gigagan-pytorch

Why we included this project

Adobe's GigaGAN set a new state of the art for GAN-based image synthesis, and this repo is a readable PyTorch implementation that includes working training loops. Researchers and ML engineers who want to train a modern GAN without jumping to diffusion will find an unconditional generator, the 1k to 4k upsamplers, and optional text conditioning via CLIP embeddings. The code adds skip-layer excitation for faster convergence and a reconstruction auxiliary loss for stability, and it supports multi-GPU training with Hugging Face Accelerate. Install it with pip, point it at a folder of images, and you can train a 256x256 generator. It's a research toolkit, though: expect to do your own training runs rather than download a pretrained model.