ComfyUI-Demucs-AudioSeparator Introduction
ComfyUI-Demucs-AudioSeparator is a custom ComfyUI node that separates a music recording into individual stems: audio tracks for parts such as vocals, drums, bass, and other instruments. It uses Meta’s Demucs audio-separation technology.
If you are making a video, animation, or other AI-assisted artwork, stems give you more control over the soundtrack. You could use the vocal track on its own, adjust the drums separately, or work with the remaining instruments without editing every sound in the original recording at once. The results are estimates, however: a separated stem may still contain traces of other sounds, especially when instruments overlap.
How ComfyUI-Demucs-AudioSeparator Works
Think of a finished song as a painting made from several layers that have been flattened into one image. Demucs tries to identify those layers in the audio and produce a separate track for each one.
You provide an audio clip to the node and choose a model. The model analyzes the recording and predicts which sounds belong to each stem. A four-stem model separates vocals, drums, bass, and other sounds. A compatible six-stem model also provides separate guitar and piano stems. “Other” means sounds that do not fit the model’s separately named categories; it is not necessarily a clean track of one instrument.
Separation involves a trade-off between processing time, memory use, and the quality you hear. For example, asking the model to make more predictions with shifts may improve a result, but takes longer. Processing an audio clip in segments with split can reduce memory pressure, which is particularly useful for long recordings.
ComfyUI-Demucs-AudioSeparator Features
- Up to six stems: Choose a compatible model to separate vocals, drums, bass, other sounds, guitar, and piano. Four-stem models do not produce separate guitar and piano tracks.
- Model choice: Try different models when a particular stem matters to your project. The available models are described below.
- GPU or CPU processing: Set
devicetocudato use a compatible CUDA GPU, orcputo process without one. A GPU can make processing faster. - Adjustable precision: The
precisionchoices arefloat32,bfloat16, andfloat16. The original documentation recommendsbfloat16for Ampere-generation GPUs, citing the RTX 3090 and 4090 as examples. If a precision setting causes an error on your system, tryfloat32. - Quality-versus-speed control:
shiftssets the number of predictions made during separation. The author recommends 1–5. Start at1for a quicker test, then try a higher value if the result needs improvement. - Segment overlap:
overlapcontrols how much neighboring processing windows share. Its documented default is 0.25. You can leave the default in place for an initial run and adjust it when experimenting with how segment boundaries sound. - Lower-memory processing: Enable
splitto process audio in segments. The author particularly recommends it for long clips or GPUs with less than 8 GB of VRAM. - Stem output controls: Use
stemsto enable or disable the instrument-specific outputs you need in your workflow. - Automatic model downloads: The extension downloads models to
ComfyUI/models/demucswhen needed. Allow time for a model to become available the first time you use it.
The node’s audio input is the clip you want to separate. A practical first test is to use a short excerpt, keep overlap at its default, set shifts to 1, and listen to the resulting stems before processing a longer recording.
ComfyUI-Demucs-AudioSeparator Models
| Model | What it is for | When to try it |
|---|---|---|
htdemucs | The standard four-stem Hybrid Transformer Demucs model. | Start here when you need vocals, drums, bass, and other sounds. |
htdemucs_ft | A fine-tuned version of the standard model. | Compare it with htdemucs when a stem is important and you want to hear which result works better for your recording. |
htdemucs_6s | A six-stem model that includes separate guitar and piano outputs. | Choose it when you specifically need to work with guitar or piano apart from the other instruments. |
hdemucs_mmi | An improved HDemucs-based model listed by the extension. | Try it as an alternative if the standard model does not give you the separation you need. |
mdx | A model based on the Music Demixing Challenge. | Compare its output with another model for your particular clip. |
mdx_extra | Another Music Demixing Challenge-based option. | Try it when comparing available separation results. |
Different recordings can produce different results with the same model. Listen to the stem you intend to use, rather than assuming that one model will be best for every song.
Troubleshooting ComfyUI-Demucs-AudioSeparator
The node runs out of GPU memory.
- Enable
split, especially for a long clip or a GPU with less than 8 GB of VRAM. - Test with a shorter excerpt to confirm that the workflow runs.
- If needed, set
devicetocpu. Processing may take longer, but it does not depend on CUDA GPU memory.
Processing is taking longer than expected.
- Check
shifts. Start with1before trying a higher value. - If you have a compatible CUDA GPU, check whether
deviceis set tocuda. - For an initial comparison, process a short excerpt before running the full recording.
There are no separate guitar or piano stems.
- Check the selected model. The documented six-stem option is
htdemucs_6s. - Check that the relevant outputs are enabled under
stems. - Listen to the results: the model predicts which sounds belong in each category, so separation may not be perfect.
A precision setting produces an error.
- Change
precisiontofloat32. - Run a short test clip.
- Once that works, try another precision setting only if you need to compare performance on your hardware.
Frequently asked questions
Will the vocal stem be completely free of instruments? Not necessarily. Demucs estimates the parts of a mixed recording, so some sound can remain in the wrong stem.
Do I need a GPU? The node offers both cuda and cpu device settings. A compatible GPU is recommended for performance, but you can select CPU processing.
Which model should I choose first? Start with htdemucs for the usual four stems. Choose htdemucs_6s when separate guitar and piano outputs are essential.
Do I need to download each model myself? According to the original documentation, models are downloaded automatically to ComfyUI/models/demucs.
Learn More about ComfyUI-Demucs-AudioSeparator
The node’s settings are best learned by comparing short clips of music you already know. Start with htdemucs, listen closely to the stem you need, and then change one setting or model at a time. This makes it easier to tell whether a change actually helps your project.
For background on the underlying audio-separation technology, look for Meta Demucs documentation. For extension-specific behavior, use the extension’s own repository documentation and issue tracker when you have its verified repository address; the repository URL shown in the original readme is a placeholder, so it is not reproduced here.
