allpass_decorrelator

NODE

Spreads the image by phase alone. The magnitude response stays flat, so the tone does not change.

Parameters

FIELD	WHAT IT DECIDES	RANGE	REQUIRED
depth	Diffusion depth (0.0 = bypassed, 1.0 = maximum diffusion).	0 〜 1	yes
low_cutoff_hz	Lower corner frequency for allpass diffusion [Hz]. Frequencies below this remain phase-aligned.	20 〜 5,000	no

{
  "depth": 0.5,
  "low_cutoff_hz": 500,
  "type": "allpass_decorrelator"
}

How it is computed

three cascaded allpass sections per channel

depth detunes the centre frequencies of L and R in opposite directions
depth = 0   L and R identical: no spreading at all
low_cutoff_hz sets where the spreading starts

|H(f)| = 1 at every frequency, in both channels

An allpass filter changes phase without changing magnitude. Because nothing is delayed and re-added, there is no comb filtering, which is what separates this from the usual widening tricks.

Using it

It lowers correlation, it does not add material.
Nothing new is generated. What changes is how alike the two channels are, which is what the ear reads as width and depth.
Check it in mono if that matters to you.
Phase differences are exactly what does not survive summing to mono. The node is transparent per channel, but a summed version will not sound the same.
Start low.
Because the tone does not change, it is easy to go too far without noticing: the image keeps opening while the recording loses its centre.