OpenCV getGaussianKernel_0

OpenCV getGaussianKernel_0 generates the 1D Gaussian convolution kernel OpenCV uses for blurring - a column vector of weights shaped like a bell curve. It's the ingredient behind a Gaussian blur, not the blur itself. The pack's GaussianBlur node does the whole job for you, so this one matters when you want the bare kernel: to apply it yourself via sepFilter2D, to combine it with another kernel, or just to see exactly what coefficients a given sigma produces.

How it works

A 2D Gaussian blur is separable - it's the outer product of a 1D kernel with itself. OpenCV's getGaussianKernel(ksize, sigma) computes that 1D kernel directly from the math: weights proportional to exp(-x²/(2σ²)), normalized so they sum to 1. If you pass sigma=0, OpenCV derives a sensible sigma from the kernel size for you - that's what the convenience blur nodes effectively do, and it's the right move unless you have a specific target.

The inputs

• ksize (INT) - kernel length, in pixels. Must be positive and odd (1, 3, 5, 7, …); an even value trips OpenCV's assertion.
• sigma (FLOAT) - the standard deviation in pixels. 0 = let OpenCV compute it. Note the asymmetry: for a fixed ksize, a large sigma just trims the tails, so you don't get a wider blur for free.
• ktype (INT) - output dtype of the kernel: -1 for same-as-source, 5 for CV_32F, 6 for CV_64F. Float is the sane choice; the kernel coefficients are fractional by nature.

Output is a single nparray - a 1D (ksize, 1) float vector, not an image. In NPARRAY form it's OpenCV numpy, so it can't be previewed directly; you either consume it with a convolution node or accept that it's a number you're using, not looking at.

Why you'd reach for it instead of GaussianBlur

Two honest reasons. First, custom application: sepFilter2D with this kernel gives you a Gaussian blur you control as two separable passes, which is the standard setup for performance-sensitive pipelines or for asymmetric smoothing (blur more in one direction than the other by using different kernels per axis - a real technique for motion-ish softening). Second, education and verification: if you're hand-building a filter pipeline or debugging why a blur looks different from the stock one, this node shows you the coefficients instead of hiding them.

If all you want is "soften this image," stop reading and use GaussianBlur - that's the one-step answer, and the KB's post-processing playbook is explicit that a Gaussian smears everything including edges, so it's the softening tool, not the edge-preserving one (that'd be bilateral). getGaussianKernel is for the days you need the parts.

Install and gotchas

ComfyUI Manager → search opencv-comfyui (display "OpenCV"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/geroldmeisinger/opencv-comfyui
pip install opencv-contrib-python

Restart, done, no models. Pack rules: outputs are NPARRAY not IMAGE, batch size stays 1, and this particular node's inputs are all plain numbers - no literal-string traps, which makes it one of the friendlier members of the family. Even ksize here is a plain INT, not a bracketed Size literal. The only real footgun is even ksize and forgetting that the output is a 1D vector, not a usable image.