The training set is mostly monochrome and comics. Maybe it will be better to generate gray pictures.
More than 300 pictures trained for 40,000 steps
hanabi(ponyv2).safetensors