ImageFilterGaussianBlurAdvanced

The regular ImageFilterGaussianBlur in the Allor pack derives its falloff from the kernel size automatically, which is fine for 99% of blurs. ImageFilterGaussianBlurAdvanced is for the 1% where you want to separate the two: you get to set the kernel size_x/size_y and the sigma sigma_x/sigma_y independently. Same blur, more control, and it's the version you want when a blur needs to look a specific way rather than just "soft."

How it works

It's OpenCV's GaussianBlur with explicit sigma values. Sigma is the standard deviation of the Gaussian - it controls how wide the falloff spreads within the kernel. Set sigma_x/sigma_y to 0 and OpenCV computes them from the kernel size automatically, which makes this node behave exactly like the non-advanced version. Set them to nonzero values and you can have a large kernel that blurs gently (small sigma) or a small kernel doing heavy work (large sigma) - the pairing that looks right for a particular image.

The axis independence carries through everything: size_x/size_y step by 2 (the node normalizes to odd values internally), and you can set different sigmas per axis for elliptical, directional falloff.

Inputs that matter

• images - the IMAGE or batch.
• size_x / size_y - kernel dimensions.
• sigma_x / sigma_y - the falloff spread. 0 = auto (computed from size). This is the whole reason the node exists.

Output is a single IMAGE.

When you'd actually want this

• You're matching a blur to a reference image and the size looks right but the falloff doesn't.
• You need a gentle, wide smoothing where a large-kernel auto-sigma would be too aggressive.
• You want different falloff on each axis without changing the kernel.

If you never find yourself wanting these, use the plain ImageFilterGaussianBlur - this node adds exactly two inputs of control and no magic.

Installing it

Part of the Allor pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Nourepide/ComfyUI-Allor

Restart ComfyUI, or use ComfyUI Manager → Allor Plugin. First launch installs rembg + onnx and creates config.json; no model downloads.

Gotchas

• Sigma interacts with size in ways that confuse people: a sigma much larger than the kernel just gets clipped by the kernel's edge. If your blur doesn't change when you raise sigma, raise size too.
• It's still CPU-bound OpenCV per image; big kernels on big batches cost time.
• Pack-wide update gotcha: Allor's repo was rebased to shrink history, so an old clone can fail on git pull; re-clone if auto-update breaks.