Trained at 768x768 from 91 images. Strong model consistency across weights which allows additions and colors to be changed.
