This version was trained with a more varied dataset of anime and realistic images.
Second version of the Oskar model. This one is definitely getting closer to what I want.