“A Style-Based Generator Architecture for Generative Adversarial Networks,” posted to arXiv in December 2018 by Tero Karras, Samuli Laine, and Timo Aila of NVIDIA, pushed image synthesis to a new level of realism. Its generator could produce high-resolution photographs of human faces so convincing that, at a glance, they were indistinguishable from real ones - even though the people did not exist.
The innovation was in how the generator is structured. Rather than feeding a random noise vector straight into the network, StyleGAN passes it through a mapping network into an intermediate “style” space, then injects those styles at every layer to control features at different scales - coarse attributes like pose and face shape at low resolution, finer ones like hair and freckles at high resolution. This gave unprecedented, disentangled control over the generated image and let the model mix styles from different sources. The authors also released FFHQ, a high-quality dataset of 70,000 face images, to train on.
StyleGAN became famous beyond research through thispersondoesnotexist.com, a website launched in 2019 that showed a fresh, entirely synthetic human face on every refresh and brought home, to a wide public, how good generative models had become. It sharpened debates about deepfakes, synthetic identities, and trust in images. Successive versions improved quality further, and although diffusion models later overtook GANs for general image generation, StyleGAN remained a benchmark for high-fidelity face synthesis and a cultural touchstone for AI-generated imagery.