OpenCV integral3_0

OpenCV integral3_0 is the deluxe version of the integral-image family. Where integral_0 gives you one summed-area table and integral2_0 gives you two, cv2.integral3 gives you three in a single pass: the plain sum table, the squared-sum table, and a tilted table computed along a 45°-rotated grid. The tilted table is the interesting one - it's what lets you sum over rotated (diamond-shaped) regions in O(1), which is exactly what old-school Haar-feature detectors (the ancestors of face-detection cascades) need when features aren't axis-aligned.

In a ComfyUI workflow, be honest with yourself about whether you'll ever use this. It's the most niche of the three integral nodes. The practical consumer-friendly results - adaptive thresholding, box blur, local contrast - all come from adaptiveThreshold_0, blur_0, or threshold_0, which ship in the same pack. integral3 is for when you're building a custom detector or analysis pipeline that genuinely needs rotated-region statistics and you want it computed at C++ speed rather than in a Python scripting loop.

How it works

cv2.integral3(src, sum, sqsum, tilted, sdepth, sqdepth) walks the source once and fills three (H+1, W+1) arrays: sum (standard top-left rectangle sums), sqsum (sums of squares, for variance), and tilted (sums over 45°-rotated rectangles, enabling diamond-region queries). Two of the three outputs are the same data as integral2_0; the tilted table is the addition. As with integral2, the squared sums grow fast, so sqdepth wants to be float (5 = CV_32F or 6 = CV_64F) rather than integer.

Inputs and outputs

• src - NPARRAY, your image via Image2Nparray.
• sdepth - INT, depth of the sum table (4 = CV_32S for 8-bit sources).
• sqdepth - INT, depth of the squared-sum table - keep it float.
• sum, sqsum, tilted - optional out-parameters; the node returns all three anyway.
• Outputs nparray_0 (sum), nparray_1 (squared sum), nparray_2 (tilted).

Installing

Same pack, one install for all ~635 nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/geroldmeisinger/opencv-comfyui
cd opencv-comfyui
pip install -r requirements.txt

or ComfyUI Manager → "opencv-comfyui". No models to download; deps are opencv-contrib-python, numpy, torch.

Gotchas

The tables aren't viewable images - sums exceed 255 and Nparrays2Image won't rescale them, so previewing gives clipped garbage; that's expected, not a bug. Keep batch_size == 1. Don't leave sqdepth integer. And integral3_0 vs integral3_1 are identical overload duplicates (MatLike vs UMat in the stubs) - there's no feature difference, so pick whichever and move on.