arkennemasis fal · MiniMax Music 3 · $0.002/s

Local music generation has a real answer - ACE-Step, which does instrumental work genuinely well on a 4GB card and has a weak spot exactly where songs are hard: vocals. If what you need is a song with lyrics that scan and a vocal that sounds like a human decided to sing, the open options still aren't there, and MiniMax Music 3 is the hosted shortcut. Up to five minutes, $0.002 per second, which means a full minute of finished audio costs twelve cents.

That pricing is the pitch. Twelve cents per minute of song makes iteration cheap enough that you can generate four takes of a chorus and keep one, which is not true of any per-image or per-second video model in this pack.

How it works

The node is a straight wrapper: FAL_KEY from the install's .env, input validation, cost estimate held against max_cost_usd, one submit, poll while the node shows status, then the file lands in output/fal/minimax_music-3/ and is shown on the node. Cancel in ComfyUI cancels the job on fal as well.

Inputs that matter

prompt is the music description - style, mood, vocals, instrumentation, arrangement. For anything you care about, write it as a structured caption: genre and BPM and key up front, then vocal details, then the arrangement section by section. Genre-only prompts get you genre-typical mush.

lyrics is where most people lose an hour. Structure tags - [intro], [verse], [pre-chorus], [chorus], [post-chorus], [bridge], [instrumental], [solo], [outro] - must each sit on their own line. The tooltip is explicit that text on the same line as a leading tag gets dropped by the model's input contract. So this is wrong:

[verse] I woke up in the city

and this is right:

[verse]
I woke up in the city

duration is a float up to 300 seconds and it's an upper bound, not a promise - the model can stop earlier, and the real length comes back on the duration output. Set 60 and get 43 seconds; that's normal, not broken. Plan your edit around the returned value, not the requested one.

num_inference_steps (default 30) is flow-matching steps per 8-second denoising chunk - more steps, better quality, slower. guidance_scale (default 1.7) is the classifier-free guidance scale; nudge it up if the prompt is being ignored, down if the result sounds over-cooked and harsh. seed at -1 lets the model pick, and the seed that was actually used comes back on the seed output - grab it if you want the same take again.

The pack-wide four: max_cost_usd ($20 default cap, 0 = off), reuse_identical_run, extra_json, max_concurrent.

Outputs: audio (wire it into the pack's assembly nodes or a Save Audio), audio_path, seed, duration, info (request id, cost estimate, saved files, fal's answer). Keep the seed; you'll want it.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis
pip install httpx

Manager → search arkennemasis. Then the key:

FAL_KEY=your-key-from-fal.ai/dashboard/keys

in the .env next to run_nvidia_gpu.bat. Restart ComfyUI.

Where it bites

• Tags sharing a line with their lyrics. Silent, and it eats the first phrase of every section.
• Expecting duration to be respected exactly. It's a ceiling.
• Repeat runs handing back the old file: reuse_identical_run is on by default, and identical settings plus identical inputs means no new call. Change the seed for a new take.
• Everything queues behind max_concurrent: 1. If you're generating ten variations, raising it on these nodes is worth it - and if one fal node in the run fails, the ones still waiting don't start, so a failure can't run up a tab.