Can you put an thumbs up / down counter for each image and use the human preferences collected to refine the model further ?