There are reasons why it would be better at generating some things better than others in a way that’s roughly proportional to the disparity of training data volume used in the model.