YuE2 Load MIDI

Everyone who writes songs has a folder of MIDI files: something you sequenced years ago, a karaoke .kar, a notation export of a melody you actually like. YuE2 Load MIDI turns any of those into something YuE2 can sing - it reads the file on the spot, downloads nothing, and a three-track file took 0.017 seconds in the author's own measurement.

It's a different job from the rest of the pack: YuE2 Transcribe listens to a recording and guesses what it heard, while this one reads notes that are already exact.

Where this fits, honestly

Local music generation in ComfyUI mostly means ACE-Step - great instrumentals, weak vocals. YuE2 is the other half of that deal, and Load MIDI is the branch for when the tune exists and only the performance is missing.

How a file becomes a singable score

The node reads the notes out of one track as the vocal line and another as the instrumental line, then folds them into the ABC score YuE2 sings from. Some of that folding is lossy, and the node tells you rather than hiding it:

• A score keeps one tempo, so a file with tempo changes gets them averaged, and the node says so. Bars come from the file's meters, and a file played in by hand is rounded to sixteenths with a notice.
• Each score line holds one note at a time, so of notes struck together the top one is kept and the node reports how many it dropped. Chords in the vocal track are sung as their top note.
• A line whose middle pitch sits outside C4 to A#5 is moved by whole octaves, because YuE2 sings the octave it is given. The key comes from the file's key signature where it fits the notes and is estimated otherwise, and markers named like sections become sections.

The settings that matter

midi is a file in ComfyUI's input folder - .mid, .midi, .kar or .rmi. Choose MIDI file... on the node uploads one, and so does dropping a file onto the node; the dropdown is empty until a file is there.

mode is melody (default) or full: the first writes the two lines without chords, the second adds chord symbols guessed from the other tracks - and each pairs with cot at melody or full on the song node. Those chords are a guess; read them over in Edit score....

vocal_track and instrument_track default to auto, and the list printed on the node shows you what auto picked, and why, before anything runs: the track a karaoke file's words fall on, then one named as the voice or the melody, then the highest line that is not a bass. instrument_track can also be none.

without_sections is on by default and is the one to understand. On, the score goes out as a bare tune with no section comments, and YuE2 Generate Song lays your lyrics along the tune's phrases before singing - splitting at breaths and bar lines, matching each line to a phrase with about as many notes as it has syllables. That matters: sung with its own sections, a wordless tune loses the words (0 of 46 words heard in order over one test theme). Turn it off when you are singing a karaoke file with its own words.

Outputs are score_abc (wire it into YuE2 Generate Song's score_abc) and lyrics (a karaoke file's words under their section tags, or a bare [Verse] to write under).

Install

The pack's install is the usual custom-node drill - its only hard dependency is tiktoken, and the models are fetched on first use by the song node, not this one:

cd ComfyUI/custom_nodes
git clone https://github.com/pytraveler/YuE2-ComfyUI
# then restart ComfyUI

Nine workflows ship with the pack under Workflow → Browse Templates; template 7, Sing a MIDI file, is this node already wired up.

Where people get burned

The older YuE-in-ComfyUI experience was rough enough that people spent a week on it and got 90 seconds of silence. What this pack changes is mostly bookkeeping: a bad score is refused with the reason on the node (No score could be written from 'x.mid': the file has no notes, for instance) instead of failing downstream. Usual suspects: a file that is not really MIDI despite the extension, a drum track picked by hand for the voice (drums have no tune - the node says so), and a lyric that fights the tune, which earns a warning about lines having half again as many notes as syllables. With max_seconds at 0, a tune-led song stops a little past the tune's own length, and the node reports which bars never got sung.