One MNIST dataset, ten GAN egos, fair-ish fight

What it does

This repository houses TensorFlow implementations of ten GAN variants—GAN, LSGAN, WGAN, WGAN-GP, DRAGAN, CGAN, InfoGAN, ACGAN, EBGAN, and BEGAN—plus VAEs. Every model is trained on MNIST and Fashion-MNIST using the same generator and discriminator blueprints (with minor tweaks for auto-encoder-based discriminators), so the only moving parts are the loss functions. The README surfaces sample grids at fixed epochs, letting you eyeball convergence, mode collapse, and conditional generation without switching repos.

The interesting bit

The project treats the sprawling GAN literature as a controlled experiment rather than a shopping list. By locking network capacity and swapping objectives, it isolates which theoretical tweaks actually move the pixel-quality needle. The author also documents honest failures—ACGAN collapses on Fashion-MNIST, InfoGAN ignores its noise vector—turning the repo into a useful cautionary atlas as much as a code collection.

Key highlights

• Side-by-side implementations of ten GAN variants and VAEs in a single TensorFlow codebase.
• Architecture-matched training on MNIST and Fashion-MNIST to isolate the impact of each loss function.
• Published sample grids for random, conditional, and continuous-code generation at set epochs.
• Openly flags failure modes like mode collapse and noise-vector saturation when moving from MNIST to Fashion-MNIST without extra tuning.
• Links to a PyTorch port and Google’s compare_gan study for readers who want FID/IS metrics and larger-scale analysis.

Caveats

• Dataset scope is narrow. Every visible result is limited to 28×28 grayscale digits and fashion items; performance on richer datasets is not shown.
• VAE coverage is thin. The title promises VAEs, but the provided README content is almost entirely devoted to GAN variants.
• Reference, not framework. This is a comparative benchmark collection rather than a reusable library or production tool.

Verdict

A solid pit stop for researchers and students who need a quick, architecture-fair reference for how each GAN flavor behaves on simple image data. Look elsewhere if you need a polished generative framework or turnkey support for high-resolution imagery.