StyleGAN

StyleGAN is an NVIDIA-developed generative model that uses a style-based architecture to produce photorealistic, high-resolution images (up to 1024x1024) with precise, disentangled control over visual features.

StyleGAN, or Style Generative Adversarial Network, revolutionized image synthesis by introducing a style-based generator architecture that separates latent space from image features. The model employs an 8-layer mapping network to transform the initial latent vector (z) into an intermediate style vector (w). This vector then controls the image synthesis process at multiple resolutions via Adaptive Instance Normalization (AdaIN). This design enables disentangled control: coarse-resolution styles affect major features like pose and general shape, while fine-resolution styles control micro-details like texture and color. Building on Progressive GANs, StyleGAN ensures stable training and high-quality output, consistently generating photorealistic results, notably on the Flickr-Faces-HQ (FFHQ) dataset.

https://nvlabs.github.io/stylegan3