Generative adversarial network
Two networks compete, one generating fakes and one spotting them, until the fakes look real.
What it is
A generative adversarial network (GAN) was the leading image generator before diffusion. GANs are still used where speed matters, such as real-time face filters, super-resolution and image-to-image translation, because they produce an image in one pass.
How it works
A generator turns random noise into an image while a discriminator tries to tell generated images from real ones. Each improves against the other. Once trained, only the generator is kept, and it creates an image in a single fast step.
What it is good for
- Real-time image effects and filters
- Super-resolution and restoration
- Translating one image style to another
- Synthetic training data
Limits to know
- Training is unstable and can collapse to a few outputs
- Harder to steer with text than diffusion
- Less diverse than modern diffusion models
