Two networks in competition
A GAN (Generative Adversarial Network), published in 2014, trains two neural networks against each other: one generates data, the other tries to tell real from fake. It led image generation before diffusion models took over, and it is a landmark in the history of generative AI.
The forger and the inspector
The usual analogy is a forger and an inspector. The generator tries to produce data indistinguishable from the real thing; the discriminator tries to catch it. As each improves to stay ahead of the other, the generator ends up producing output the inspector cannot detect. That competition is the training process.
What it produced
- Photorealistic faces of people who do not exist
- Upscaling photographs and colorizing black-and-white images
- Style transfer, such as rendering a photograph in the manner of a painter
- The technical basis of deepfakes
The shock of the late 2010s — AI-generated faces indistinguishable from photographs — came largely from GANs.
Where it stands now
Diffusion models took over as the mainstream for image generation, being more stable to train and more varied in output. GANs remain in use where generation speed matters, such as local editing and upscaling. More lastingly, the idea of setting models in competition shaped a great deal of subsequent research. It is worth knowing as part of the history of the field.
The story that Ian Goodfellow implemented the idea in a single night after an argument with colleagues is one of the field’s best-known anecdotes.
Note that Generative appeared in the name well before generative AI became a common phrase — GANs were an early instance of the idea.