Abstract
Generative models can produce impressively realistic images. This paperdemonstrates that generated images have geometric features different from thoseof real images. We build a set of collections of generated images, prequalifiedto fool simple, signal-based classifiers into believing they are real. We thenshow that prequalified generated images can be identified reliably byclassifiers that only look at geometric properties. We use three suchclassifiers. All three classifiers are denied access to image pixels, and lookonly at derived geometric features. The first classifier looks at theperspective field of the image, the second looks at lines detected in theimage, and the third looks at relations between detected objects and shadows.Our procedure detects generated images more reliably than SOTA local signalbased detectors, for images from a number of distinct generators. Saliency mapssuggest that the classifiers can identify geometric problems reliably. Weconclude that current generators cannot reliably reproduce geometric propertiesof real images.