ComfyUI UVR MDX

ComfyUI custom node for separating instrumental and vocal stems with the bundled UVR-MDX-NET-Inst_HQ_5.onnx model.

Simplified Chinese README

Features

• Ready to use inside ComfyUI without depending on the UVR desktop application
• Supports auto, cuda, directml, and cpu execution providers
• Splits input audio into instrumental_audio and vocals_audio
• Automatically resamples audio to 44.1 kHz for inference and restores the original sample rate on output
• Handles mono input by duplicating it to stereo before separation

Node

• Node name: UVR MDX Inst HQ5 Separator
• Category: audio/uvr

Inputs

• audio: ComfyUI AUDIO
• device: auto / cuda / directml / cpu
• segment_seconds: chunk size for long audio, default 15
• margin_seconds: overlap between chunks, default 1
• denoise: enables the common UVR denoise inference trick
• compensation: output gain compensation, default 1.01
• batch_index: which item to read from a batched AUDIO input

Outputs

• instrumental_audio
• vocals_audio
• info

Requirements

ComfyUI usually already provides torch. This node additionally needs:

• numpy
• One ONNX Runtime package:
  • onnxruntime for CPU
  • onnxruntime-gpu for NVIDIA CUDA
  • onnxruntime-directml for DirectML on Windows

Installation

1. Copy or clone this repository into your ComfyUI/custom_nodes/ folder.
2. Make sure the model file exists at models/UVR-MDX-NET-Inst_HQ_5.onnx.
3. Install dependencies in the same Python environment used by ComfyUI:

pip install -r requirements.txt

If you want GPU acceleration, replace the default ONNX Runtime package with the variant that matches your system.

Repository Layout

comfyui_uvr_mdx/
|- __init__.py
|- nodes.py
|- README.md
|- README.zh-CN.md
|- requirements.txt
|- LICENSE
`- models/
   `- UVR-MDX-NET-Inst_HQ_5.onnx

Model Attribution

This node uses the bundled UVR-MDX-NET-Inst_HQ_5.onnx model from the Ultimate Vocal Remover (UVR) distribution.

Sources:

• Official website: ultimatevocalremover.com
• Official repository: Anjok07/ultimatevocalremovergui

According to the UVR repository, the models provided in the package were trained by the UVR core developers unless otherwise noted, and third-party developers using UVR models should provide credit to UVR and its developers.

MDX-Net architecture credits listed by UVR:

• Kuielab
• Woosung Choi

Please verify upstream model redistribution terms before republishing bundled weights.

Notes

• The node loads the model from its local models/ directory and does not rely on absolute paths.
• The bundled model file is large enough that you may prefer GitHub Releases or another distribution method in the future, but it is still below GitHub's 100 MB file limit.
• The LICENSE file applies to the source code in this repository unless otherwise noted.
• Before redistributing the bundled model, verify the upstream model's license and redistribution terms.

Troubleshooting

• Bundled model not found: place UVR-MDX-NET-Inst_HQ_5.onnx inside models/
• This node requires onnxruntime: install a compatible ONNX Runtime package in the ComfyUI Python environment
• Provider errors on cuda or directml: switch device to auto or install the correct runtime package for your hardware

License

MIT for the repository source code. See LICENSE.