Training GAN is hard. Models may never converge and mode collapses are common. To move forward, we can make incremental improvements or embrace a new path for a new cost function. Do cost functions matter in the GAN training? This article is part of the GAN series which looks into the Wasserstein GAN (WGAN) and the WGAN-Gradient penalty in details. The equations for Wasserstein GAN look very unapp