This repository offers an unofficial PyTorch implementation of the "Progressive Growing of GANs" paper, designed for researchers and developers working with generative adversarial networks. It provides a robust framework for generating high-quality images with improved training stability and variation, supporting resolutions up to 1024x1024.

How It Works

The project leverages the progressive growing technique, which incrementally adds layers to both the generator and discriminator during training. This method starts with low-resolution image generation and progressively increases resolution, enhancing training stability, quality, and the diversity of generated outputs.

Quick Start & Requirements

• Requirements: Ubuntu 20.04.3+, Python 3.8.3, NVIDIA GPU (GeForce 1080 Ti or higher with 8GB VRAM), NVIDIA drivers >= 470.86. CUDA 11.3 is recommended, though PyTorch includes CUDA support.
• Installation: Create a virtual environment and install via pip: pip install pro-gan-pth.
• Usage: The package provides command-line tools (progan_train, progan_lsid, progan_fid) and can be imported as a Python package (import pro_gan_pytorch as pg).
• Development: Clone the repository, install in editable mode (pip install -e .), and install development requirements (pip install -r requirements-dev.txt).
• Resources: The repository contains large binary assets, which may impact cloning times.

Highlighted Details

• Command-line utilities: Includes tools for training GANs, generating latent-space walk demo videos, and computing FID scores.
• Training flexibility: Supports configurable epochs, batch sizes, learning rates, loss functions (WGAN-GP, standard GAN), equalized learning rate, and exponential moving averages (EMA).
• Pre-trained models: Offers links to trained models, including a Metfaces model achieving an FID score of 101.624 at 1024x1024 resolution.
• Educational content: Features a Medium blog post explaining the progressive growing technique and a YouTube video showcasing training timelapses.

Maintenance & Community

The project is described as "fairly tiny" with no formal CI, automated testing, or documentation building. Contributions are welcomed via PRs and issues. @owang is credited for a Metfaces trained model. No community channels (e.g., Discord, Slack) are listed.

Licensing & Compatibility

The license type is not explicitly stated in the provided README, which may pose a challenge for commercial or closed-source integration without further clarification.

Limitations & Caveats

Training resumption is noted as not currently supported. The latent-space walk demo tool (progan_lsid) is restricted to .mp4 output format. The project lacks automated testing and formal documentation, relying heavily on the README for guidance.