The reason for this, is for models like U2Net where the input image size is 320x320, meaning that the resulting output needs to be resized, and can have sharp edges; blurring before resizing can smooth this out.

This is the code that I used to generate the images.

constkernels=[5,9,17];constsigmas=[2,3,5];constchunk=4;for(leti=0;i<kernels.length;i++){constkernel=kernels[i];constsigma=sigmas[i];image.gaussianBlur(kernel,sigma,chunk)then(image=>image.save(`blurred-(${kernel}-${sigma}).png`));}

Sigma should roughly be a third of kernel size (from what I found researching).
Kernel must be odd, since we want an even square around the pixel.
For example, if we had a 3x3 kernel grid, then when we place this on our pixel there is an even space all around that pixel (in this case 1 pixel).

Here is a demo of what this implementation currently looks like.

Image	Kernel	Sigma	1 chunk	4 chunks	12 chunks
	N/A	N/A	N/A	N/A	N/A
	5	2	9 seconds	3 seconds	3 seconds
	9	3	31 seconds	4 seconds	5 seconds
	17	5	1 minute, 42 seconds	8 seconds	7 seconds

EDIT: This has since been updated so that a number of chunks are not passed, and instead we will use thenavigator.hardwareConcurrency property to find the ideal number of chunks.
As you can see above, there was no time save between 4 and 12 because my PC had 4 cores.