A robust model can effectively detect AI-generated faces from a variety of GAN and diffusion-based synthesis engines, even at low resolutions and compressed image qualities.