YuE2 音频转谱

This is the node that makes the whole "keep the tune, change everything else" workflow possible. Give it a song - any song you have as a file - and it produces ABC notation and MIDI, optionally rendered as an actual score you can look at. Feed that to YuE2 旋律重制 and you have a cover with new lyrics and a new arrangement, sung in YuE2's own voice.

The pitch is not "rip the melody perfectly". It's "get a machine-readable melody you can correct".

Inputs

• model - from YuE2 模型服务.
• audio - one audio item. Batch more than one and the node raises rather than guessing; split first. WAV, FLAC, MP3, M4A, OGG and AAC are all accepted (FFmpeg handles the decode).
• melody_only - default true, and you probably want to leave it there. This keeps the melody lines and drops the chords, which is what the cover node downstream expects; with it on, the worker errors out loudly if it couldn't produce a usable melody ABC instead of handing you a half-empty score.
• render_score - none, pdf, png or svg. Nicety, not a necessity: the score is generated offline by the bundled renderer, so it needs no network and no LaTeX install.

What comes out

transcription - a YUE2_TRANSCRIPTION handle, which is what wires into YuE2 旋律重制. It carries the ABC text, an error field, and the melody-only flag, so a downstream failure can tell you why there was no melody. abc - the score as a plain string. metadata - job JSON.

The engine underneath is SheetSage2 doing the transcription with MERT-v2 as the feature model - that's the pair downloaded at install. It requires the pack's transcription capability, meaning the transcription models, the isolated runtime and FFmpeg all present. If they aren't, you get a straight "转谱环境未就绪" and a pointer to re-run the installer.

Be honest about accuracy

The author's own guide says transcription may have wrong pitch, wrong time signature and wrong segmentation, and that you must check it against the original audio by ear. That's not false modesty - it's a model doing automatic music transcription, which is a genuinely hard problem, on a mix where the melody is buried under drums and reverb.

And it doesn't give you lyrics. Those you paste in yourself, organised into sections. The pipeline assumes a human is standing between the transcription and the cover.

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 fetches the SheetSage2 and MERT weights plus the offline score renderer - then restart ComfyUI. Registry: comfy node install yue2-t8. Windows 10/11 + NVIDIA GPU; 24 GB VRAM and ~60 GB disk recommended for the whole install (transcription itself is light next to generation).

Where people get burned

"乐谱渲染器未安装". If you built the runtime with the -SkipRenderer flag, score rendering is gone and only none works. Re-run the installer without that flag.

An empty ABC. With melody_only on and nothing usable detected, the node fails rather than handing over garbage - which is correct behaviour, but it means a quiet or heavily processed track can simply refuse to transcribe. Try a cleaner source, or a section of the song with the melody exposed.

Downstream silence. YuE2 旋律重制 will not invent a melody: if the transcription has no ABC and you didn't paste a corrected one into its text field, it raises. The abc output here is a string specifically so you can put eyes on it first - that hand-off is the whole workflow.

Source-length patience. The pack is a full-song system, so it will happily chew through a four-minute track. Don't test-tune on a whole album.