This version uses the new train-text-encoder setting and improves the quality and edibility of the model immensely. Trained on 95 images from the show in 8000 steps.