StyleGAN2

NVIDIA's benchmark generative adversarial network for synthesizing high-resolution, photorealistic imagery through decoupled style control.

StyleGAN2 improves upon its predecessor by redesigning generator normalization and adopting progressive growing alternatives to eliminate droplet artifacts. Developed by Tero Karras and the NVIDIA Research team, it utilizes weight demodulation and path length regularization to achieve superior image quality and smoother latent space interpolation. The architecture excels at generating 1024x1024 faces (FFHQ dataset) and cars (LSUN dataset) with precise control over stochastic variation. It remains a foundational tool for researchers using PyTorch and TensorFlow to push the boundaries of unconditional image synthesis.

https://github.com/NVlabs/stylegan2