As requested:
Instead of relying on random text-to-music generation, I fed an existing track into the model and generated a new song inspired by it.