// the find
soumith/ganhacks
starter from "How to Train a GAN?" at NIPS2016
A no-longer-maintained list of empirical tricks for training GANs, compiled by Soumith Chintala and others around 2016. It's for someone doing GAN work in an older framework or studying GAN training dynamics historically, not for anyone starting a new generative model project today.
The tricks are specific and testable (spherical Z sampling, separate real/fake batches for BatchNorm, label smoothing ranges), not vague advice. Several items cite the actual papers they came from (Salimans et al., Radford et al.), so you can go verify the claim instead of taking it on faith. Item 10 on tracking failure modes (D loss collapsing to 0, gradient norms over 100) is still useful diagnostic intuition for any adversarial training setup.
The README says outright it's unmaintained and the author isn't sure it's relevant past 2020 — that disclaimer is now six years stale itself. There's no code, just prose advice, so you can't run anything or check it against a reference implementation. GANs have largely been superseded by diffusion models for most generative tasks the tricks target (image synthesis), so half the value here is historical context rather than something you'd apply to a current project.