🎵 MuAPI Audio

Audio is the pack's quiet corner, and 🎵 MuAPI Audio keeps it simple to the point of being blunt: pick a model, write a prompt, get a URL. Fourteen endpoints, one prompt box.

What's behind the dropdown is a decent spread of the hosted audio world. Seven Suno jobs - suno-create-music (the default), suno-remix-music, suno-extend-music, suno-add-instrumental, suno-add-vocals, suno-generate-mashup, suno-generate-sounds. Three from MMAudio - mmaudio-text-to-audio, mmaudio-video-to-audio, mmaudio-v2/text-to-audio. And two MiniMax TTS endpoints, minimax-speech-2.6-hd and minimax-speech-2.6-turbo. Roughly: song generation, foley, and speech.

The thing to know before you wire it up

The outputs are audio_url (STRING) and request_id. There is no AUDIO output. Nothing here produces a waveform tensor, so you cannot wire this into ComfyUI's audio nodes, a video-sync node, or a latent audio pipeline. You get a URL.

Practically that means: read it out of the console or preview it in a browser, and if you need the file, fetch that URL yourself. The pack has no audio downloader - 🎬 MuAPI Save Video explicitly validates for cdn.muapi.ai video and decodes with OpenCV, so don't expect it to save an mp3 cleanly. This is the least integrated node of the thirteen, and if you assumed it would behave like a local audio node, that assumption is the bug.

Inputs

Two required fields: model and prompt (multiline, defaulting to "An upbeat electronic track with driving bass"). That's it. Optional: api_key, custom_endpoint (used when model is custom), and extra_params_json.

That last one is not optional in practice. Notice the mismatch: half these endpoints exist to modify an existing track - remix it, extend it, add vocals to it - and the node has no audio input field at all. The source track has to go through extra_params_json as a URL keyed the way that endpoint expects. Same for mmaudio-video-to-audio, which by definition needs a video the node can't accept. Check muapi's docs for the field name; the node merges your JSON straight into the request body and doesn't validate it.

Also absent: duration, tempo, lyrics, instrumental/vocal flags, language. A Suno call with real lyrics in it means writing that JSON yourself.

Where this sits next to local

The KB's audio-generation.md is clear about the open alternative: ACE-Step is the local Suno, and as of the 1.5 release it runs in about 4GB of VRAM with LoRA training support. Its instrumental output is genuinely good; its lyrics are not - "The lyrics are absolute garbage" is the recording in the KB, and Suno has the same complaint. So the honest split is: for instrumentals and loops you'll iterate on, run ACE-Step locally and pay nothing. For a song with vocals that needs to be done today, the hosted models are ahead, and this node is the shortcut.

The MMAudio endpoints are the cloud twin of a job the KB already treats as local plumbing - MMAudio is the model people bolt onto the end of a Wan or Hunyuan render to add foley (HunyuanVideo-Foley being the Tencent alternative). If you already run video locally and only want sound, a local MMAudio node is the cheaper long-term answer. If you don't, mmaudio-video-to-audio is one JSON payload away.

And the two MiniMax speech endpoints are closed, hosted TTS - fine for a quick voice line, but the KB's TTS section has the local options (Kokoro, Piper, Chatterbox, Higgs) if cloning or bulk narration is the actual task.

Install

Manager → Install via Git URL → https://github.com/SamurAIGPT/muapi-comfyui → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/SamurAIGPT/muapi-comfyui
pip install -r muapi-comfyui/requirements.txt

requests, Pillow, numpy, torch, opencv-python - no model files, nothing in models/audio/.

Troubleshooting

If a run fails, the useful message is in the console: the pack prints [MuAPI Audio] <endpoint> before submitting and the job status on every poll. A 400 usually means the endpoint wanted a field you didn't provide - for the remix/extend/vocals variants that's almost always the source track in extra_params_json. And remember the poll loop gives up after 900 seconds; a long music generation can outlast it while still finishing server-side, so check your dashboard before re-running and paying twice.