comfyui_uvr_mdx Introduction
comfyui_uvr_mdx adds an audio-separation node to ComfyUI. Give it a piece of music, and it produces two audio outputs: one intended to contain the instrumental backing and one intended to contain the vocals. You can use those outputs as separate parts of an audio workflow instead of working with a single mixed recording.
The extension runs inside ComfyUI and does not require the Ultimate Vocal Remover (UVR) desktop application. It uses one bundled model, UVR-MDX-NET-Inst_HQ_5.onnx, for the separation.
How comfyui_uvr_mdx Works
A finished song combines sounds that you hear together. The separation model examines that combined audio and estimates which parts belong in the instrumental stem and which belong in the vocal stem. Think of it as making two working layers from a mixed image: the layers can be useful independently, even though the original was supplied as one piece.
In ComfyUI, the UVR MDX Inst HQ5 Separator node takes an AUDIO input and returns instrumental_audio, vocals_audio, and info. For a long recording, it processes the audio in chunks. A small overlap between chunks gives the node audio on both sides of each boundary.
The node prepares audio at 44.1 kHz for the model, then restores the original sample rate for its outputs. If the input is mono, it duplicates it into stereo before separation. These steps happen automatically.
comfyui_uvr_mdx Features
Find the node under audio/uvr in ComfyUI. Its controls let you choose how the audio is processed:
| Input | What it does | When you might change it |
|---|---|---|
audio | Supplies the ComfyUI AUDIO item to separate. | Connect the recording you want to split. |
device | Chooses where inference runs: auto, cuda, directml, or cpu. | Use auto to let the node select a provider, or choose one that matches your available runtime and hardware. |
segment_seconds | Sets the chunk length for long audio. The default is 15 seconds. | Adjust it when you need a different chunk size for a longer recording. |
margin_seconds | Sets the overlap between chunks. The default is 1 second. | Adjust how much neighboring audio is shared at chunk boundaries. |
denoise | Enables a common UVR denoise inference technique. | Compare results with it on and off to see which you prefer for your recording. |
compensation | Adjusts output gain. The default is 1.01. | Change it if you need to adjust the output level. |
batch_index | Selects which item to read from a batched AUDIO input. | Set it to the item you want when your input contains a batch rather than one recording. |
The two named audio outputs let you continue working with the instrumental and vocal stems separately. The node also returns an info output; the provided documentation does not specify its contents.
comfyui_uvr_mdx Models
The extension includes one model: UVR-MDX-NET-Inst_HQ_5.onnx. It is the model used to estimate the instrumental and vocal stems; the supplied documentation does not describe a model-selection option or other bundled models.
The model comes from the Ultimate Vocal Remover distribution. The author credits UVR and its developers for the model and notes that UVR lists Kuielab and Woosung Choi in its MDX-Net architecture credits. For background, see the Ultimate Vocal Remover website (https://ultimatevocalremover.com/) and official UVR repository.
Troubleshooting comfyui_uvr_mdx
The node reports “Bundled model not found.”
- Check that
UVR-MDX-NET-Inst_HQ_5.onnxis present in the extension’smodels/folder. - Check the filename carefully, including the
.onnxending. - Run the node again once the model is in the expected location.
The node reports “This node requires onnxruntime.”
- Confirm that a compatible ONNX Runtime package is available in the Python environment ComfyUI uses.
- Run the node again after that environment has the required package. The runtime variants named in the original documentation are
onnxruntimefor CPU,onnxruntime-gpufor NVIDIA CUDA, andonnxruntime-directmlfor DirectML on Windows.
The node reports a provider error with cuda or directml.
- Change
devicetoautoand try again. - If you specifically need the selected provider, check that its matching ONNX Runtime variant is available for your hardware and ComfyUI environment.
Frequently Asked Questions
Do I need the UVR desktop application?
No. This is a ComfyUI node that runs without the UVR desktop application.
Can I provide mono audio?
Yes. The node duplicates mono input into stereo before separation.
Do I need to convert my recording to 44.1 kHz first?
No. The node resamples audio for inference and restores the original sample rate on output.
Can I choose a different separation model?
The supplied documentation describes only the bundled UVR-MDX-NET-Inst_HQ_5.onnx model, not a model-selection feature.
Which output should I use?
Use instrumental_audio when you want the estimated backing track, and vocals_audio when you want the estimated vocal stem. Listen to both outputs to decide how well they serve your project.
Learn More about comfyui_uvr_mdx
- Ultimate Vocal Remover website (https://ultimatevocalremover.com/): Background on the UVR project behind the bundled model.
- Official Ultimate Vocal Remover repository: Project information and model attribution.
- Simplified Chinese README: An alternate-language version of the extension’s original documentation.
The repository’s MIT license applies to its source code unless otherwise noted. The author advises checking the upstream model’s terms before redistributing its bundled weights.
