Generative Adversarial Network (GAN) - Page 2

A generative adversarial network, or GAN, contains two neural networks trained through competition. The generator produces synthetic examples, while the discriminator tries to distinguish generated samples from real training data. As each network improves against the other, the generator can learn to create convincing images, audio, or other structured content. GANs enabled major advances in photorealistic synthesis and data augmentation, but training can be unstable and may collapse to a limited variety of outputs. Results also inherit biases and privacy risks from the source data. Diffusion models have replaced GANs in some applications, though GANs remain useful when fast generation or specialized architectures matter.