Another attempt at a coherent model for characters.
This is a bit over-trained, and over-fit.
Strengths of less than 1.0 work best.
993735_training_data.zip
mst3cStyle.safetensors