Generative Adversarial Network

A generative adversarial network (GAN) is a deep learning framework introduced by Ian Goodfellow in 2014 that trains two neural networks simultaneously in a minimax game. The generator network learns to produce synthetic data (such as images) that resemble the training data, while the discriminator network learns to distinguish between real and generated samples. Through this adversarial process, both networks improve, and the generator eventually produces highly realistic outputs.

Training GANs is notoriously challenging due to issues like mode collapse (where the generator produces limited variety), training instability, and the difficulty of evaluating generation quality. Numerous GAN variants have been developed to address these challenges, including DCGAN, Wasserstein GAN, StyleGAN, and Progressive GAN. Techniques such as spectral normalization, progressive growing, and careful architectural choices have significantly improved GAN training stability.

GANs have produced remarkable results in image synthesis, producing photorealistic faces, artwork, and scenes. They have been applied to image super-resolution, style transfer, data augmentation, drug discovery, and video generation. While diffusion models have recently surpassed GANs in some image generation benchmarks, GANs remain influential and continue to be used in applications requiring fast inference and real-time generation.