MiniMax Music 2.6

This node turns a text prompt into an actual song - vocals, arrangement, the works - and it does it without you downloading a single model file or worrying about VRAM. MiniMax Music 2.6 runs on Runware's cloud, so the whole exchange is: you type a prompt, the node fires off a request, and a finished audio file comes back as a native AUDIO tensor. If you've ever wanted to score a video or demo an idea but balked at the 10GB model download and the fiddly local setup that music gen usually demands, this is the cheat code.

It's part of the Runware pack, which is less a "custom node" and more a thin client for every model the Runware API hosts. The pack installs once (search Runware in ComfyUI Manager, or clone https://github.com/Runware/ComfyUI-Runware into custom_nodes/, run pip install -r requirements.txt, restart) and then you just need an API key from the Runware dashboard - Settings → Runware API key in the UI, the RUNWARE_API_KEY env var, or runware auth login. Nothing else. No checkpoints to place, no venv hell.

How it works

This is MiniMax's music-generation model (the minimax:[email protected] AIR) doing text-to-music. You describe the vibe in plain language - genre, tempo feel, mood, instruments - and the model composes from scratch. What makes it usable as a real tool rather than a toy is the lyrics path: drop structured lyrics into settings.lyrics using section tags like [Intro], [Verse], [Chorus], [Bridge], [Outro], and [Inst], and the model respects the structure. That's the difference between "vague synthwave" and "a song with a verse-chorus-verse form."

The inputs that matter

Nothing here is required, which is both the point and the trap - an empty prompt means the model decides everything, so give it something to work with.

• positivePrompt - the style and content of the song. This is your one real creative input.
• settings.lyrics - structured lyrics with the section tags above. settings.lyricsOptimizer will clean up or even generate lyrics from the prompt when it's on, which is handy when you want a song but don't want to write one.
• settings.instrumental - drop vocals entirely. Good for beds and underscores.
• numberResults (1–4) - batch out variations; each one uses a different seed.
• outputFormat - MP3 by default; WAV, FLAC, and OGG are there if you need lossless.
• audioSettings - a typed socket you can feed from the Runware Audio Settings builder to pin sample rate, channels, and bitrate instead of trusting the default.

The single output, audio (AUDIO), wires straight into a Core SaveAudio node or anything that consumes the native audio type. After each run the node's title bar shows the cost - for a short song you're usually talking fractions of a cent, which is the whole appeal of the pay-per-generation model.

Where people get burned

• No API key → the node dies with "No Runware API key. Set RUNWARE_API_KEY or run runware auth login." It's clean and unambiguous, but it still trips people up on first use.
• Forgotten lyricsOptimizer semantics - leave it off if you wrote careful lyrics; it's an optimizer, not a pass-through, and it will rewrite your words.
• The toggles. Many params in this pack are "enable to set X" booleans paired with a _value field - flip the toggle, then set the value. Leave it off and the model's default wins. If a setting looks like it's being ignored, check the toggle, not the field.

One honest caveat: as with all text-to-music right now, results are lottery-grade. You'll get usable takes, and you'll get stuff that's impressively wrong. But for zero setup, zero VRAM, and sub-cent generations, it's the easiest way to iterate on song ideas inside your existing ComfyUI graph.