Ten GAN papers, one PyTorch training loop

What it does

This repository is a PyTorch re-implementation of a well-known TensorFlow collection. It houses ten GAN variants—from vanilla GAN to WGAN-GP and BEGAN—trained with the same generator and discriminator blueprints wherever possible. The author uses MNIST and Fashion-MNIST as the primary testbeds, producing epoch-by-epoch image grids and loss curves for each model.

The interesting bit

The value is in the controlled experiment. By keeping network capacity and architecture nearly identical across variants, the project isolates the algorithmic tweaks each paper introduces. It is essentially a laboratory for loss-function archaeology, complete with animated GIFs that show how quickly each variant collapses or converges.

Key highlights

• Covers ten named GAN variants derived from their original papers.
• Holds network architecture constant across models to isolate core ideas; only EBGAN and BEGAN receive necessary structural changes for their auto-encoder discriminators.
• Produces fixed-generation grids, conditional-generation samples, loss plots, and animated GIFs for MNIST and Fashion-MNIST.
• Includes conditional and continuous-code manipulation demos for CGAN, ACGAN, and infoGAN.
• Code for CPU mode exists, though the author only tested GPU mode.

Caveats

• Only MNIST and Fashion-MNIST were actually tested; CIFAR-10, SVHN, STL-10, and LSUN-bed are listed but unverified.
• Some models differ slightly from the original TensorFlow implementations.

Verdict

Useful for students and researchers who want to compare GAN mechanics without rebuilding ten codebases. Not the right choice if you need state-of-the-art image synthesis or validated results on complex datasets.