Adding a normal loss to the training of a generative model significantly slows down the training process, which is not acceptable in practice.