nvlabs/stylegan3

StyleGAN3 is NVIDIA's research code for generating photorealistic synthetic images using a type of AI model called a generative adversarial network. The project was published as a paper at NeurIPS 2021. The core improvement over the previous version, StyleGAN2, is fixing a problem called aliasing: in earlier models, fine details in generated faces or objects would appear to be stuck to fixed positions in the image frame rather than moving naturally with the depicted surfaces. StyleGAN3 changes the internal math to make generated content behave more like objects in the real world, which matters especially for animations and video where you want consistent motion. Pre-trained models are available from NVIDIA's model catalog for several datasets: human faces at 256x256 and 1024x1024 resolution, artistic portraits from the MetFaces dataset, and animal photos from the AFHQ dataset. Two model variants are offered, one that handles translation and one that handles both translation and rotation. Running a model to generate images is a single Python command pointing at a model file and specifying a seed number. The code also includes tools for interactive visualization, generating video interpolations between generated images, and analyzing the frequency characteristics of trained networks. The hardware requirements are significant: the README recommends one to eight high-end NVIDIA GPUs with at least 12 GB of memory each. CUDA 11.1 or later, Python 3.8, and PyTorch 1.9 or later are also required. A conda environment file and a Docker configuration are both provided to simplify setup. The code compiles custom GPU extensions automatically when run, which means you also need a C++ compiler installed. This code is primarily a research release. It is meant for studying or building on the techniques in the paper rather than as a turnkey tool for casual image generation. Business use requires a separate license from NVIDIA, and a link for licensing inquiries is included in the README.