In this paper, we present a systematic study on GANs with categorical discriminator, especially their impact on the optimization scheme of the generator. We derive class-aware gradients and cross-entropy decomposition, to theoretically reveal how they help GAN training and the inherent problems in previous models. Based on the analysis, we propose an advanced model AM-GAN, along with an interesting dynamic labeling mechanism. We show mathematically that the proposed AM-GAN is a general one covering several major existing solutions that exploit categorical discriminator. Empirical experiments demonstrate the effectiveness of the proposed method, with state-of-the-art sample quality and fast convergence.