Audio Invert Phase

What it does

Multiplies the waveform by -1. That's the entire node. Input a (a ComfyUI AUDIO), output -a. Same sample rate, same length, same metadata - every sample's sign flipped.

You cannot hear polarity. Play an inverted file on its own and it is indistinguishable from the original. The node only does anything interesting when its output gets summed with something else, and then it does three useful things.

The three real uses

Fix a stem with flipped polarity. Occasionally a separation or a conversion comes back inverted relative to the original mix. Lay the two on top of each other and instead of reinforcing, they partially cancel - the low end thins out and everything sounds hollow. Inverting one side fixes it. This is the difference between a mix that sounds wrong and one that doesn't, and it's almost impossible to diagnose by ear alone.

Null-test a separation. Send your original into Audio Ensemble alongside an inverted copy of a separated stem and set weights to 1 and -0.999 - avg_wave with a negative weight gives you the difference signal. Near silence means the stem and the source are identical; whatever's left is exactly what the model changed. It's the fastest way to see what a model actually did rather than trusting the preview. (Exact 1 and -1 gets rejected - a zero weight sum raises an error by design.)

The karaoke trick. Core ComfyUI ships Split Audio Channels and Join Audio Channels. Load a stereo track, split it, invert one channel, rejoin: you now have L and −R. Summed to mono, anything panned dead-centre - which is where lead vocals usually sit - cancels out. It's the oldest vocal-removal hack there is, and it's genuinely useful for a quick check. It's also destructive: the bass, the kick and half the stereo image go with the vocal, which is exactly why people moved on to the models in this pack. Use it as a diagnostic, not a deliverable.

Inputs and outputs

• Input a - the AUDIO stream. Required; an empty one gives you audio input a is required.
• Output -a - the inverted stream.

That's the full schema. No widgets, no settings, nothing to tune.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/pymss-project/comfy-mss
python -m pip install pymss   # the Python env ComfyUI runs in

Or ComfyUI Manager → Comfy-MSS, then restart. It's part of the same pack as the separators, so install it once and you get the whole family.

Gotchas

• Nothing will sound different. If you wired it in and A/B'd the two outputs, they're identical until they're summed. That's physics, not a broken node.
• The null test needs matching lengths and rates. Audio Ensemble truncates everything to the shortest input and resamples to the first input's rate, so slight mismatches show up as a residual tail rather than being errors.
• It flips the whole waveform, not one channel. For the stereo trick you need Split Audio Channels first - this node has no channel selector.
• v1.0.x workflows. Node types were renamed in v1.1.0, so older saved graphs need rebuilding.