YuE2 旋律重制

"旋律重制" is a better name than the English one it usually gets translated to. This node takes a melody you already have - normally a transcription of an existing song - and regenerates a whole track around it in a new style, with new lyrics, sung in YuE2's own voice.

It's the difference between "write me a song" and "write me a different song using this tune".

Inputs

• model - from YuE2 模型服务.
• transcription - the YUE2_TRANSCRIPTION output of YuE2 音频转谱. This is where the melody comes from.
• style - the target genre and instrumentation, in prose. This is the field that makes it a remake rather than a re-render: same tune, different band.
• lyrics - your new words, with [Verse] / [Chorus] tags. Not extracted from the original; nobody does that here, and the transcription node doesn't try.
• seed - the usual re-roll lever.
• reviewed_abc (optional) - your corrected score. This wins over the transcription's own ABC whenever it's non-empty.

The mechanism

Under the hood this node calls the same generation pipeline as YuE2 生成歌曲, but locked down: melody-only planning, CFG at 1.0, exactly one candidate. So it's not a special model - it's the full YuE2 pipeline conditioned on a melody you didn't let it invent.

Before it does that, it picks the ABC: your reviewed_abc if you typed one, otherwise the transcription's. If neither exists, it raises and surfaces the transcription's own error. Which is the right behaviour and also the single most common failure here - a bad transcript upstream looks like a broken cover node downstream.

Chaining it

The shipped 04_参考音色翻唱.json workflow wires Load Audio → YuE2 音频转谱 → this node → Preview Audio, with a second branch for voice conversion. Load it and follow the links. The chain is: transcribe the original, eyeball the ABC, fix what's wrong, paste the corrected version into reviewed_abc, set your style and lyrics, generate.

Why the manual step matters more here than anywhere else in the pack: the melody is now a condition, and the model will faithfully sing whatever errors you feed it. A wrong key in the transcript doesn't get smoothed over - it becomes the song.

Also worth knowing: the studio's own assistant is strict about what counts as a sendable score, and it enforces specific checks (stripping chords, keeping melody and rhythm consistent across voices). The ComfyUI path doesn't have that safety net. You're the validator.

Outputs

audio (48 kHz stereo, straight to Preview/Save Audio), result (for YuE2 导出工件), metadata (job JSON with timings and the truncation flag).

Install

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-YuE2-T8.git

Run install_runtime.bat once from the node directory - it grabs both the generation models and the SheetSage2/MERT transcription pair, since this workflow needs both - then restart ComfyUI and pass YuE2 模型服务 first. Registry: comfy node install yue2-t8. Windows 10/11, NVIDIA, 24 GB VRAM and 60 GB free disk recommended.

Where people get burned

Melody-only, enforced. Full transcriptions carry chords, and chords aren't what this node conditions on. If your ABC has them, strip them before pasting. With melody_only on in the transcription node (the default) they're dropped for you.

AI vocal, not your vocal. This node uses YuE2's own singing voice. If you wanted the reference timbre, that's the next node in the pair - YuE2 参考音色翻唱, which takes this node's output and runs Demucs plus Seed-VC over it.

Everything is non-commercial. YuE2's weights and first-party inference code ship under CC BY-NC 4.0. A remade melody is still YuE2 output. And the melody itself may belong to someone else - this node makes it trivially easy to do something you shouldn't, so don't.

One candidate, no batch. Unlike the song node, there's no candidates field. Re-queue with a new seed if you didn't like it; the plan reuse means the melody stays put either way, which is the good news.