Explore the official PyTorch implementation of the groundbreaking NeurIPS 2021 paper titled "Alias-Free Generative Adversarial Networks (StyleGAN3)." This implementation, authored by Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila, introduces a novel approach to generative adversarial networks that eliminates aliasing issues, offering improved synthesis for images and paving the way for more suitable generative models for video and animation.
The architecture ensures alias-free generation through small, generally applicable architectural changes, resulting in networks that match the FID of StyleGAN2 but differ significantly in internal representations. The networks are fully equivariant to translation and rotation, even at subpixel scales. The implementation includes tools for interactive visualization, spectral analysis, and video generation, making it a comprehensive platform for creative AI applications.
Alias-free generator architecture and training configurations (stylegan3-t, stylegan3-r).
Tools for interactive visualization, spectral analysis, and video generation.
Equivariance metrics for translation and rotation.
General improvements: reduced memory usage, faster training, and bug fixes.
Compatibility with old StyleGAN2 models and training configurations.
Synthetic image detection collaboration for AI forensics research.
Pre-trained models for various configurations available for download.
Supported on Linux and Windows (Linux recommended for performance).
Requires 1–8 high-end NVIDIA GPUs (Tesla V100 and A100 recommended).
64-bit Python 3.8 and PyTorch 1.9.0 or later.
CUDA toolkit 11.1 or later.
GCC 7 or later (Linux) or Visual Studio (Windows) compilers.
Utilize pre-trained networks for image and video generation.
Interactive visualization tool for model exploration.
Easily integrate pre-trained networks into Python code.
Train new networks using the provided train.py script.
Fine-tune pre-trained networks for specific datasets.
Monitor training progress with quality metrics such as FID and KID.
FID50k_full, KID50k_full, PR50k3_full, PPL2_wend, EQT50k_int, EQT50k_frac, and EQR50k recommended.
Legacy metrics such as FID50k, KID50k, PR50k3, and IS50k also supported.
Visualize spectral properties using built-in FFT mode in visualizer.py.
Calculate and display average 2D power spectra for training data and pre-trained generators.
Copyright © 2021, NVIDIA Corporation & affiliates. All rights reserved.
Licensed under the Nvidia Source Code License.
Citation: @inproceedings{Karras2021, ...}
