What it does

This repo implements “Improved Training of Wasserstein GANs” in PyTorch, swapping the original TensorFlow code for PyTorch’s autograd. It covers toy 2D datasets, MNIST, CIFAR-10, and a character-level language model — all using the gradient-penalty trick to keep discriminator training stable without weight clipping.

The interesting bit

The language model is a curiosity: it uses 1D convolutions for both generator and discriminator on character data, producing output that reads like “The witl is f ont UAy Y nalence.” The author admits they “haven’t run enough epochs” because it’s “very time-comsuming” — a refreshing honesty rarely seen in READMEs.

Key highlights

• Four working scripts: toy datasets, MNIST, CIFAR-10, and character-level language
• CIFAR-10 support came via community PR, not the original author
• Directly acknowledges parent implementations by Ishaan Gulrajani and Martin Arjovsky
• Results folder contains actual sample images, not just loss curves
• Requires “a latest master version of Pytorch” — very 2017 energy

Caveats

• 64×64 architecture is unimplemented and marked “Looking forward to your pull request”
• MNIST generator and discriminator both use nn.Conv1d, which is unconventional for image data
• Language model results are extremely preliminary (8,699 epochs, still gibberish)

Verdict

Worth a look if you’re studying GAN history or need a minimal PyTorch reference for WGAN-GP. Skip it if you want production code or modern diffusion models — this is a 2017 paper implementation that stayed in 2017.