Learn AI · Level 2 · How AI works · step 13 of 18 · word 2 of 3
What is a GAN (generative adversarial network)?
A GAN is a pair of neural networks that compete. One makes fakes, the other tries to spot them, and each round the fakes get more convincing.
In plain English
Think of a forger and a detective. The forger (the generator) makes fake pictures. The detective (the discriminator) is shown real and fake ones and guesses which is which. Every time the detective catches a fake, the forger learns to do better, until the fakes are hard to tell from real.
An example
Ian Goodfellow and colleagues introduced GANs in 2014, describing them as a two-player game between the two networks. Within a few years, GANs could make photos of faces of people who don’t exist, and they powered many early deepfakes.
Watch out
GANs are hard to train and tend to make only a narrow range of images. Most of today’s image generators use diffusion models instead, though GANs are still used where speed matters, such as sharpening video in real time.