The paper investigates the possibility of creating a "universal" detector to distinguish real images from those generated by Convolutional Neural Networks (CNNs), regardless of the specific architecture or dataset used. The authors collect a dataset of 11 different CNN-based image generator models, including ProGAN, StyleGAN, BigGAN, CycleGAN, StarGAN, GauGAN, DeepFakes, cascaded refinement networks, implicit maximum likelihood estimation, second-order attention super-resolution, and Seeing-in-the-dark. They train a standard image classifier on ProGAN-generated images and evaluate its performance on other models. The results show that the classifier can generalize surprisingly well to unseen architectures, datasets, and training methods, suggesting that CNN-generated images may share common systematic flaws. The study also highlights the importance of data augmentation, diversity in training images, and robustness to post-processing operations such as JPEG compression, blurring, and resizing. The findings have implications for image forensics and the detection of synthetic images in various applications.The paper investigates the possibility of creating a "universal" detector to distinguish real images from those generated by Convolutional Neural Networks (CNNs), regardless of the specific architecture or dataset used. The authors collect a dataset of 11 different CNN-based image generator models, including ProGAN, StyleGAN, BigGAN, CycleGAN, StarGAN, GauGAN, DeepFakes, cascaded refinement networks, implicit maximum likelihood estimation, second-order attention super-resolution, and Seeing-in-the-dark. They train a standard image classifier on ProGAN-generated images and evaluate its performance on other models. The results show that the classifier can generalize surprisingly well to unseen architectures, datasets, and training methods, suggesting that CNN-generated images may share common systematic flaws. The study also highlights the importance of data augmentation, diversity in training images, and robustness to post-processing operations such as JPEG compression, blurring, and resizing. The findings have implications for image forensics and the detection of synthetic images in various applications.