In my melody generation system I'm already including melodies that I've judged as "good" (https://www.youtube.com/playlist?list=PLoCzMRqh5SkFwkumE578Y...) in the updated training set. Since the number of catchy melodies that have been created by humans is much, much lower than the number of pretty images, it makes a significant difference. But I'd expect that including AI-generated images without human quality judgement scores in the training set won't be any better than other augmentation techniques.