So there's this model of StableDiffusion that is trained on spectogram images, then you send the prompt and the model generates a spectogram that can be turned into audio.