GANs: generator versus discriminator

In short: A GAN consists of two networks that compete: the generator produces fakes, the discriminator tries to tell real from fake data. Both become better and better as a result.

In more detail: Ian Goodfellow presented the idea in 2014. The picture: a forger and an art examiner train each other.

In detail

• Process: the generator receives random noise and produces an image. The discriminator rates real/fake. The generator is rewarded when it deceives, the discriminator when it is right.
• Strengths: very sharp images, fast generation in one step.
• Weaknesses: unstable training, “mode collapse” (the generator produces only a few variants), hard to control.
• Known models: StyleGAN (faces), CycleGAN (style transfer), Pix2Pix.
• Replacement: diffusion models are more stable and better controllable and have largely replaced GANs for image generation.
• Deepfakes: face and voice swaps use GAN-like methods. Misuse is hard to detect, see Deepfakes and AI detection.

See also: Autoencoders and VAE, Image generation with diffusion models