AQ_ColorMatchImage

Ever generated twelve images and wished they all shared one photo's look? That's this node's entire job: force the color of a whole batch to match a reference image. It's a color-grade transfer by statistics - the classic "make my AI output look like the moodboard" trick, but it works on a full batch at once instead of one image at a time.

How it works

The mechanism is per-channel mean normalization, with a local option. With blur_size set to 0 it computes each image's per-channel mean and scales every channel so the mean lands on the reference's mean. That's global color-cast transfer: warm photos make warm output, teal shadows become teal. With blur_size above 0 it goes local - it blurs both images (gaussian blur, or a guided filter if you pick guidedFilter) and divides/multiplies per-pixel by those blurred values. Local transfer preserves contrast while matching color neighborhood by neighborhood, which handles images that have a color cast in one region but not another. factor then lerps between the original and the result, so 0.5 gives a half-strength match and negative values push away from the reference.

If the reference batch is shorter than your image batch, the first reference frame is repeated to cover it.

Inputs

• images - your batch.
• reference - the look you're after.
• blur_type - blur (gaussian) or guidedFilter. Guided filter is usually better at preserving edges; it needs cv2.ximgproc, see below.
• blur_size - 0 means global; higher means more local. Default 0.
• factor - -10 to 10, default 1.

Output

One IMAGE per input frame, same dimensions.

Install

It's part of AQnodes:

cd ComfyUI/custom_nodes
git clone https://github.com/2frames/ComfyUI-AQnodes
cd ComfyUI-AQnodes
pip install -r requirements.txt

or search "AQnodes" in ComfyUI Manager and restart. The pack imports cv2 at startup, so OpenCV has to be present - normally fine, since insightface pulls in opencv-python as a dependency. But the guided filter specifically imports from cv2.ximgproc, which lives in opencv-contrib-python, not the plain build. If you hit an import error on guidedFilter, pip install opencv-contrib-python fixes it.

Gotchas

Be realistic about what it does: it transfers color statistics, not a "grade." You get the cast, not contrast curves or a learned style. It's a batch-friendly Reinhard-style match, which is the right tool when you need consistency - unifying a set of renders to one reference - and the wrong tool when you want creative color grading. Large blur_size values internally downscale the image to keep it fast, so don't expect perfect fidelity at the extreme end.