PyTorch-GAN
PyTorch implementations of Generative Adversarial Networks.
A collection of 30+ Python implementations of GAN research paper variants, from vanilla GAN to CycleGAN, StarGAN, and Super-Resolution GAN, each runnable with a single command, for studying or experimenting with generative AI models.
PyTorch-GAN is a collection of Python implementations of Generative Adversarial Networks (GANs) from research papers. A GAN is a type of machine-learning model in which two neural networks are trained together: one generates new samples (usually images) and the other tries to tell the generated samples apart from real ones, with both improving as they compete. Different GAN papers tweak the architecture or training objective to produce different effects, and this repository gathers many of those variants in one place.
The table of contents in the README lists more than thirty implementations side by side, including the original GAN, Deep Convolutional GAN, Conditional GAN, InfoGAN, CycleGAN, Pix2Pix, BicycleGAN, BEGAN, DiscoGAN, DualGAN, Energy-Based GAN, Least Squares GAN, MUNIT, UNIT, StarGAN, Super-Resolution GAN, Wasserstein GAN with its gradient-penalty and divergence variants, and several others. Each entry includes the paper abstract, a link to the paper, and a runnable Python script. The author notes that model architectures do not always mirror the papers exactly, since the goal is to cover the core ideas rather than reproduce every layer configuration.
Someone would use this as a reference library for studying or experimenting with GAN variants without having to reimplement each paper from scratch. Installation is by cloning the repo and pip installing requirements; each variant is run with a single python3 command from its implementation folder. The tech stack stated is Python and PyTorch. The author also notes the repository has gone stale and is no longer being actively maintained. The full README is longer than what was provided.
Where it fits
- Run a named GAN architecture from a research paper with a single command to see it generate images before implementing it yourself.
- Compare multiple GAN variants such as Wasserstein GAN versus LSGAN side by side using the repo's reference implementations.
- Use a working CycleGAN or Pix2Pix implementation as a starting point for an image style transfer or domain adaptation project.