Generative Adversarial Network
A generative adversarial network (GAN) uses two competing neural nets—one making fake data, one spotting it—to train a model that creates hyper-realistic images from pure noise.
A generative adversarial network (GAN) is a deep learning framework comprising two neural networks—a generator and a discriminator—that compete in an adversarial game. The generator creates synthetic data from random noise, while the discriminator evaluates whether the data is real or fake, pushing the generator to produce increasingly realistic outputs. Unlike diffusion models that optimize likelihood-based objectives, GANs rely exclusively on adversarial training to learn complex data distributions.
What this means in real life
A forger practices creating fake paintings while an art expert gets better at spotting them. Over time, the forger's fakes become nearly indistinguishable from real works, and the expert's eye sharpens—both improve through competition.
What it isn’t
It is not a single neural network that generates data. A GAN requires two separate networks working against each other; a standalone generative model (like a VAE) lacks this adversarial dynamic and trains differently.
Commonly misused online
People often use 'GAN' to mean any AI-generated image or deepfake, when technically a GAN is only one architecture among many that can create synthetic media. Not all AI fakes are GAN-produced.