gigagan-pytorch
This implementation of a State-of-the-Art GAN from Adobe is enhanced for faster convergence and improved stability, leveraging lightweight GAN technologies. It features 1k to 4k upsamplers, skip layer excitation, and auxiliary reconstruction loss in the discriminator for high-resolution image synthesis. The project supports unconditional settings and integrates multi-GPU training via Huggingface's Accelerator, ensuring effective multi-scale input processing and stable training with an efficient gradient penalty application.
