ComfyUI_YingMusic-SVC

ComfyUI_YingMusic-SVC

---

This project is based on the code from https://github.com/yourusername/ComfyUI_YingMusic-SVC.git. It has been adapted into a ComfyUI custom node and is intended for use within ComfyUI workflows.

ch01. Installation

1. Clone the repository into the ComfyUI custom_nodes directory:

cd /path/to/ComfyUI
git clone https://github.com/yourusername/ComfyUI_YingMusic-SVC custom_nodes/ComfyUI_YingMusic-SVC

# cd custom_nodes/ComfyUI_YingMusic-SVC
# git mv requirements.txt requirements.txt.bk
# git mv README.md README.md.bk
# awk -F "[ =<>]" '$1=="torch" || $1=="torchvision" || $1=="torchaudio"{next} {print $1} \
#   END{print "\ndemucs\nrequests"}' requirements.txt.bk > requirements.txt
# ls -1 README.md __init__.py nodes.py

2. Install dependencies in the ComfyUI environment

conda install -c conda-forge -y sox      # @conda
# sudo apt install -y sox libsox-fmt-all # @debian/venv

pip install -r custom_nodes/ComfyUI_YingMusic-SVC/requirements.txt

Note: libsox is required for audio I/O and resampling via torchaudio.

3. Download models

• install commandline tool hf and create the models dir

pip install -U "huggingface_hub[cli]"
mkdir -p ./models/YingMusic-SVC

• Download GiantAILab/YingMusic-SVC

hf download --local-dir ./models/YingMusic-SVC GiantAILab/YingMusic-SVC YingMusic-SVC-full.pt

• Download funasr/campplus

hf download --local-dir ./models/YingMusic-SVC/funasr/campplus funasr/campplus

• Download lj1995/VoiceConversionWebUI

hf download --local-dir ./models/YingMusic-SVC/lj1995/VoiceConversionWebUI \
  lj1995/VoiceConversionWebUI rmvpe.pt

• Download nvidia/bigvgan_v2_44khz_128band_512x

hf download --local-dir ./models/YingMusic-SVC/nvidia/bigvgan_v2_44khz_128band_512x \
  nvidia/bigvgan_v2_44khz_128band_512x bigvgan_generator.pt config.json

• Download openai/whisper-small

hf download --local-dir ./models/YingMusic-SVC/openai/whisper-small openai/whisper-small

• Download for htdemucs for Demucs

mkdir -p ~/.cache/torch/hub/checkpoints

wget -O ~/.cache/torch/hub/checkpoints/955717e8-8726e21a.th \
  https://dl.fbaipublicfiles.com/demucs/hybrid_transformer/955717e8-8726e21a.th

4. Copy the configuration files

• Copy the configuration file to models/configs(the filename must be converted to lowercase):

cp custom_nodes/ComfyUI_YingMusic-SVC/configs/YingMusic-SVC.yml models/configs/yingmusic-svc.yaml

ch02. Usage

1. Add nodes to your workflow

• Add a YingMusicSVC: Load Models node as LoadModels
• Add a Upload Audio(Upload) node as UploadAudio
• Add a Upload Audio(Upload) node as UploadVolcals
• Add a YingMusicSVC: Audio Seperator node as AudioSeperator
• Add a YingMusicSVC: Voice Convert node as VoiceConvert
• Add a Preview Audio node as PreviewVocals
• Add a Preview Audio node as Preview Audio
• Add a Preview Any node as VoiceConvert_Metrics

2. workflow

• docs/workflow.a01.png

3. Performing Voice Conversion

• Connect LoadModels model_bundle -> Voice_Convert model_bundle
• Connect UploadAudio audio -> AduioSeparator audio
• Connect AduioSeparator vocals -> VoiceConvert input_vocals
• Connect AduioSeparator instructmental -> VoiceConvert instructmental_audio
• Connect UploadVocals audio -> VoiceConvert ref_vocals
• Connect VoiceConvert vocals -> PreviewVocals audio
• Connect VoiceConvert audio -> PreviewAudio audio
• Connect VoiceConvert metrics -> VoiceConvert_Metrics source
• Adjust parameters as needed:
  • diffusion_steps: Number of diffusion steps (higher = better quality but slower). Typical range: 20–60.
  • inference_cfg_rate: Classifier-free guidance rate. Controls how strongly the model follows the reference (target) timbre/style versus the source content.
    • Lower: more “source-like” / less reference influence
    • Higher: more “target-like” / stronger reference influence
    • Typical: 0.5–1.0.
  • f0_condition: Enable/disable pitch (F0) conditioning.
    • True: better pitch accuracy / singing & expressive speech usually improves
    • False: can sound smoother for some speech, but pitch may drift
  • semi_tone_shift: Shift the output pitch in semitones.
    • Positive = higher pitch, negative = lower pitch
    • 0 (or None): no manual shift (uses adaptive/auto pitch adjustment if enabled in your pipeline)
  • vocal_gain: Gain applied to the converted vocal before mixing with accompaniment.
    • 1.0 = unchanged
    • Increase to make vocal louder in the final mix (e.g. 1.2–1.8)
  • instructmental_gain: Gain applied to accompaniment before mixing.
    • 1.0 = unchanged
    • Decrease to duck the backing track (e.g. 0.6–0.9) so vocals sit on top
  • save_debug_wav: Save intermediate/final WAV files to the ComfyUI output folder for debugging.
    • False: no files written, output only flows through the graph
    • True: writes a timestamped WAV (and optionally an additional mixed/accompany version, if you kept that logic)
  • fp16: Use half precision (float16) inference to reduce VRAM and speed up on NVIDIA GPUs.
    • Recommended True on CUDA GPUs
    • Automatically should be forced to False on CPU (fp16 typically doesn’t help and may break)

4. Debug Output(save_debug_wav=true), output/YingMusicSVC/:

• mixed_{pitch_shift}.{ts}.wav
• accompany.{ts}.wav
• vc_{pitch_shift}.{ts}.wav

ch04. Notes

• The model requires significant GPU memory for inference
• For best results, use high-quality input audio
• The reference audio should be clean speech without background noise

ch05. Troubleshooting

• If you encounter CUDA out of memory errors, try reducing the batch size or using a smaller model
• Make sure all required model files are downloaded and in the correct locations
• Check the ComfyUI console for error messages

ch06. License

This project is licensed under the same terms as the original YingMusic-SVC model.