GANs pit two neural networks against each other in a genuinely clever adversarial setup — the mechanism behind a wave of realistic AI-generated imagery before diffusion models took over as the dominant approach.
The Adversarial Setup
Generator: creates fake images from random noise, trying to fool the Discriminator Discriminator: tries to tell real images apart from the Generator's fakes Both networks train simultaneously, each getting better in response to the other: - Generator improves at fooling the Discriminator - Discriminator improves at catching fakes - This adversarial pressure pushes the Generator toward increasingly realistic output
Why This Setup Is Genuinely Clever
Neither network is trained against a fixed target — they’re trained against each other, in a constantly shifting equilibrium. This is what let GANs generate images with a level of realism that direct, single-network approaches struggled to match at the time, since the discriminator effectively serves as an automatically-improving, adaptive critic rather than a fixed loss function.
Why Diffusion Models Largely Superseded GANs
GANs are notoriously unstable to train (a phenomenon called “mode collapse,” where the generator finds a narrow set of outputs that reliably fool the discriminator and stops exploring further) — diffusion models proved more stable to train and scale, and largely displaced GANs as the dominant approach for state-of-the-art image generation by the early 2020s, though GANs remain relevant for specific applications like fast, real-time generation.