Complexity Matters: Rethinking the Latent Space for Generative Modeling

Our investigation starts with the classic generative adversarial networks (GANs). Inspired by the GAN training objective, we propose a novel "distance" between the latent and data distributions, whose minimization