Twenty GANs, five VAEs, and a Helmholtz Machine

What it does

This repo is a curated cabinet of reference implementations covering the generative-modeling canon from the mid-2010s. It collects standalone scripts for twenty GAN variants, five VAE flavors, two RBMs, and a Helmholtz Machine, written in PyTorch and/or TensorFlow. Think of it as a museum exhibit where each landmark paper gets its own minimal working script rather than a unified library.

The interesting bit

The collection doesn’t just chase the hype cycle; it preserves oddities like the Wake-Sleep Helmholtz Machine and Binary RBMs with Persistent Contrastive Divergence that most modern toolkits ignore. That breadth makes it a useful time capsule of the field’s mid-2010s generative landscape.

Key highlights

• 20 GAN architectures, from Vanilla and Wasserstein to DiscoGAN, BEGAN, and GibbsNet
• VAE variants including Conditional, Denoising, Adversarial Autoencoder, and Adversarial Variational Bayes
• Old-school energy models: Binary RBMs (CD and PCD) plus a Binary Helmholtz Machine
• Dual PyTorch/TensorFlow coverage (the README notes both frameworks, though it doesn’t map which model uses which)
• ~7.5k stars, suggesting it has served as a de facto textbook for many practitioners

Caveats

• The README lists both PyTorch and TensorFlow as dependencies but never clarifies which model uses which framework
• Each model lives in its own directory with its own output folder, so expect standalone scripts rather than a shared library or API

Verdict

Worth bookmarking if you’re a student or researcher who needs to dissect a specific classic paper’s implementation. Skip it if you want a maintained, batteries-included generative framework or state-of-the-art diffusion transformers.