Multi input? Infilling? First generative audio model I've seen that starts to close the gap with image models.