Introducing AuraSR - An open reproduction of the GigaGAN Upscaler
Today we are releasing AuraSR, a 600M parameter upsampler model derived from the GigaGAN paper. This model can upscale low-res images to 4x the resolution, and can be applied repeatedly. We are publishing this model under a truly open source license.
AuraSR excels in upscaling images generated by text-to-image models. This model does not have any limitations on resolution or upscaling factor.
Why did we train a GAN upscaler?
GANs generate images in a single forward pass through the generator network. Diffusion models use an iterative process of gradually denoising an image, which requires multiple steps. Generating and upscaling images with GANs can be orders of magnitude faster than diffusion and autoregressive models. For example AuraSR can generate 1024px images (4x upscale) in 0.25 seconds.
Special thanks to lucidrains for the https://github.com/lucidrains/gigagan-pytorch implementation.
If you are interested in training models for open source, we are hiring at fal.