GANs Illustrated

This animation illustrates the training process of a system of GAN neural networks. In this simplified but accurate representation, there are four main steps in each training cycle, or epoch. A GAN network may require thousands of epochs to reach sufficient quality. The principal steps are:

1. POPULATE: real images are placed into an image set to be evaluated
2. GENERATE: the generator neural network produces fake images that are mixed at random with the real images
3. DISCRIMINATE: the discriminator network attempts to determine whether each image is real or fake
4. VALIDATE: the discriminator network is shown which of its judgments were correct and which were incorrect

After an epoch, the generator and discriminator networks are both adjusted based on the outcome of the just-completed training cycle. In each epoch, the generator’s ability to produce high-quality fakes improves, as does the discriminator’s accuracy. Together, they train each other to refine their abilities, often achieving highly realistic and believable results after numerous cycles.

Isabella Pu, GANs Illustrated (excerpts), 2022, digital animation, Design, programming and animation by Isabella Pu, Princeton University; Advisor–Adam Finkelstein, Professor of Computer Science, Princeton University


︎ View GANs Illustrated