“Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks,” posted to arXiv on November 19, 2015 by Alec Radford, Luke Metz, and Soumith Chintala, took the still-fragile GAN idea from 2014 and made it work reliably. The original adversarial framework was elegant in theory but notoriously hard to train; DCGAN proposed a specific set of architectural rules that produced stable training and convincing images.
Those rules became standard practice: replace pooling layers with strided convolutions, use batch normalization in both the generator and discriminator, drop fully connected hidden layers, and use particular activation functions in each network. The result was a class of convolutional networks that learned a hierarchy of image features from unlabeled data and could generate new images from a random input vector. The authors also showed that the learned features were useful for ordinary classification tasks and that arithmetic on the input vectors produced meaningful changes in the output, such as adding glasses to a generated face.
DCGAN is important because it turned generative adversarial networks from a striking proof of concept into a tool researchers and practitioners could actually use and build on. Nearly every later GAN, from progressive growing to StyleGAN to BigGAN, inherited architectural choices first laid out here. For a general reader, it marks the moment adversarial image generation became repeatable rather than a one-off curiosity.