Learning to navigate image manifolds induced by generative adversarial networks for unsupervised video generation