Loading…
Glossary term
A pair of networks, a generator and a discriminator, trained against each other to produce realistic data.
A GAN trains two networks against each other. The generator makes fake examples; the discriminator tries to tell fake from real. As each improves, the fakes become more convincing.
GANs produced the first photorealistic generated faces, though they can be unstable to train.
Simplified GAN training step
fake = G(noise())
d_loss = bce(D(real), 1) + bce(D(fake.detach()), 0) # D: spot real vs fake
g_loss = bce(D(fake), 1) # G: fool DThey popularized AI-generated images, and are a historical root of today's deepfake debate. Diffusion models have since taken over much of this work.