skin tone. While that textual inversion works fine, it is limited to generating just one skin tone, with the only way of varying the tone being to change weights, but that changes other parameters in the generation as well - including the 'looks' of the characters.
This LoRA, on the other hand gives you the flexibility to prompt for three skin tones through trigger words:
desilatte - this produces light skin-tones - but still something distinct from white or asian tones
desimocha - this produces medium skin-tones - what some might call 'olive' skin
desiespresso - this produces dark skin-tones, but still distinct from black or African skin tones
Since this is trained using a private collection of images of real people, there is an association of skin tone and facial features, but nothing too crazy for the most part.
Seems to work well with most photorealistic checkpoints, though I need to experiment some more.
Best settings, so far:
CFG: 5.5-7
Steps: ~20
Weight: 0.7-1.0
Sampler: euler a
As usual, comments and suggestions are welcome, as are questions. (Except ones on whether I can share the training data set, which I cannot - sorry.)