Wasserstein GAN (WGAN)

A Wasserstein GAN, or WGAN, is a generative adversarial network designed to make training more stable by using a distance measure related to the Earth Mover distance. A generator creates samples while a critic scores how closely generated and real data distributions match. Unlike a traditional GAN discriminator, the critic does not simply classify samples as real or fake. The resulting training signal can remain useful even when the distributions barely overlap. WGAN variants are used for images, audio, and synthetic data, but they still require careful architecture choices, regularization, evaluation, and checks for bias, memorization, and unrealistic outputs.