A Generative Adversarial Network (GAN) is a machine learning architecture consisting of two neural networks, a generator and a discriminator, that are trained in opposition to one another. The generator creates synthetic data attempting to mimic real data, while the discriminator tries to distinguish between real and generated samples. Through this adversarial process, both networks improve until the generator produces outputs convincing enough to fool the discriminator.

GANs were introduced in 2014 and became the dominant approach for AI image generation throughout the late 2010s, producing significant advances in photorealistic synthesis, style transfer, and image-to-image translation. Notable GAN architectures include StyleGAN for high-quality face generation, CycleGAN for unpaired image translation, and Pix2Pix for paired image transformation. However, GANs have historically been difficult to train, prone to mode collapse where they generate limited variety, and require careful balancing to prevent training instability.

While diffusion models have largely displaced GANs as the preferred architecture for modern image generation due to their superior stability and quality, GANs remain important in AI history and continue to be used in specialized applications. Understanding GANs provides context for how generative AI evolved and helps explain why current diffusion-based models represent a significant architectural shift in the field.