Generative adversarial network(GAN) is an active branch of deep learning field, which has become a popular research direction in the field of artificial intelligence. GAN adopts an unsupervised learning method and automatically learns from the source data, which can produce amazing effects without artificially labeling data. In this paper, we present the background, basic idea of GAN and comb its related theory, training mechanism and stateoftheart applications. We also summarize the common network architectures, training skills and model evaluation metrics, and compareGAN with other generative model VAE and GAN variants. Finally, we point out the advantages and disadvantages of the GAN and look forward to the further development direction.