YuE2 训练 RVC 音色

Training a voice model from inside ComfyUI is the odd part of this node, and the odd part is the point: it has no training parameters. No epochs, no sample rate, no batch size. You pick a project you already built in the pack's studio, tick a rights checkbox, and it runs. Everything that decides what gets trained was decided earlier, in a place with play buttons and progress bars. Expecting a trainer node? This is a reproducible launcher for a run you prepared by hand.

How it works

The node does nothing model-related itself. It ensures the worker on 127.0.0.1:8189 is up, checks that RVC training components are installed, then calls the project's preflight endpoint before it will start anything. If preflight isn't clean it raises with the actual reasons joined together - missing base models, unlistened material, not enough disk. Only then does it submit the training job and wait.

The pipeline behind it is standard RVC: material is sliced into segments, pitch is extracted with RMVPE, HuBERT features are computed, the model trains against the v1 or v2 pretrained generator/discriminator for your chosen sample rate, and a FAISS retrieval index is built at the end. The studio reports progress per stage - preprocessing, pitch, features, training, index - so a long run isn't a black box. That index is what index_rate in the cover node blends against later.

The parameters live in the workbench for discipline, not laziness: material, settings and the rights confirmation get frozen into the project, so a run is something you can describe later. And voice training is where the community's advice is boringly consistent - ten minutes or more of clean, consistent singing beats a clever parameter sweep, which is why this pack makes the data step the only step you control.

Inputs and outputs

training_project is a combo read from the studio's RVC projects - name plus a 32-character ID. If nothing's been prepared, the single entry literally reads 请先在 YuE2 工作台建立 RVC 训练项目, which is the author telling you where to go.

confirmed_materials_and_rights is a BOOLEAN, default off. Leave it off and the node refuses to run: 请先在工作台逐段试听素材、确认训练权利，再勾选确认. It's not a legal formality - it's the pack's guard against the single most common way people make a bad voice model, which is training on tracks they never actually listened to segment by segment. Bleed, reverb tails and clipping all get learned.

The node is marked as an output node, so it runs as a terminal in your graph even if nothing consumes its result. voice is a YUE2_RVC_VOICE handle - the same type the loader emits - so you can wire it straight into YuE2 RVC 翻唱 and never touch a dropdown. It points at the first speaker of the trained model. metadata is a STRING with the job record.

Preparing the project (this is the actual work)

Open the studio's 我的音色 / 训练 workspace, create a project, import your material, and play every segment. Separate the backing music off any track that still has it, and mark vocals-only material as such - the pack requires it, because a model trained on a vocal with accompaniment baked in learns to output screech and hum. The guide suggests 10–50 minutes of clean, timbre-consistent material and says to check noise, reverb and target range before committing. Then pick settings, run 检查训练条件, and train. Changing material or architecture later means a new project.

The shipped example is a real reference point: the CSD Korean Female v1 voice was trained on 46 segments, 68.5 minutes, 100 epochs. That's the shape of a working dataset, not a minimum.

Install

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

Then install_runtime.bat once from the node directory and restart ComfyUI. That script is what pulls the RVC pretrained base models (both network versions at 32k/40k/48k, plus HuBERT and RMVPE) that training needs - the node will flatly refuse to start without them. Manager users can search the pack title. Windows + NVIDIA, ~24 GB VRAM and 60 GB free disk recommended; the training data and checkpoints are extra.

Where it goes wrong

• Preflight says material isn't reviewed. You have to open each segment in the workbench and mark it. There's no bulk "yes, all fine" because that's exactly the shortcut that produces a bad model.
• 训练目录空间不足. The preflight estimates what the run needs on disk and compares it to free space. Free some up; failing midway leaves you resuming instead of finishing.
• Under five minutes of material. A warning, not an error: fine for checking that the plumbing works, not for judging quality, and judging a voice at 20 epochs is a waste of an afternoon.
• Missing pretrained base for the version/sample rate you picked. Hard error, named file. Rerun the installer.
• It stops halfway. Cancel keeps the saved checkpoints and the project continues from them, reusing the finished stages. You can't switch material mid-flight, though - new project, new run.
• The project list is stale. Refresh the ComfyUI page after creating or renaming a project.

Finally, the licensing note that comes with anything trained in this pack: YuE2's weights and first-party inference code are CC BY-NC 4.0, i.e. non-commercial, and the pack's own example voice is CC BY-NC-SA 4.0. RVC itself is a separate project with its own terms, but an exported voice inherits whatever constraints your source material carries. Read the per-file manifests in the model directory rather than trusting a summary - including this one.