From GAN to WGAN

Discriminator: tries to detect if an image is from the generator or from the legitimate test set. Error gradients from this are passed down to the generator.

Generator: tries to generate images from an input feature set.

Can be unstable. Discriminator's errors are used to train the Generator. A perfect discriminator has no errors to train the generator with (vanishing gradient.)

Adding light noise to discriminators while training seems to be one way around ensuring there is always some error to transmit.