From GAN to WGAN

A model that easily distinguishes generated images from real ones may give an image generator little useful guidance about how to improve. Lilian Weng's GAN tutorial explains this training difficulty: confident classification can coincide with a nearly vanished learning signal.

Wasserstein GAN changes what the judging network learns. It produces scores used to estimate the cost of moving the generated distribution toward the real one, subject to a constraint on how rapidly those scores can change. The original paper shows why that measure can provide useful feedback even when the distributions do not overlap. The contribution is a more informative training objective, rather than a universal guarantee of successful image generation.

“It is inspired by game theory: two models, a generator and a critic, are competing with each other while making each other stronger at the same time.”

“However, it is rather challenging to train a GAN model, as people are facing issues like training instability or failure to converge.”