UVR MDX Inst HQ5 Separator:
UVRMDXInstHQ5 separates a recording into an instrumental track and a vocal track using a bundled MDX audio model. It is useful when you need music without the prominent vocals, or a vocal track for further editing, remixing, or audio-driven creative work. The node processes the recording in manageable sections when needed, then reconstructs the tracks at the recording’s original sample rate. Separation is an estimate rather than a perfect extraction: instruments may remain audible in the vocal track, and some vocal sound may remain in the instrumental track.
UVR MDX Inst HQ5 Separator Input Parameters:
audio
The required AUDIO recording to separate. Connect a ComfyUI audio output containing waveform data and a sample rate; there is no default, minimum, or maximum stated for this input. The node processes one batch item at a time. If the selected item is mono, it duplicates that channel to make stereo audio; if it has more than two channels, it uses only the first two. An empty recording cannot be processed.
device
Chooses where the model runs. The available options are auto, cuda, directml, and cpu; the default is the configured DEFAULT_DEVICE, whose actual value is not shown in the supplied source. Use auto to let the node’s session setup choose a provider, or select a specific option when you know it is available in your environment. The selected provider can affect whether processing starts and how quickly it runs. The returned info text reports the provider actually used.
segment_seconds
Controls the approximate amount of audio processed per main segment. The allowed range is 0.0 to 120.0 seconds, in 0.5-second steps; the default is DEFAULT_SEGMENT_SECONDS, whose numeric value is not shown in the supplied source. Shorter segments can make long recordings easier to process in pieces, while longer segments reduce the number of segment boundaries. A value of 0.0 processes the entire recording as one segment; the same happens when the recording is shorter than the requested segment length. The source does not establish a guaranteed memory or speed improvement for any particular value.
margin_seconds
Sets the amount of neighboring audio included around segment boundaries during processing. The allowed range is 0.0 to 10.0 seconds, in 0.1-second steps; the default is DEFAULT_MARGIN_SECONDS, whose numeric value is not shown in the supplied source. This overlap gives the model context near a cut between segments; the overlapping portions are trimmed when the results are joined. A larger margin means more overlapping audio is processed. If the margin exceeds the segment length, the node limits it to that segment length.
denoise
Controls an additional model-processing pass. The options are True and False, with True as the default. When enabled, the node runs the model on both the audio representation and its negation, then combines the predictions; when disabled, it makes one model run instead. Enabling it may change the separation result and requires the additional run, but the supplied code does not guarantee that it will remove every noise or artifact.
compensation
Scales the model’s estimated instrumental track before the vocal track is calculated. The allowed range is 0.0 to 4.0, in 0.001 steps; the default is MODEL_COMPENSATION, whose numeric value is not shown in the supplied source. Increasing a positive value makes the estimated instrumental louder before it is subtracted from the original mix, so it changes both returned tracks. At 0.0, no positive compensation is applied, but the instrumental estimate is still subtracted. Treat this as a separation-balance adjustment, not as an independent volume control for the finished vocal track.
batch_index
Selects which item to process when audio contains a batch. The allowed setting is an integer from 0 to 4096, in steps of 1, and the default is 0. Index 0 selects the first item, index 1 the second, and so on. The selected index must also be smaller than the batch’s actual size; increasing the setting does not make the node process multiple items at once.
UVR MDX Inst HQ5 Separator Output Parameters:
instrumental_audio
An AUDIO output containing the model’s estimated instrumental track after any positive compensation is applied. You can connect it to playback, export, or further audio-processing nodes when you want a backing track. It is returned as stereo audio at the selected source recording’s sample rate and, if resampling was needed internally, with its original sample count restored. Because this is a model estimate, some vocal sound may remain.
vocals_audio
An AUDIO output calculated by subtracting the compensated instrumental estimate from the original mix used for separation. You can use it when you want to work with the voice separately. It is returned as stereo audio at the selected source recording’s sample rate and, if resampling was needed internally, with its original sample count restored. Its content depends directly on the instrumental estimate and compensation setting, so it may include instrumental bleed or separation artifacts.
info
A STRING status summary in the form model=… | provider=… | sample_rate=… | denoise=… | compensation=…. It identifies the bundled model file’s name, the provider used for inference, the source audio’s sample rate, whether denoise was enabled, and the compensation value displayed to three decimal places. Use it to check which configuration produced a result or to compare runs; it is not a quality score or a report of separation accuracy.
UVR MDX Inst HQ5 Separator Usage Tips:
- Start with the node’s defaults and listen to both tracks before changing settings. The supplied source names configuration constants for several defaults but does not reveal their numeric values, so check the values shown in your installed node.
- For a batched audio input, set
batch_indexto the item you want and keep it below the number of items in the batch. Run the node separately for other items. - For a long recording, try a finite
segment_secondsvalue instead of 0.0 so the node divides it into sections. Keep a nonzeromargin_secondsif you hear changes around the joins, then compare the result by listening; a larger margin also processes more overlapping audio. - Compare
denoise=Trueanddenoise=Falseon the same excerpt when judging artifacts. The former performs two model runs per inference call, while the latter performs one. - Change
compensationin small steps and check both outputs after every adjustment. It changes the instrumental estimate and the vocal remainder together, so improving one track may make the other less useful. - Check
infowhen comparing runs, especially after changingdevice. It records the provider actually used and the settings most directly relevant to the returned tracks.
UVR MDX Inst HQ5 Separator Common Errors and Solutions:
batch_index `{batch_index}` is out of range for batch size `{batch_size}`.
- Explanation: The requested batch item does not exist. For example, a batch of one item accepts only index 0.
- Solution: Set
batch_indexto a value from 0 through the batch size minus one, or supply a batch containing the item you intended to select.
Audio input must have at least one channel.
- Explanation: The selected audio item has no audio channels, so the node cannot prepare it for stereo processing.
- Solution: Check the upstream audio node and provide a recording with at least one channel.
Input audio is empty.
- Explanation: The selected item contains no audio samples.
- Solution: Confirm that the upstream node loaded or generated audible-length audio, then retry with a nonempty recording.
Bundled model not found: `{path} followed by Place {MODEL_FILENAME}` in the node's models folder before starting ComfyUI.
- Explanation: The model file expected by this node is absent from its
modelsfolder. - Solution: Obtain the exact model file required by your installation, place it in the node’s
modelsfolder as directed by the error, and restart ComfyUI. Check the displayed path and filename rather than substituting a differently named model.
Expected audio array with shape [C, T].
- Explanation: The separator received audio in a shape other than channels by time. This indicates a preparation or integration problem rather than a normal setting you can fix with a slider.
- Solution: Supply audio through the node’s
audiosocket using a compatible ComfyUI audio-producing node. If the error persists with valid audio, check the custom node installation or report the failing workflow to its maintainer.
Model session or provider error mentioning cuda, directml, cpu, or an execution provider
- Explanation: The chosen processing backend may be unavailable or may have failed while loading or running the model. The exact message depends on components whose implementations are not included in the supplied source.
- Solution: Try
device=autoordevice=cpu. If a specific GPU option is needed, verify that its required runtime and drivers are installed and that ComfyUI can use them; retain the full error text when seeking support.
Audio validation or missing waveform/sample_rate error
- Explanation: The node relies on an audio-input validator whose implementation is not supplied, so its exact error wording cannot be stated here. Such an error indicates that the incoming value is not an acceptable ComfyUI audio object.
- Solution: Connect a valid
AUDIOoutput, confirm that it contains waveform data and a sample rate, and check the upstream node for loading or generation errors.
