RC Gaussian Blur
A blur node that respects your alpha (and your cutouts)
- image
- IMAGE
Gaussian blur sounds like the most boring node you could possibly install, and mostly it is - until you need to blur an image that has transparency and the blur smears the alpha into a gray halo around your cutout. That's exactly the failure this node is built around. RC_GaussianBlur is a professional-grade blur with a preserve_alpha switch that keeps the transparency edge sharp while the color blurs underneath, which matters far more than it sounds like.
It's part of the RC Image Compositor pack, the 36-node Photoshop-in-ComfyUI suite from kj863257. This is the blur you wire in when you're making a soft drop shadow (blur the alpha, keep the color solid), defocusing a background for a fake depth-of-field, or pre-blurring before an upscale pass.
How it works
The node takes an image plus three controls, and honestly the algorithm choice is the interesting part:
- radius (0–100 px) - how much blur. The kernel scales with the radius (about 6× radius, forced odd), so bigger radius = wider spread.
- method -
OpenCVorPIL. OpenCV is the default and the README frames it as the "precise, professional grade" path; PIL is the faster general-purpose one. On a normal still they look close; on huge batches OpenCV's C-backed kernel tends to hold up better. - preserve_alpha (default on) - blurs the RGB channels but leaves the alpha untouched. Turn it off and the alpha blurs too, which is what you want for a shadow layer, not for a cutout.
Output is a single IMAGE, same size, per-batch-frame - so a video sequence gets blurred identically across every frame. There's no radius auto-fit or "keep edges" trickery here; it's a straight Gaussian, which is refreshingly honest.
Installing it
Same drill as every node in this pack. Via ComfyUI Manager, search "RC Image Compositor" and install, then restart. Or manually:
cd ComfyUI/custom_nodes
git clone https://github.com/kj863257/ComfyUI_RC_Image_Compositor
cd ComfyUI_RC_Image_Compositor
pip install -r requirements.txt
Dependencies are just pillow, numpy, and opencv-python - nothing to download, no models. If you have any other image node working, this one works.
Where people get burned
preserve_alphaoff by default would have been a trap, and it isn't. The default is on, which is the right call. The person who turns it off to blur a shadow layer and then forgets to turn it back on is the person who wonders why their next composite has gray edges.- Radius is in pixels, not "amount." A radius of 100 on a 4K image is a heavy, slow blur. If it feels like the node is hanging, it's probably not - it's genuinely convolving a big kernel. Drop the radius.
- This blurs, it doesn't denoise. If you're blurring to hide noise, you're trading one artifact for another; you'd be better off in an upscaler or a real denoise pass.
It's a small node and it knows it. Nothing flashy - but preserve_alpha is the kind of detail that tells you this pack actually composites cutouts all day, and once you've had a halo ruined by a dumb blur, you'll reach for this one on purpose.
Inputs (4)
| Name | Type | Default | Description |
|---|---|---|---|
| image | IMAGE | — | |
| radius | FLOAT | 2.00–100 | Blur radius in pixels |
| method | COMBO | OpenCV | Blur algorithm: - PIL: Fast, general purpose - OpenCV: Precise, professional grade |
| preserve_alpha | BOOLEAN | true | Preserve alpha channel without blurring |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| IMAGE | IMAGE | — |