lightweight-gan

Why we included this project

For teams with small image collections and limited GPU time, this PyTorch implementation of the ICLR 2021 lightweight GAN offers a realistic path to high-resolution synthesis. Its core trick, skip-layer excitation in the generator paired with self-supervised learning in the discriminator, lets it converge on a single GPU in hours from a dataset of under a hundred images. The tool runs from a single command, pointing it at a folder and an output size, with checkpoints and samples saved automatically. Built-in augmentation, essential for low-data training, is configurable from the same CLI. It is a research codebase, so expect rough edges, but it is direct and usable for hands-on GAN work without a big compute budget.