StyleGAN

StyleGAN is a family of generative adversarial networks developed by NVIDIA researchers that produces realistic images with controllable stylistic variation. The core idea is a style-based generator in which a latent code maps to an intermediate space and is injected into the synthesis network at multiple layers. This allows influence over coarse attributes at early layers and details at later layers, enabling style mixing and adjustable fidelity.

The original StyleGAN, introduced in 2019, used a mapping network and adaptive instance normalization to blend styles across layers, with progressive growing to stabilize training on high-resolution images. StyleGAN2, released in 2020, reduced artifacts by replacing AdaIN with a modulated, demodulated weight mechanism and adding path length regularization, yielding sharper and more consistent samples.

StyleGAN3, announced in 2021–2022, improved robustness to geometric transformations and aliasing, producing more stable images under translation, rotation, and sampling changes. The family has been trained on diverse datasets such as FFHQ and applied to portraits, art, and synthetic data augmentation. Its latent spaces and manipulation techniques have influenced research on disentangled representations and controllable synthesis.

While widely used for creative and research purposes, StyleGAN also raises concerns about misuse and privacy given realistic synthetic faces. Researchers emphasize ethical considerations, data provenance, and governance alongside technical advances.