Couldnt the same be done with the generation model? I.e, The detection model can emulate the generation model to inform detection.