Blur Gaussian 🐦‍🔥

Gaussian blur is the default answer to "this image is too noisy to detect anything." Where the pack's plain Blur node averages flatly, Gaussian weights each neighbor by distance from the center, so it smooths without that harsh "smeared window" look. It's the blur you run before Canny - the node's own description for Canny says it explicitly: usually needs Gaussian blur first.

The inputs

• nparrays: an NPARRAY from To Nparray.
• ksize_x (default 3) and ksize_y (default 1): kernel size, both must be odd. Same convention as the pack's plain Blur - ksize_y = 1 means "use the same as width," so the default is a symmetric 3×3.
• sigma_x and sigma_y (defaults 0): the standard deviations. The 0 default is a feature, not a bug - leave both at 0 and OpenCV computes sensible sigmas from the kernel size. Set sigma_x and leave sigma_y at 0 and it mirrors. Only set both when you want anisotropic softening.
• borderType: edge pixel extrapolation (DEFAULT / CONSTANT / REPLICATE / WRAP).

How to actually use it

The pipeline that never goes out of style: load an architectural or mechanical photo → To Nparray → Gaussian Blur → Canny → Hough Lines. The blur's job is to kill the sensor or compression noise so Canny's hysteresis doesn't invent edges out of random brightness jumps. Start at a 3×3 or 5×5 kernel and raise it only if edges look spaghetti-fied - a big Gaussian will start erasing the thin lines you actually care about.

It's also the honest softening filter for "I want the image to look less harshly sharp" without the boxy artifacts of a plain blur.

Output

dst, the blurred NPARRAY - same shape, ready for the next OpenCV node or back through To Image for a preview.

Install

ComfyUI Manager → search ComfyUI-ArchiGraph, or:

cd ComfyUI/custom_nodes
git clone https://github.com/vincentfs/ComfyUI-ArchiGraph

Restart and run the pack's install script once. OpenCV is the only real dependency; no model downloads.

Gotchas

• ksize_y = 1 is shorthand for "same as width," not a 1-pixel kernel.
• Sigma 0 means "auto from kernel size." If you raise sigma manually without touching ksize, you can blur more than the kernel width suggests - which sometimes looks great and sometimes just wipes detail. Tune both together.
• CPU-only, like the rest of the pack's OpenCV set.

If there's one blur in this pack to reach for by default, this is it - it's the version that makes the next detection node look smart.