StyleGAN3

StyleGAN3 is NVIDIA's local AI image generation and training software for researchers and developers. Its architecture addresses generated detail that stays fixed to pixel positions instead of following the depicted objects as they move.

• Consistent detail during movement: StyleGAN3-T keeps generation consistent under translation, while StyleGAN3-R also supports rotation.
• Pretrained image models: Generate images with FFHQ, MetFaces and AFHQv2 models, or use compatible StyleGAN2 pretrained models.
• Custom training: Train networks on your own image collections, including labeled datasets, and evaluate saved model snapshots with FID.
• Interactive exploration: Inspect the characteristics of trained models with the included visualizer, or use pretrained networks in your own Python code.

Generation and training run on your hardware through PyTorch. Linux and Windows are supported, with Linux recommended for performance and compatibility; Docker support is also provided. The stated hardware requirement is high-end NVIDIA GPUs with at least 12 GB of GPU memory. Pretrained networks use .pkl files.

This listing covers the StyleGAN3 reference implementation and its StyleGAN3-T and StyleGAN3-R pretrained networks. The NVIDIA Source Code License for StyleGAN3 restricts use to noncommercial research and evaluation. It is not a permissively licensed commercial image generator.