Extensions/ComfyUI_RH_RVC

ComfyUI Extension

ComfyUI_RH_RVC

ComfyUI custom nodes for Retrieval-based Voice Conversion (RVC), including model loading, ZIP upload, voice conversion, and training.

By RH-RunningHub·Created 3 months ago·Updated 3 months ago· 0

RH-RunningHub/ComfyUI_RH_RVC

Nodes—

On cloudLocal install

Stars0

Updated3 months ago

Readme

ComfyUI RH RVC

ComfyUI custom nodes for Retrieval-based Voice Conversion (RVC). This plugin wraps inference and training code from Retrieval-based-Voice-Conversion-WebUI and exposes model loading, ZIP model upload, voice conversion, and one-click training nodes in ComfyUI.

Features

Load existing RVC .pth models from ComfyUI/models/RVC.
Upload an RVC model ZIP from the ComfyUI node UI and extract .pth plus optional .index files into ComfyUI/models/RVC/_uploaded.
Convert ComfyUI AUDIO inputs and return ComfyUI AUDIO outputs for downstream save, mix, or video nodes.
Run RVC preprocessing, F0/feature extraction, model training, and index training from a local dataset folder.
Save training outputs as <save_name>.pth, <save_name>.index, and <save_name>.zip under ComfyUI/output/RVC/<save_name>/.
Use HuBERT from ComfyUI/models/RVC/_assets/hubert.
Support optional RMVPE pitch extraction with ComfyUI/models/RVC/_assets/rmvpe/rmvpe.pt.
Keep model binaries outside the plugin repository.

Installation

Clone this repository into ComfyUI/custom_nodes:

cd ComfyUI/custom_nodes
git clone <repository-url> ComfyUI_RH_RVC
cd ComfyUI_RH_RVC
pip install -r requirements.txt

Restart ComfyUI after installation.

The plugin includes the RVC inference source under:

ComfyUI_RH_RVC/rvc_source

Set RVC_PROJECT_ROOT=/absolute/path/to/rvc_source only if you intentionally want to use a different RVC source tree.

Model Download and Installation

The plugin repository does not include model binaries. Put all RVC assets under ComfyUI/models/RVC.

Model Directory Structure

ComfyUI/
└── models/
    └── RVC/
        ├── _assets/
        │   ├── hubert/
        │   │   └── hubert_base.pt
        │   └── rmvpe/
        │       └── rmvpe.pt
        ├── your_voice_model/
        │   ├── your_voice_model.pth
        │   └── your_voice_model.index
        └── _uploaded/
            └── uploaded_zip_models_are_extracted_here/

Required files:

| File | Required | Destination | Description | | --- | --- | --- | --- | | RVC .pth voice model | Yes | ComfyUI/models/RVC/<model-folder>/ | Main voice conversion checkpoint. | | HuBERT | Yes | ComfyUI/models/RVC/_assets/hubert/hubert_base.pt | Content feature model. | | RVC .index | Optional | Same folder as .pth or any subfolder under models/RVC | Retrieval index for better timbre matching. | | RMVPE | Optional | ComfyUI/models/RVC/_assets/rmvpe/rmvpe.pt | Required only when f0_method=rmvpe. | | Pretrained G/D | Optional | ComfyUI/models/RVC/_assets/pretrained or _assets/pretrained_v2 | Training auto-detects these paths when the pretrained inputs are empty; missing files fall back to training from scratch. |

Download Methods

Method 1: Download core assets with aria2

Run these commands from ComfyUI/models/RVC:

mkdir -p _assets/hubert _assets/rmvpe
aria2c -x 16 -s 16 -k 1M -c -o _assets/hubert/hubert_base.pt \
  https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/hubert_base.pt
aria2c -x 16 -s 16 -k 1M -c -o _assets/rmvpe/rmvpe.pt \
  https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/rmvpe.pt

Method 2: Download core assets with hf

cd ComfyUI/models/RVC
mkdir -p _assets/hubert _assets/rmvpe
hf download lj1995/VoiceConversionWebUI hubert_base.pt --local-dir _assets/hubert
hf download lj1995/VoiceConversionWebUI rmvpe.pt --local-dir _assets/rmvpe

Method 3: Manual download

| Model | Link | Destination | | --- | --- | --- | | HuBERT | https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/hubert_base.pt | ComfyUI/models/RVC/_assets/hubert/hubert_base.pt | | RMVPE | https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/rmvpe.pt | ComfyUI/models/RVC/_assets/rmvpe/rmvpe.pt |

Bring your own RVC .pth and optional .index voice model files. The .pth and .index file names do not have to be identical, but using the same stem is recommended because it makes automatic matching more reliable.

Model Selection Guide

| Use case | Recommended setup | Notes | | --- | --- | --- | | Basic voice conversion | .pth + HuBERT | Works without an index, but quality may be lower. | | Better timbre matching | .pth + matching .index + HuBERT | Put the .index near the .pth or use the ZIP loader. | | RMVPE pitch extraction | .pth + HuBERT + RMVPE | Required only when f0_method=rmvpe. |

F0 Method Notes

| Method | Extra requirement | | --- | --- | | harvest | Uses pyworld; installed from requirements.txt. | | dio | Uses pyworld; installed from requirements.txt. | | pm | Uses praat-parselmouth; installed from requirements.txt. | | crepe | Uses torchcrepe; installed from requirements.txt. | | rmvpe | Requires models/RVC/_assets/rmvpe/rmvpe.pt. |

Usage

Example Workflow

Import this workflow into ComfyUI:

The workflow demonstrates:

Load or upload an RVC model ZIP with RunningHub RVC ZIP Model Loader.
Load audio, trim it to a default end time of 20 seconds, and separate stems with AudioSeparation.
Convert the vocal stem with RVC.
Mix the converted vocal back with the other stems.
Save the final audio with a standard audio save node.

The training example uses RunningHub RVC One-Click Train. Training data can be provided in two ways:

trainset_dir: dataset audio folder, empty by default; if filled, it takes priority.
audio: optional audio input, used when trainset_dir is empty. It accepts ComfyUI AUDIO, lists of AUDIO, audio file paths, or path lists from upstream nodes.

The training node enables audio_auto_clean by default. It converts training audio to mono, trims silence, skips clips that are too short or too quiet, and peak-normalizes the kept clips before RVC preprocessing. If the source has backing music, enable auto_extract_vocals to run Demucs/HDEMUCS vocal extraction first; keep it disabled for dry vocal datasets to avoid extra separation artifacts. Multi-speaker, heavy-reverb, or noisy sources should still be cleaned upstream so training sees clean single-speaker vocals.

trainset_dir examples:

Absolute path: /workspace/ComfyUI/input/my_rvc_trainset
Relative to ComfyUI input: my_rvc_trainset

The save_name input controls the final output file names. A completed run writes:

ComfyUI/output/RVC/<save_name>/<save_name>.pth
ComfyUI/output/RVC/<save_name>/<save_name>.index
ComfyUI/output/RVC/<save_name>/<save_name>.zip

Node Reference

RunningHub RVC Model Loader

Loads an existing .pth model from ComfyUI/models/RVC and returns an RVC_MODEL handle.

RunningHub RVC ZIP Model Loader

Uploads or selects a .zip file from ComfyUI input and extracts .pth plus optional .index files into ComfyUI/models/RVC/_uploaded.

Upload limit: 150 MB per ZIP.

RunningHub RVC Voice Conversion

Converts a ComfyUI AUDIO input with an RVC_MODEL and returns converted AUDIO plus runtime information. Use ComfyUI audio save nodes for file output.

RunningHub RVC One-Click Train

Runs the complete RVC training pipeline from a dataset folder or optional audio input. The node returns info, which includes output .pth, .index, .zip paths plus a log summary. It also reports the final .zip as a ComfyUI output file so RunningHub post-processing can upload it to COS. save_name controls the final output files; experiment_name controls the training log and intermediate folder.

Notes:

f0_method=rmvpe requires ComfyUI/models/RVC/_assets/rmvpe/rmvpe.pt; use harvest if RMVPE is not prepared.
Training requires HuBERT at ComfyUI/models/RVC/_assets/hubert/hubert_base.pt.
Runtime and VRAM use depend on dataset length, batch_size, total_epoch, and GPU.
The output ZIP can be used directly with the ZIP Model Loader and contains same-name .pth plus optional .index.

License

This project is licensed under the Apache License 2.0. Redistribution and modified versions must retain the license and attribution notices, including the notices in NOTICE.

The bundled RVC inference source is based on Retrieval-based-Voice-Conversion-WebUI, which is distributed by its upstream project under the MIT License.

Source and Related Links

RunningHub China
RunningHub International
Plugin repository: ComfyUI_RH_RVC
Upstream repository: Retrieval-based-Voice-Conversion-WebUI
Core model assets: lj1995/VoiceConversionWebUI
Plugin license: Apache License 2.0
Upstream license: MIT License

Acknowledgements

This project is based on Retrieval-based-Voice-Conversion-WebUI by the RVC Project contributors.