StyleGAN

In December 2018, NVIDIA researchers released StyleGAN, a GAN whose faces were good enough to fool most people. In February 2019, software engineer Phillip Wang used it to launch This Person Does Not Exist, which showed a new invented face on every page reload and made the technology famous overnight.

The key idea is in the name: style.

1. A random vector first passes through a small “mapping network” into a better-organized latent space.
2. That latent then sets the “style” of each layer of the image generator, a CNN that grows the image from coarse to fine.
3. Early layers decide pose and face shape, middle layers features like hair and eyes, late layers color and skin texture. Mixing styles from two faces at different layers blends them in controllable ways.

Mode collapse

GANs are notoriously hard to train. A classic failure is mode collapse: the generator finds a few outputs that fool the discriminator and keeps producing just those, losing diversity. StyleGAN’s careful design and training tricks reduced such problems, but they remain a key reason diffusion models later took over most image generation.

StyleGAN’s success also showed the downside: realistic fake faces became a cheap tool for fake accounts, an early warning for today’s deepfake debates.