Browsing through their code, it looks like they are using a VQGAN for image generation and ESRGAN for super resolution.