I discovered the optimal training parameters in SDXL 1.0. The generated LORA now resembles a real person in real life.
Train based on SDXL 1.0
Dataset using over 100 carefully selected high quality images
The prompt must include: ichika
Best settings:
weight: 0.8 - 1
sampling steps: 25-40
sampling method: DPM++ 2M SDE KARRAS or DPM++ 2M KARRAS
CFG Scale: 7
Visit https://ko-fi.com/ai_imaging/shop for more exclusive popular idols
*****************************************************
Get access to numerous exclusive hand-picked LORA for free by joining our membership for only $5 per month! Visit https://ko-fi.com/ai_imaging to sign up