“Analyzing and Improving the Image Quality of StyleGAN,” posted to arXiv on December 3, 2019 by Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila of NVIDIA, refined the celebrated StyleGAN architecture into the version known as StyleGAN2. The original StyleGAN produced strikingly realistic faces but suffered from characteristic blob-shaped artifacts and other telltale flaws that the authors diagnosed and removed.
The improvements were both architectural and methodological. They redesigned how the generator applies its style information and normalization to eliminate the droplet artifacts, reconsidered the progressive-growing strategy inherited from earlier work, and added a path-length regularizer that encourages smooth, predictable changes in the output as the input varies. These changes raised image quality, made the generator easier to invert so that a real photo could be mapped back into the model’s latent space, and improved the ability to attribute a generated image to the model that made it.
StyleGAN2 became the reference point for high-fidelity face generation and the engine behind countless deepfake demonstrations, synthetic-portrait tools, and research on detecting AI-generated images. For a general reader, it underscores how much of the realism in modern generated imagery came from careful, iterative engineering rather than a single breakthrough, and it marks roughly the high-water mark of GANs just before diffusion models took over general image generation.