HyperGAN is an open-source, composable framework for building and training generative adversarial networks (GANs) in PyTorch, aimed at developers, researchers and artists. The GitHub project has gathered about 1,200 stars and remains under active development, with updates continuing into October 2026.
Training GANs usually means stitching together generators, discriminators, losses and samplers from scattered code. HyperGAN treats these as composable building blocks: you assemble a GAN from interchangeable components through configuration or its API, then train, sample and experiment through a consistent interface and user interface. The framework ships with implementations of common GAN components and example configurations for tasks such as image generation and colorization, so you can reproduce known setups and then swap pieces to test new ideas. Because everything is modular, it is equally useful as a learning tool โ each part of the adversarial setup is visible and replaceable โ and as a research sandbox for prototyping architectures without rewriting boilerplate training loops.
HyperGAN has been in pre-release and open beta, with its component implementations refined over several years of experimentation. Its configuration files make experiments easy to share and reproduce, and example projects in the ecosystem show the framework applied to creative image tasks as well as research benchmarks.
HyperGAN is free and open source under the MIT license. There is no paid edition or hosted service; you run it on your own hardware, with training costs limited to your own compute.
Machine learning researchers, students and creative coders who want to experiment with GAN architectures in a structured framework instead of assembling every training loop from scratch.