XB-BOX - 🎵 YuE2 乐谱工作台

Most music models give you a black box: prompt in, audio out, bad bar in the middle, re-roll and pray. YuE2 isn't built that way. It plans an editable ABC score first and renders audio from that score, so there's an intermediate object you can read and edit - which is exactly what makes this node useful. XB-BOX - 🎵 YuE2 乐谱工作台 (Score Workbench) takes a score and does a handful of symbolic operations on it, then hands it back. No render, no re-roll.

One thing to internalise before you start: the score carries no instruments. If you want a guitar instead of a piano, that's the Style Prompt node's job. The workbench only touches notes, chords and timing.

Inputs and outputs

The required inputs are 操作 / Operation and its sidekicks, plus the editing box; abc is the optional upstream score (the abc output of YuE2GenerateABC, or whatever the transcription intake gave you). Outputs are abc - for the official music node, or back into another pass - and 报告 / Report.

The report's first line is the mode suggestion. That's the practical bit: read it before you render, because it tells you whether this score wants full (harmony present) or melody (a bare melody line). Widgets that don't apply to your chosen operation hide themselves, so the panel stays short.

The five operations

只查看 / Inspect validates the score and writes a report without touching it. The most valuable diagnostic here is truncation: a score cut off mid-phrase fails structure validation outright, which is how you find out the planner ran out of tokens before you blame the renderer.

去掉和弦 / Strip Chords deletes the chord symbols from the Vocal voice. Do that and the downstream mode has to be melody, since melody mode rejects any chord symbol. Your accompaniment gets regenerated freely in whatever style you chose.

人声移到器乐轨 / Vocal to Ins moves the sung melody onto the Ins track and leaves Vocal holding chords and rests. Only the melody moves; nothing about instruments changes. Two conditionals appear with it:

• 和声处理 / Harmony - keep the chords on the score (then downstream mode must be full, and you can keep editing the harmony) or drop them (then melody).
• 重叠处理 / Overlap - what to do when the moved melody collides with existing Ins notes. Vocal Priority is the official default: Ins gets shortened to make room. Error Out fails instead, so you can go fix the collision by hand rather than silently clipping someone's riff.

用编辑框 / Use Edited takes a score you edited by hand in 编辑框 / Edited ABC, revalidates it, and diffs it against the upstream score so you can see exactly what you changed.

编译事件 / Compile Events is the odd one out: paste an official events.json into the edit box and it compiles a legal score from the events, using the same event compiler the instrumental path uses. The format, per the widget's own tooltip, with times measured in quarter notes as integers or fraction strings like "3/2":

{"bpm":81,"key":"Bm","bars":[{"meter":"4/4","section":"verse"},{}],
 "notes":[[0,1,71],[1,"1/2",74]],
 "chords":[[0,"Bm"]]}

Single melody only - no stacked or overlapping notes. Out-of-range pitches, conflicting chords and durations the notation can't express all raise errors rather than getting quietly rounded off.

Where people get burned

• A score still holding chords fed into melody mode. Immediate rejection, and it looks like the render node is broken.
• Expecting the diff to prove musical correctness. It proves the notes you intended are in the score you're rendering. It says nothing about whether the score sounds like what you imagined.
• Reaching for the workbench to change instruments. It won't. Style node, every time.
• Assuming the error messages are the node's own. They're the official validator's, passed through verbatim - which is deliberate, and worth reading rather than skimming.

Install

Search XB_ToolBox in ComfyUI Manager, or clone it and restart:

cd ComfyUI/custom_nodes
git clone https://github.com/wjluoxiao/XB_ToolBox.git

You'll need a ComfyUI build with the native YuE2GenerateABC / YuE2GenerateMusic nodes, since this only emits strings. The ABC parsing is an in-node port of the official bounded-dialect implementation and uses the standard library only, so it works offline and doesn't drag in the pack's optional dependency pile.