A PyTorch Rosetta Stone for GAN papers, now seeking a caretaker

What it does This repository houses over two dozen standalone PyTorch implementations of Generative Adversarial Network architectures drawn from research papers. Each model—whether a basic gan.py, a cyclegan.py for image translation, or a wgan_gp.py for stable training—lives in its own directory as a self-contained script. The goal is to provide readable code that captures the central idea of each paper without forcing you to reconstruct the original TensorFlow or Lua source.

The interesting bit The author explicitly trades architectural perfection for pedagogical clarity, noting that models “will not always mirror the ones proposed in the papers.” That makes this less a production framework and more a well-indexed set of lab notes: you can compare how began.py, infogan.py, and relativistic_gan.py handle loss functions side-by-side in a single codebase.

Key highlights

• Broad coverage, from foundational models like dcgan.py to specialized variants like stargan.py, munit.py, and esrgan.py.
• Self-contained scripts: each implementation is an isolated file you can read and run independently without navigating a complex package hierarchy.
• Includes visual results—GIFs and stills—for many models, so you can see expected outputs before committing GPU hours.
• Explicitly educational focus: prioritizes readable code that captures a paper’s core insight over pixel-perfect layer matching.
• Companion to a similar Keras-GAN repository if you prefer that framework.

Caveats

• The repository is explicitly unmaintained; the author posted a “gone stale” notice and is seeking a collaborator to take over.
• Architectures are simplified distillations, so don’t expect to reproduce exact paper benchmarks or use these as drop-in production models.
• No unified API or library structure—each script is its own island, which means shared utilities or consistent hyperparameter handling are absent.

Verdict Grab this if you are a student, researcher, or hobbyist who wants to see how Wasserstein loss or adversarial autoencoders look in clean PyTorch before building your own variant. Skip it if you need a maintained, pip-installable GAN framework with reproducible paper results.