What is a generative adversarial network used for?#
This system generates high-quality, original data from scratch, providing a cost-effective way to create training materials for other AI, design new products, or produce unique marketing assets without needing vast real-world datasets.
How does a generative adversarial network work?#
A GAN consists of two competing AI models. The first, called the "Generator," creates new data, like an image of a person's face. The second, the "Discriminator," acts as a judge, trying to determine if the image is a real photo from a training set or a fake one created by the Generator.
The two models are trained together in a continuous feedback loop. The Generator constantly tries to create more realistic images to fool the Discriminator, while the Discriminator gets better at spotting fakes. This adversarial process forces the Generator to become extremely skilled at producing highly realistic outputs that are often indistinguishable from real-world examples.
How is a GAN different from other generative AI?#
Competition between two networks is what sets a GAN apart from the models behind most of today's generative AI, and it also marks the limits of what GANs are used for.
Text is the clearest example. Writing is generated by transformer models such as large language models, which predict one word at a time, not by GANs. Training a GAN depends on adjusting its output by very small amounts to see whether the judge is more easily fooled, and there is no halfway point between one word and another — so the method that works well for the continuous values in an image does not transfer to language.
Images have largely moved on too: diffusion models, which start from random noise and repeatedly remove it, now produce most state-of-the-art image generation. GANs remain in use where generation has to be fast and the model small, including synthetic training data, image enhancement and audio.