Music Prompt Pixaroma

What it does

A music model needs two different pieces of writing. The caption describes how the song should SOUND: the genre, the tempo, the key, what the voice is like and which instruments play. The lyrics are the words that actually get sung.

This node writes both from one idea. It runs a language model you already have, on your own machine, twice on a single load: once for the caption and once for the lyrics. Nothing is sent anywhere.

Wire caption and lyrics straight into MiniMax Music 3 Text Encode, and duration into its max_duration. That last wire means you set the length ONCE, here, where the words are written for it.

Getting started

• Put a language model in your ComfyUI/models/text_encoders folder. This node only reads and writes words, so it does NOT need a vision model.
• Press the gear on the node and pick it. Both formulas were measured on qwen3.5_4b_int8_convrot.
• Type your idea in plain words, like a slow acoustic song about coming home in the rain.
• Set the length, then press Generate.

Length, and the duration wire

The music model treats length as a CEILING. It can end a song early, but anything past the limit is simply cut off part way through. So a lyric written for three minutes against a thirty second setting gets chopped.

Wire the duration output into the music node's max_duration and the two can never disagree. The highest it accepts is 360 seconds.

A song can still come out shorter than you asked. That is the music model stopping when the words run out, not a fault: it only sings what it was given. Measured over ten songs, nine used the whole length and one stopped at three quarters, because that seed happened to write half as many lines.

If a song comes out short, press Re-roll. How many lines get written varies from one seed to the next, and a fresh seed usually fills the time.

Verses are a request, not a promise

Left on Auto, the length alone decides the shape, and that is the most reliable way to run it. Under forty seconds you get a verse and a chorus. Around a minute adds a second verse. Around two minutes adds a bridge and a final chorus.

Up to about a minute and a half, Auto states that shape outright rather than leaving the model to work it out, because that is where it was measured to get it wrong in both directions: minute-long songs kept coming back with the single verse of a thirty second one, and thirty second songs kept opening with an instrumental section they had no room for. Saying it plainly took the first from three runs in five to five in five, and the second from one in four to four in four.

Longer than that, Auto says nothing, because there it already gets it right and naming a number made it worse.

Ask for a number instead and the model usually obeys, but not always. One and two come back exactly as asked. Three sometimes comes back as two. That is why the chips stop at three: asking for more does not get you more.

Asking for verses also OVERRIDES the length shape, so three verses at three minutes gives a shorter song than three minutes on Auto would.

Auto

The length decides everything. The most dependable setting.

1 to 3

Ask for that many verses, each with a chorus.

Bridge

Ask for a bridge: one different section, usually near the end.

Break

Ask for one instrumental break inside the song, where the band plays and nobody sings. It uses up singing time. Do not use it on a 30 second song: in testing the chorus was never sung at all. It was called Instr. before, and was renamed because it sat next to the Instrumental setting and the two mean different things.

Voice / Instrumental

Voice writes a song with singing. Instrumental writes a piece with no singing at all, which runs the model once instead of twice and is about two and a half times faster.

Using a different model

Everything this node writes comes from a FORMULA SET: the two instructions, plus the sampling that makes them work, under a name saying what they are for. The one that ships is MiniMax Music 3 (Qwen3.5 4B int8), which names the music model it writes for and the language model it was measured on.

On a different language model the wording may need changing. Press the gear and you will find it all there, nothing hidden.

A set also remembers which language model it was measured on. If you have that model it is chosen for you; if you do not, the line under the picker says so in amber and your own model is left alone, so a set from somebody else's machine can never point this node at something that is not there. A model on the clip wire always wins.

The set picker

Every set you have. An orange dot ships with Pixaroma, a grey one is yours. Picking one copies its wording, its numbers AND its model onto the node, so this is also how you go back after changing something.

Edit

Opens that instruction in a full screen box. It starts from whatever the node is using now, so you can change a line rather than write from nothing. Save it back unchanged and the node keeps following the built-in one.

Save as

Keeps the current wording and numbers under a name of your own. It suggests (mine) on the end when you started from the shipped set. After changing one of your own sets, its name is already filled in: press Enter to update it, or type a new name to keep both. Typing the name of another set of yours asks before replacing it. Your sets live in one file in your ComfyUI user folder, so a reinstall of the node does not touch them.

Delete

Removes one of yours. The set that ships with Pixaroma cannot be deleted or overwritten, so there is always a way back.

temp

How much the model improvises. The caption wants a low number so it states facts; the lyrics want a high one or every song rhymes the same way.

max len

How much it may write before stopping. Raise it a lot for a model that thinks out loud before answering, or it spends the whole budget thinking and writes nothing.

Writing your own instruction

The two instructions do different jobs and each has a few things it must keep saying, whatever else you change.

THE CAPTION describes the SOUND and never the words. It has to come back as three labelled parts, in this order: Global Metadata (genre, a tempo in BPM, a key, the mood, how the recording should sound), Vocal Details (whose voice, what it is like, any harmonies), and Arrangement (which instruments carry it, what the bass and drums do). It must never write lyrics, section tags or quoted words, because those go in the other field.

THE LYRICS are sung out loud, every single line of them. Anything in brackets gets sung too, so a stage direction like a slow piano begins will be sung by a person. Lay it out with a section tag on its own line before each part, from [Intro] [Verse] [Pre-Chorus] [Chorus] [Post-Chorus] [Bridge] [Instrumental] [Solo] [Outro]. A tag can stand alone with nothing under it, which means the band plays and nobody sings, and that still uses up time.

Both instructions should say to write nothing else: no explanation, no markdown, no repeating the idea back.

• The node adds your length and shape to the end of the idea before it asks, so your instruction does not need to mention either.
• Change one thing at a time and run the same idea on a couple of seeds. Wording that reads better is not always wording that works better.
• If a result looks wrong, try the temperature before rewriting anything. The shipped caption wording writes cleanly at 0.3 and rambles at 0.7 with not a word changed.
• Do not put examples inside a rule. Naming a few instruments tends to put those exact instruments into every song.
• Say what you DO want rather than what you do not. Telling this model not to leave a section empty was measured making things worse, once producing 26 sung lines for a thirty second song.

The buttons on the node

Caption / Lyrics

Which of the two you are looking at. Both are written every run, this only picks what the box shows.

Expand

Write the idea in a full screen box.

The bar between the boxes

Drag it to give the idea more or less room. Double click puts it back.

The seed

Click the number to type one. F keeps the same seed, so an unchanged node is cached and Run is instant. R rolls a new one every run, so every run is a different song.

Tags

Opens your tag library, the same one Prompt Pixaroma and AI Prompt use. Anything you save there can be used in all three.

Re-roll

A new seed, then run again. The quickest way to try another version of the same idea.

Copy

Copies whichever of the two you are looking at.

Free VRAM

Unloads the model when this node finishes. Turn it on when a music model has to fit in the same card afterwards.

Generate

Queues the whole workflow, the same as pressing Run.

Tags in your idea

Your idea box understands the same @tags as Prompt Pixaroma and AI Prompt, out of the same library, so a phrase you saved once can be used in all three. Press Tags to open it.

They are swapped for their real words before the model is asked anything, so it reads finished English and never sees an @ sign. Only your IDEA is read for tags: the instructions in the settings are left alone, because prose written for a model often contains things like "step #1" that are not meant to be tags.

@name

Your saved snippet, exactly as you wrote it.

*Category

One random tag out of that category, rolled fresh on every Run.

#name

One random line out of that tag, rolled fresh on every Run.

Reading the colours as you type

The colour tells you what a word will become, so you can see at a glance whether a tag is real before you Run anything.

Orange

An @tag that exists.

Green

A *category that exists and has something in it.

Violet

A #list that exists and has lines in it.

A red wavy underline

That name is not in your library, so it will be sent to the model as the literal text you typed. Every name looks like this while you are still typing it, and it takes its real colour the moment it matches.

Voice or Instrumental

The Music row switches between a song with singing and a piece of instrumental music. On Instrumental the node runs the model ONCE instead of twice, so it is about twice as fast: there are no words to write, so it only writes the caption and sends [Instrumental] as the lyrics.

• The verses, Bridge and Break controls dim on Instrumental, because there are no sections to ask for. They keep your settings and come back when you switch to Voice.
• Do not confuse Instrumental with the Break button. Break adds one instrumental section to a song that still has singing. Instrumental means no singing anywhere in the piece.
• Usually there is no voice at all, but sometimes a stray sound gets through: across four test renders in two very different styles, three had no voice and one had a single “aah” near the end. Re-roll the seed if you get one.
• It works because the caption never mentions voices at all. That is not an oversight: a caption that said no singing, no humming, no vowel sounds produced humming every time it was tried. Naming the thing summons it. If you write your own caption for an instrumental, say nothing about vocals rather than ruling them out.
• With no singer, something has to carry the tune, so the instrumental caption names one instrument that states the theme, develops it and returns to it at the end.

What the caption controls, and what the lyrics control

These are two different jobs and mixing them makes songs worse. The caption decides how the music SOUNDS. The lyrics decide the words and the order of the sections, and nothing else.

• Backing vocals, harmonies, instruments, tempo and production all go in the CAPTION. MiniMax does not support the round brackets some other music tools use for backing vocals: if you write (love) in the lyrics, the singer sings the word love. Ask for harmonies in the caption instead and you get them.
• Square brackets in the lyrics are SECTION TAGS and nothing else: [Intro] [Verse] [Pre-Chorus] [Chorus] [Post-Chorus] [Bridge] [Instrumental] [Solo] [Outro]. Writing a description inside them, such as [soft piano plays alone], does not give you soft piano. MiniMax's own guide warns that describing instruments inside the lyrics confuses where the sections start and end.
• A section tag on its own with no words under it means the band plays and nobody sings, which is how you get an intro or a break. It still uses up time.
• Do NOT use the Break button on a 30 second song. Tested by ear: the same song without it sang all four lines in 23 seconds, and with it the chorus was never sung at all and the song cut off. The instrumental time does not politely sit at the end either, it spreads through the whole song, so a 15 second opening arrives before the first word. There is more room at 60 seconds and above. Also avoid asking for a bridge and an instrumental section together at 60 seconds, because that is already five sections.
• The lyrics do not rhyme reliably, and that is the model rather than the wording. Asking it to rhyme was measured across ninety runs, along with a lower temperature and a worked example, and none of them beat chance. It rhymes about four line pairs in ten whatever you do, and one seed can give you a fully rhymed song where the next gives you none. Re-roll until you get one you like.

Good to know

• With no model chosen it passes your text straight through to both outputs, so you can drop it into a working graph and set it up afterwards without breaking anything. The banner always says which it is about to do.
• Two runs on one load is why it takes about twice as long as a single prompt node. The model is only loaded once.
• The lyrics are written knowing the caption, so the words match the mood the caption just described.
• Wire a model into the clip input and it is used instead of the one in the settings. Free VRAM is skipped then, because that model belongs to the loader you placed.
• The wording of both instructions is built in and was measured rather than guessed, so there is no formula to write. If you want to write your own, AI Prompt Pixaroma is the node for that.