Testing a lora for different curl types. Probably need to refine the tagging and maybe not try and lump the full 2A - 4C into one lora but this is the starting point.
Works well with Ebony Bliss
Description
same images but removed all tags that didn't pertain to hair.
FAQ
Comments (11)
VERY promising for a first concept! There's some style impact towards realism, but otherwise I'd say it's already working really well with a variety of skin and hair colours and even different hair types on different characters in the same image ^^
One thing I've noticed is that tagging dark skin and black hair really strengthens the effect of the heavier curl types while the effect can be pretty weak otherwise, which makes complete sense because there can't be many pictures of people without dark skin in the training data, but something to note nonetheless 😅
Thanks for the feedback!
@Kandrosaurus Correct assessment, once you get out of the 2C range of the sample data was of folk with dark skin.Â
@Kandrosaurus New version less realistic can do no dark skin with prompting
@guythis31773Â Awesome, I'll give it a try and post some images of my results!
@Kandrosaurus coiled, tightly, loosely, curled various combinations of those words make different curl patterns, 2ACurls, 4BCurls doesn't do much tho.
@guythis31773Â I think something's gone wrong with V2, my UI is showing that it's for SDXL rather than anima and it's having zero impact on image generation - I mean literally outputting identical images with the LORA enabled and disabled.
@Kandrosaurus Yup don't know how I pulled that off and my SD Forge setup doesn't make same image with or without the lora so I didn't notice. I'll post a properly trained model as soon as it is ready. Sorry for the inconvenience and thanks for the help and feedback!
@guythis31773Â Not a problem at all, looking forward to it!
@Kandrosaurus This might be as good as I can get this. Tried a 4C only with higher res training data and it came out worse than what I have posted now.