Introduce:
Model has been fixed Asian face and enhanced Asian-****, tried your prompt to create beautiful highly realistic nudity Asian girls.
The Base model DOESN'T incorporate any LoRa. (Prevent a decrease in LoRa performance.)
If you want good LoRa performance, need to retrain LoRa.
Model has been enhanced realistic, if you want to render animated characters into real-life cosplay, you can use the official model training LoRa.
Recommend:
Train LoRa using the base model and accelerate image generation using the LCM version.
In the case of the sampler Euler a/LCM, only 8 steps can be used to generate clear images. (CFG is set to 2.5 or less.)
When multiple prompts (>75 tokens), facial features become unstable. Recommended to enable After Detailer yolov8n_v2 plugin to keep an Asian feature.
Parameters setting:
Sampler: Euler a/LCM
Steps: 6~10
Denoising strength: 0.2~0.5
Resolution : fixed one side 1024 (eg. 1024×1024, 1024×1280,768×1024)
CFG scale: 1.5-3
clip skip: 2
Does not require VEA.
And Enjoy it!
Prohibited images:
❌Violent expressions
❌Excessive sexual
❌Child pornography
Base formula:
0.1ver(unreleased):
Fixed asian face.
-Addition 800 pictures -Training Steps 150,000
(Test model, fine-tuning by SDXL_Dream_XL_v2.0 )
0.2ver(unreleased):
Enhanced asian-****.
-Addition 1200 pictures -Training Steps 150,000
(Test model, fine-tuning by afroditexl_11Bkdvae)
Base 1.0ver:
Mixed 0.1ver and 0.2ver and afroditexl
I have tried numerous mixed combinations to adjust the model generation effect.