CV Image Moments

Here's a distinction that confuses people for years: a contour node and a moments node can both tell you a region's orientation, but they're not measuring the same thing. Contour moments integrate a uniform density inside an outline - the outline is all that matters. cv2.moments on a raster defaults to binaryImage=False, where a pixel's mass is its value. That's Hu's original density formulation, and it's the case the contour nodes in this pack structurally cannot reach: two regions with identical outlines but different internal shading are the same shape to CV Shape Moments and different shapes to CV Image Moments.

If that's your problem, this is the node. If it isn't, use the shape one.

The one input you'll tune

images takes frames (IMAGE/MASK/NPARRAY, one output row per frame) and reduces each to a single channel - luminance for colour, since cv2.moments rejects 3-channel data.

weighting decides the density function, and it's the interesting part:

• intensity, peak-normalized (the default) - divide by the frame's own maximum before integrating. Hu moments are not invariant to a brightness gain, and the tooltips report the measured cost of ignoring that: a ×0.5 exposure change moved the I1 error from 0.0394 raw to 0.0146 normalized. That's a real improvement, and honestly reported as imperfect - dimming a soft-edged region also shrinks its effective support, and no normalisation fixes that.
• intensity - the raw formulation, mass equals pixel value.
• binary - ignores values and uses the silhouette, the closest thing to what the contour nodes measure. But every non-zero pixel counts as shape, so any background haze swallows the frame. On a mask this is fine; on a photo it usually isn't.

You get ten outputs, deliberately mirroring CV Shape Moments one-for-one so the two nodes are swappable downstream: centroids (N,1,2 - plugs into CV Draw Points and Draw Labels), mass (the m00 moment: sum of pixel values, not an area - unless you chose binary, where it is the non-zero pixel count), angles (major-axis orientation in degrees, [-90, 90), Y pointing down), eccentricities (0 = radially symmetric, →1 = a line), axes (major-axis segments, 2 standard deviations per side, ready for CV Draw Segments), hu_moments (7 log-scaled invariants), count, plus orientations, chirality and pose_confidence.

Those last three are the advanced ones and the tooltips are refreshingly blunt. orientations is a canonical 0–360 direction where the third-order moments tell head from tail - and it's meaningless when pose_confidence is near 0. chirality is the signed seventh Hu invariant, the one whose sign flips under reflection and which everyone else discards as too noisy - and on intensity fields its magnitudes are far smaller than on a hard silhouette (~1e-5 versus 0.02–0.05 for a glyph), so any downstream matcher's chirality gate needs loosening on this path.

Install

ComfyUI Manager, search comfyui_cv (repo bmad4ever/comfyui_cv). Or:

cd ComfyUI/custom_nodes
git clone https://github.com/bmad4ever/comfyui_cv

Restart. Requirements: Python ≥ 3.12 plus a ComfyUI recent enough for the V3 node API - every node here is an io.ComfyNode with an io.Schema, and there's no NODE_CLASS_MAPPINGS to fall back on, so older builds just don't load the pack. Dependency:

pip install "opencv-contrib-python-headless~=5.0.0.93"

Pinned because the pack is curated against it, and contrib because the contrib submodules have to be present. All four OpenCV wheels share one site-packages/cv2, so a non-contrib install over a contrib one silently strips them and contrib nodes disappear from the menu with no error. tools/repair_opencv_contrib.py --check / --apply is included for that.

Practical notes

Feed it crops, not whole frames, if you're measuring regions of a busy scene - CV Crop by Masks and CV Crop by BBoxes exist for exactly that. A non-zero background is fatal for the intensity modes, and zero frames is a valid result rather than an error, so don't build error handling around "must return data."

And the pack-wide caveat, which the README states itself: heavy LLM assistance in development, test-driven overfitting risk in some pipelines, updates not planned, and not production-ready without your own review. The moments maths is textbook OpenCV; the failure modes above are the ones the node's own tooltips document, which is more guidance than most packs give you.