I wanted to expand the diversity of my collection. So, I've added some East-Asian flair with Mei.
Description
FAQ
Comments (4)
Curious to your methodology, if I may ask?
Do you gen off model (Say with a ZIT/SDXL) to get your face, then train in krea to better lock in the particulars?
The realization I had was that each model has it's own definition of asian, with none of them really understanding the difference between the types of asian. I could prompt for Chinese, Vietnamese, Japanese (et al) and see the exact same facial features, which is bizarre because those people all look different.
Yes, I use an SDXL megamerge (lost track of which models are in there) + a ZIT or Krea2 as refiner, then I train on Krea2 RAW. SDXL is an imperfect tool but it does have much more ethnic diversity than Krea or ZIT as the latter have a strong tendency to produce the "generic Asian". But most models were trained on data where some ethnic groups are overrepresented, so they render those with lots of individuating detail, and underrepresented ones get pulled toward a smoother, more averaged type. My solution is to alternate/mix tokens, explicitly describe facial features, and, because I'm using an SDXL-based checkpoint, use strong negative prompting.
@dr0s Appreciate the answer, and love your work, Thank you
@badhandproductions Thanks!