Definition
A GAN pits a generator (creates fakes) against a discriminator (spots fakes). Training pushes the generator toward the real data distribution.
GANs pioneered photorealistic image synthesis; diffusion models later overtook many image-generation benchmarks.
In simple terms
A forger and a detective constantly one-up each other — the forger gets better until the detective can barely tell real from fake.
Where you see it
- StyleGAN face and art generation.
- Data augmentation in research settings.
- Image-to-image translation (pix2pix, CycleGAN).
How it works
1.Generate fakes
Generator maps noise to candidate samples.
2.Discriminate
Discriminator scores real vs fake.
3.Adversarial update
Both networks improve until equilibrium.
Why it matters
- GANs proved deep generative models could produce stunning samples — shaping the generative AI wave.
Often confused
All generative AI uses GANs.
LLMs use next-token prediction; modern image models often use diffusion.