Moviegan

They trained a GAN on a dataset of over 2 million frames extracted from unlabeled movie trailers (specifically "sunny," "rainy," and "action" scenes).

1. Introduction: Beyond Still Images Generative Adversarial Networks (GANs) have revolutionized artificial intelligence by enabling machines to create hyper-realistic images, from fake human faces to artificial landscapes. However, video generation remains a significantly harder challenge. A video is not just a collection of images; it is a sequence of images bound by temporal coherence —the property that consecutive frames must flow smoothly and logically. moviegan