🎵 Demucs Audio Separator ⚡:
DemucsAudioSeparator separates a piece of audio into musical parts using a pretrained Demucs model. This can help you isolate a voice, work with individual instruments, or make a version without the vocal for editing and creative projects. The node prepares the audio for the selected model, runs the separation, and can adjust the level of the resulting audio. Separation is an estimate: how cleanly a sound is isolated depends on the recording and the model you choose.
🎵 Demucs Audio Separator ⚡ Input Parameters:
audio
The audio you want to separate, supplied as an AUDIO value containing a waveform and sample rate. The node accepts audio with one channel by duplicating it into stereo; if it has more than two channels, it keeps only the first two. It also resamples audio when its sample rate differs from the selected model’s required rate. No default, minimum, or maximum is specified.
model
Chooses the pretrained Demucs model used for separation. Available options are htdemucs, htdemucs_ft, htdemucs_6s, hdemucs_mmi, mdx, mdx_extra, mdx_q, mdx_extra_q, mdxc, and mdxc_fb_ft. The model determines which source types are actually available; selecting an output does not make a model produce a source it lacks. No explicit default is specified.
device
Chooses where separation runs: cuda for a compatible GPU or cpu for the processor. A GPU may be preferable for demanding jobs, but if you select cuda when CUDA is unavailable, the node prints a notice and runs on the CPU instead. No explicit default is specified.
precision
Offers float32 and float16, with float32 as the default. The node uses this setting while preparing the waveform, but converts the waveform to float32 immediately before model inference. Consequently, selecting float16 does not mean the separation itself runs in half precision. The code also handles a legacy bfloat16 value by changing it to float32, although bfloat16 is not an interface option.
shifts
Sets the number of shifts passed to Demucs during separation. The default is 1; the minimum is 1 and the maximum is 10. Higher values ask the model to do more work and can take longer, so start with 1 and increase it when you have time to compare results on a difficult recording.
overlap
Sets the amount of overlap used when Demucs processes split audio. The default is 0.25; the minimum is 0.1, the maximum is 0.9, and the interface step is 0.05. More overlap increases the amount of neighboring audio considered around segment boundaries, at a potential processing cost. It is most relevant when split is enabled.
split
Controls whether Demucs processes the audio in segments. The default is True. Keeping it enabled is a practical starting point for longer recordings; turning it off asks the model to process without that segmentation and may require more memory. The interface does not specify numeric limits.
normalize
Controls peak normalization of each enabled, available source after separation. The default is True. When enabled, the node divides a non-silent stem by its highest absolute sample value, bringing that peak to a magnitude of 1 before applying gain_multiplier. This changes the level of each stem independently, so the processed stems should not be assumed to preserve their original relative volumes.
gain_multiplier
Scales the waveform of each enabled, available stem after optional normalization; it also scales the enabled instrumental output. The default is 1.0, the minimum is 0.0, the maximum is 10.0, and the interface step is 0.05. 1.0 leaves the level unchanged at this stage, 0.0 makes processed audio silent, and values above 1.0 increase sample levels and may produce clipping in later playback or export.
vocals
Controls whether the separated vocal source is returned as audible audio. The default is True. If disabled, or if the chosen model does not provide a source named vocals, the vocals output contains silence of the corresponding shape and duration. Disabling it does not stop the model from performing separation.
drums
Controls whether the model’s drums source is returned as audible audio. The default is True. If disabled or unavailable from the model, the drums output is silent; the other output sockets remain available.
bass
Controls whether the model’s bass source is returned as audible audio. The default is True. If disabled or unavailable from the model, the bass output is silent. This setting selects what you receive, rather than changing which sources the model calculates.
other
Controls whether the model’s other source is returned as audible audio. The default is True. This source represents audio the model assigns to its other category, rather than a guaranteed single instrument. If disabled or unavailable, the output is silent.
guitar
Controls whether a source named guitar is returned as audible audio. The default is False. Turn it on when you want that stem, but check that your selected model actually lists guitar among its available sources; otherwise, the output remains silent.
piano
Controls whether a source named piano is returned as audible audio. The default is False. Enabling the switch requests that output, but it is silent when the chosen model does not provide a source named piano.
instrumental
Controls whether the node produces its vocals-removed audio. The default is True. When enabled and the model provides vocals, the node subtracts the vocal waveform from the prepared input waveform; if no vocal source exists, it returns the prepared input instead. When disabled, its output is silent. Because the node can process the vocal result before this subtraction, normalization or gain settings may also affect the instrumental result when vocals is enabled.
🎵 Demucs Audio Separator ⚡ Output Parameters:
vocals
An AUDIO value containing the model’s vocal estimate when vocals is enabled and that source exists. Use it when you need to edit or process a voice separately. Its waveform is returned on the CPU at the model’s sample rate, with optional normalization and gain applied. Otherwise, it contains silence rather than an absent output.
drums
An AUDIO value containing the model’s drum estimate when drums is enabled and available. It lets you work on percussion separately. The output uses the model’s sample rate and contains a zero-valued waveform if the switch is off or the source is unavailable.
bass
An AUDIO value containing the model’s bass estimate when bass is enabled and available. It is useful for processing or listening to the low-end part independently. If disabled or unavailable, the waveform is silent.
other
An AUDIO value containing the model’s other estimate when enabled and available. It can contain multiple sounds grouped by the model rather than one guaranteed instrument. If disabled or unavailable, the waveform is silent.
guitar
An AUDIO value containing a guitar estimate only if guitar is enabled and the chosen model supplies a guitar source. Its default switch setting is off, so a silent result is expected until you enable it; it also remains silent when the source is unavailable.
piano
An AUDIO value containing a piano estimate only if piano is enabled and the chosen model supplies a piano source. A silent result can mean the switch is off, as it is by default, or that the selected model has no piano source.
instrumental
An AUDIO value intended to represent the prepared input with the estimated vocals removed. When enabled, the node subtracts the vocal source if one exists, or returns the prepared input if it does not; when disabled, it returns silence. It applies optional normalization and gain. This is a subtraction-based result, not a guaranteed perfect sum of the other output stems, and its sound may depend on whether the vocals output was processed first.
metadata
A JSON value describing the completed run. It includes the selected model, actual processing device, reported precision, shifts, overlap, split, the model’s available_sources, input and output sample rates, and status set to success. Check available_sources when an enabled stem is silent: it tells you which source names the loaded model actually provides. The metadata records the precision setting, but inference is still fed a float32 waveform.
🎵 Demucs Audio Separator ⚡ Usage Tips:
- Start with
shiftsset to1,overlapat0.25, andsplitenabled. Listen to a short representative passage before increasing processing settings for a full recording. - Check
metadata.available_sourcesbefore relying onguitarorpiano. An enabled switch cannot create a stem that the selected model does not provide. - For a quiet or long recording, keep
splitenabled. If an unsplit run encounters memory trouble, turnsplitback on, reduceshifts, or try a shorter audio excerpt. - Leave
normalizeon when you want each selected result brought to a readily audible peak level. Turn it off when preserving the model’s relative stem levels matters to your mix. - Start with
gain_multiplierat1.0. Increase it cautiously and listen for distortion, particularly whennormalizeis on, because gain is applied after normalization. - For a vocals-removed track, enable
instrumentaland compare it by ear with the individual stems. If you need to evaluate the subtraction against the unmodified vocal estimate, turn offnormalizeand leavegain_multiplierat1.0. - If CUDA is unavailable, choose
cpuknowingly or allow the node’s automatic fallback. Selectfloat32when you want the precision setting to reflect the waveform type used for inference.
🎵 Demucs Audio Separator ⚡ Common Errors and Solutions:
Failed to load Demucs model [model]: [details]
- Explanation: The selected model could not be obtained or initialized. The details appended to this message provide the underlying cause.
- Solution: Check the detailed error, confirm the model selection and that the Demucs installation and model files are accessible, then retry. If the model needs to be retrieved, check your connection and storage space.
⚡ [Demucs Pro] Error during Demucs inference: [details]
- Explanation: The model loaded, but separation failed while it was running. The appended details are needed to distinguish causes such as insufficient memory from an audio or installation problem.
- Solution: Read the details first. For a memory-related failure, enable
split, lowershifts, try a shorter excerpt, or use another available device.
⚡ CUDA not available, falling back to CPU
- Explanation: This is a notice, not a failure. You selected
cuda, but PyTorch could not use CUDA, so the node switched tocpu. - Solution: Continue on the CPU if the speed is acceptable. To use GPU processing, check that your system has a compatible GPU and a working CUDA-enabled PyTorch setup.
⚠️ [Demucs Pro] BFloat16 not supported for FFT. Upgrading to Float32 for inference.
- Explanation: A legacy configuration supplied
bfloat16, which is not one of this node’s listed interface choices. The node changes it tofloat32. - Solution: Select
float32in the current interface. No further action is needed if separation completes successfully.
KeyError: 'waveform' or KeyError: 'sample_rate'
- Explanation: The supplied
audiovalue is missing a field that the node reads when preparing the recording. - Solution: Connect a compatible
AUDIOoutput that supplies both a waveform and its sample rate, and check any upstream node that constructs or modifies the audio.
Audio shape or channel-related error
- Explanation: The node expects a waveform it can treat as channels and samples, optionally with a batch dimension. It handles mono and more-than-stereo audio, but an unexpected waveform layout may still fail.
- Solution: Use an upstream audio node that produces a standard
AUDIOwaveform, and test with an ordinary mono or stereo file.
An enabled stem is silent, with no error message
- Explanation: A source output is filled with zeros when its switch is off or the loaded model does not provide that source. A genuinely quiet source is also possible.
- Solution: Confirm the switch is on, then inspect
metadata.available_sourcesfor the exact source name. Try a model that provides the source you need.
The instrumental output sounds unexpectedly different
- Explanation: The node constructs it by subtracting the vocal estimate from the prepared input. When
vocalsis enabled, the vocal waveform may have been normalized or gain-adjusted before subtraction, which can change the result. - Solution: For a baseline comparison, set
normalizetoFalseandgain_multiplierto1.0. Compare the result with the original audio and the vocal stem.
Playback is distorted after increasing gain_multiplier
- Explanation: The multiplier can push waveform values beyond a typical playback or export range, especially after normalization.
- Solution: Reduce
gain_multipliertoward1.0and check the level in the next audio-processing or export step.
