AI glossary · Basic · Generative media
GAN
A GAN (generative adversarial network) is a method that pits two neural networks against each other, a generator that creates fake data and a discriminator that tells real from fake, so that the generator produces increasingly realistic images and other data.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Models & generation
- Fields
- Generative media
- Added
- 2026-10-08
How it works
The generator takes random numbers (noise) as input and creates fake images. The discriminator looks at real images and the generator's fake images in turn and judges which is which. The discriminator's judgments become feedback: the generator learns to fool the discriminator better, and the discriminator learns to tell them apart better. Repeating this competition makes the generator's output steadily closer to the real thing.
Why it matters
GANs showed that a model can learn what data looks like and create new data without people labeling every answer. Technologies such as photos of faces of people who do not exist, style transfer and sharpening low-quality photos advanced greatly thanks to GANs. Before generative AI drew wide attention, it was one of the first technologies to leave a strong public impression that 'AI can make pictures'.
Good to know
If the balance between the two networks breaks down, training easily becomes unstable. A common problem is also that the generator keeps producing just a few outputs that easily fool the discriminator (mode collapse). For these reasons diffusion models have become mainstream in image generation today, but GAN ideas are still used where fast generation is needed. Because GANs can be misused for realistic fake videos (deepfakes), countermeasures such as AI watermarks are discussed alongside them.
Example
GAN-based techniques have been used in photo app features such as 'restore old photos sharply' and filters that turn selfies into cartoon styles or different ages. They are also used to create synthetic data, such as generating face images for training without real personal information.