The authors generated 100 images each from Dreamlike Photoreal 2.0, OpenJourney, and Stable Diffusion 1.5, plus 100 real COCO images, all from the same 100 captions. They computed CAS using each model on all images. CAS is significantly higher when the scoring model matches the generating model (diagonal: 189.90, 171.81, 163.24) than when it does not (off-diagonal: 32.39–121.80). This enables source detection at 0.90 accuracy and fake detection at 0.92 accuracy using a 3-layer MLP on CAS values.
Evaluated on 100 captions from COCO; the MLP classifier is trained on only 100 samples per model; the authors note this is a small-scale experiment demonstrating potential.