What it does InfoGAN is a variant of GANs that learns disentangled, interpretable representations without any labels. The generator gets extra “latent codes” alongside the usual noise vector, and the discriminator tries not only to spot fakes but to recover those codes from generated images. If the discriminator can read the code back out, the generator is forced to use it consistently.

The interesting bit The trick is an extra mutual-information term in the loss. The authors approximate it with a variational lower bound, which is the kind of elegant hack that makes variational methods tolerable. The result: one continuous code might control rotation, another digit identity, all emergently.

Key highlights

• Reproduces the MNIST results from the 2016 paper (Xi Chen et al.)
• Uses a now-ancient TensorFlow dev build (commit 79174a, circa r0.9rc0)
• Ships with Docker instructions for the brave
• Only ~1,070 stars, which feels low for a paper with Sutskever and Schulman on it
• OpenAI has archived it: “code is provided as-is, no updates expected”

Caveats

• Dependencies are frozen in 2016 amber: prettytensor, progressbar, and a specific TF commit
• Only the MNIST launcher is explicitly documented; other datasets are presumably DIY
• No candidate images provided, so you’ll have to run TensorBoard yourself to see the pretty spirals

Verdict Worth a look if you’re studying disentanglement or the history of generative models. Skip it if you want something that runs on modern PyTorch without archaeology.