Yu would train it in an RL setting rather than actually use generated images in the train set.