OpenCV calcCovarMatrix_0

Let's set expectations: OpenCV calcCovarMatrix_0 is a statistics function, not an image operation. It takes a set of samples, computes their covariance matrix - the table of how every dimension varies together with every other - and also returns the mean. It's the mathematical foundation under PCA, whitening, Mahalanobis distance, and a lot of feature-analysis pipelines. If your mental model of ComfyUI is "images in, images out," this node is the one in the pack you'll almost certainly never wire, and knowing that is the useful part of this article.

It exists here for the same reason every node in this pack exists: the author auto-generated a node for all 635 top-level standalone cv2 functions, including the linear-algebra utilities that OpenCV ships for computer-vision researchers, not for diffusion users. The README is upfront that "not every function is useful within ComfyUI without further processing." This is that warning embodied.

Inputs (read these only if you're doing real stats work)

• samples - NPARRAY: your data, with rows/columns meaning samples vs. features depending on the flags you choose (that's what the CV_COVAR_* flag variants are about).
• mean - NPARRAY, and here's a pack quirk worth knowing: mean is an out-parameter in OpenCV (the function writes the mean into it) but the generator surfaced it as a required input. You have to wire something in even though it gets overwritten - classic auto-gen awkwardness. Feed it any same-shape nparray.
• flags - INT: the cv2.COVAR_* mode (e.g. 1 = CV_COVAR_NORMAL, 0 = CV_COVAR_SCRAMBLED).
• ctype - INT: the output array type (e.g. CV_32F = 5, CV_64F = 6).
• covar (optional) - the other out-parameter; leave unwired.

Outputs: nparray_0 (the covariance matrix) and nparray_1 (the mean). Neither is an image - they're square float matrices and a vector - so don't feed them into Nparrays2Image expecting a picture; that's the 'NoneType' object has no attribute 'shape' error waiting to happen.

Install

ComfyUI Manager → opencv-comfyui, or git clone https://github.com/geroldmeisinger/opencv-comfyui into ComfyUI/custom_nodes, restart. Needs opencv-contrib-python.

The real takeaway: if you're doing PCA-style analysis on features inside a ComfyUI graph - a genuinely exotic workflow - this is the primitive. If you're not, bookmark it as "the statistics node," don't lose sleep over the out-parameter weirdness, and move on. It's not broken; it's just not for you.