SEC(

Software Engineering Courses (SECourses)•3y ago

mikemenders

I just trained a Lora model on RealisticVision with 32/32 (NR/NA) and Prodigy with these parameters:

I just trained a Lora model on RealisticVision with 32/32 (NR/NA) and Prodigy with these parameters:
decouple=True use_bias_correction=True safeguard_warmup=True

It's amazing to achieve professional results with such low weights on a SD 1.5 model. Here is a sample (original and hires fixed):