I know nothing about generative fill, but several of the examples look like cropped photos. Could they be cropped photos with the original in the dataset, making them far less jaw dropping?
With the Beatles it generates a psychedelic scene and an actual yellow submarine from what is really a very mundane street scene, indicating the model has 'knowledge' of who the Beatles are and what their aesthetic is.
With Master of Puppets, it utterly fails to determine the tone or even content of the scene.