This model can learn the high image quality of Japanese women and display it in a condition suitable for SD XL.
We do not use any specific Japanese idols, so randomness is important.
Recommended settings:
Make sure you are aware on the usage instructions of LORA
Trigger: jpn-girl
VAE: none
Sampler: DPM++ 2M Karras (Recommended for best quality, you may try other samplers)
Steps: 20 to 35
Clipskip: 1 or 2
Upscaler : Latent (bicubic antialiased)
CFG Scale : 4 to 9
LORA weight for txt2img: anywhere between 0.2 to 0.8 are recommended
Description
Learning model was trained with sdxl vae fix
FAQ
Comments (2)
Does this work in a negative fashion as well i.e. can you use it in the negative prompt to reduce the chance of getting a Japanese character?
My checkpoint isn't asian centric, but it cannot get anything from this lora in the way of Japanese girl, unlike some of the other Asian loras. (Also I noticed a lot of ppl using 2 or more "asian" loras etc to make their images, which did work for me)