With the rise of generative AI and the infiltration of AI-generated images into our lives (particularly through social media), we’ve all developed a new skill: separating the “real” from the “fake”. Sometimes the “fake” are easy to identify by inspection, but these generative models continue to improve at a staggering pace. Moreover, the sheer volume of images is prohibitive for manual identification.

My graduate teammates and I developed several competing CNN architectures and trained them to complete this task. We compare the effectiveness of this solution to the state-of-the-art and discuss the models’ generalizeability to new test sets.