The training images for this were generated by mixing up an auto-caption and using it as a prompt/training tags.