Compare HeartMuLa and ACE-Step 1.5, both running locally on a ZeroGPU-allocated GPU (no Inference API involved).