arkennemasis fal · MiniMax Music v1.5 · $0.03/run

Three cents per generation - thirty-three songs for a dollar. Before you get excited, check which box you're typing in, because on this node prompt is not the prompt. It's the lyrics. The style description goes in lyrics_prompt, which is backwards from the name and the most likely reason your first run sounds nothing like what you asked for.

What it is

The fal wrapper for minimax-music/v1.5, one of a family of MiniMax music endpoints the pack ships as separate nodes. Unlike the plain MiniMax Music node next to it, this one needs no reference track - you supply words and let it make the song. That makes it the entry point for the "I wrote a verse and I want to hear it" job, and at 3 cents the trial-and-error is basically free.

Worth knowing where it stands against local tools: ACE-Step running on your own card is excellent at instrumental music and weak on vocals, which is exactly the gap a hosted song model fills. The community barely discusses MiniMax's music models at all - no verdict exists to borrow - so treat the first few runs as your own listening test.

How it works

Read FAL_KEY from the .env beside your launcher, validate the inputs, estimate the cost, submit one billed request to fal, poll, download into output/fal/fal-ai_minimax-music_v1.5/. No media upload on this one - it's text in, audio out - so nothing before the submit could possibly bill you. The node's price badge and its pre-run estimate both track the per-run rate.

Inputs - mind the naming

• prompt - the lyrics. 10 to 600 characters, and the tooltip names the section tags it understands: [intro], [verse], [chorus], [bridge], [outro]. Use them; they're how you get structure instead of a continuous wall of singing.
• lyrics_prompt - the music description. 10 to 3000 characters, and the tooltip's whole explanation is "Control music generation." Genre, instrumentation, tempo, mood, production feel. Despite the name, this is where "lo-fi hip-hop, dusty drums, warm Rhodes" belongs.
• audio_setting_sample_rate - 8000, 16000, 22050, 24000, 32000, 44100. Leave it at 44100 unless you have a reason.
• audio_setting_bitrate - 32000, 64000, 128000, 256000.
• audio_setting_format - mp3, pcm, flac. Pick flac if you're going to master it; pcm is a raw dump you'll have to wrap yourself.

Optional, same four as everywhere in the pack: max_cost_usd (default $20; stops before sending when the estimate is over, 0 = off), max_concurrent (default 1 - one paid fal call at a time in a run), reuse_identical_run (default on - identical inputs return the already-paid file, restart included), extra_json (JSON merged into fal's request last).

Outputs: audio (waveform + sample rate, with a player on the node), audio_path, and info - request id, estimate, saved files, fal's full answer.

Install

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

Manager users can install via Git URL instead. Restart, then:

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

in the .env next to run_nvidia_gpu.bat. Menu: arkennemasis/fal/Audio/MiniMax, titled with its live price.

Getting a usable take

Write the description first and the lyrics second. Put the genre and the arrangement in lyrics_prompt, then keep the lyrics box tight - 600 characters is roughly two verses and a chorus, so don't paste a whole poem and hope. If the vocal comes back mush, it's usually too many syllables per line rather than a broken model; shorten the lines.

Then the repeats: reuse_identical_run means an unchanged run returns your previous file for free, which is great until you forget you changed nothing and conclude the node is broken. Flip it off for a genuinely new take, and note that a penny-price model is exactly the place where rolling the same lyrics five times is a reasonable strategy - five songs is fifteen cents.