Towards Universal Fake Image Detectors that Generalize Across Generative Models