YuE2 参考音色翻唱

The trickiest node in the pack, and the one people will search for by name. Give it a finished song and a short clean vocal sample, and it swaps the singing voice to match the sample, then puts the track back together.

It's a two-model pipeline bolted onto generation: Demucs separates the song into vocals and accompaniment, Seed-VC converts the vocal to your reference timbre, and then the two are remixed into a 48 kHz stereo FLAC. The pack needs the voice runtime plus its own Seed-VC and Demucs weights on top of the generation models - if the voice_conversion capability is off, the node tells you so and points at the voice runtime rather than failing weirdly downstream.

The reference sample is the whole ballgame

Required: 1–30 seconds of clear single-speaker dry vocal. Recommended: 5–25 seconds, no accompaniment, little reverb. That's from the author's own guide, and it's the honest constraint - zero-shot voice conversion is a similarity search over what it can hear, and if what it hears includes a snare drum and a room, that's part of the voice it'll learn.

Getting this right matters more than any dial below. A clean phone-voice-memo of someone singing a phrase beats two minutes of a live recording every time.

The dials

• diffusion_steps - default 30, range 4–50. Quality versus time. 30 is a sane default; dropping to 15 is a reasonable "is this even the right direction" test.
• timbre_strength - default 0.7, range 0–1.5. This is Seed-VC's inference CFG rate. Push it up for a stronger match and more artifacts; pull it down for something smoother and less convincing. When articulation goes weird, this is the first knob to back off.
• auto_match_pitch - default off. Turn it on when the reference sits in a different range than the song's original vocal and you want the engine to handle the transposition. Manual control is the next field.
• semitone_shift - default 0, −12 to +12. Note that this shifts the converted vocal's sung pitch, not the accompaniment. So a shift changes how the performance sounds against an unchanged backing track; it is not a key change.
• vocal_gain_db / accompaniment_gain_db - both default 0, −18 to +12. Straight scalar gains on the separated stem and the backing during the remix. Vocal drowned out? Raise this. Don't reach for timbre_strength to fix a level problem.

Outputs

audio (the remixed FLAC), result (for YuE2 导出工件), metadata. The artifact bundle keeps the intermediate stages too, which is how a failed or cancelled run can resume from the saved separation instead of redoing Demucs.

Install

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

Run install_runtime.bat once from the node directory - the voice stage pulls the Seed-VC checkpoint, its pitch extractor, Whisper, BigVGAN vocoder and Demucs weights, and the pack keeps all seven model subfolders plus the manifests. Restart ComfyUI, and run YuE2 模型服务 first. Registry: comfy node install yue2-t8. Windows 10/11 + NVIDIA; 24 GB VRAM is the recommendation for the install as a whole.

Where people get burned

It's a background-size job. The pack treats voice conversion as a persistent task with saved stages, so a long conversion can outlive your ComfyUI session - but killing the service process kills the task. Let it finish; progress lives on the WebUI's task page.

The memory budget is a ceiling, not a reservation. Setting it to 23.5 GiB doesn't reserve 23.5 GiB. If something else is on the card, you can still run out. The author says this outright in the guide.

Other apps reduce your headroom, and your patience will too. Demucs plus Seed-VC on a long track is slow, and the pitch-conversion stage is where the time goes.

Licensing and consent. YuE2's weights and first-party inference code are CC BY-NC 4.0 - non-commercial. Seed-VC is GPL-3.0, with the third-party components keeping their own terms. And the human side: use your own voice, or one you have explicit permission to use. A node that clones a voice from 30 seconds of audio is not a place to be casual about it.