Discriminating Against Unrealistic Interpolations in Generative Adversarial Networks