cv2.checkRange

Every graphics-adjacent bug you've ever chased ends the same way: a NaN got into the buffer and something downstream turned it into black, white, or a frame of solid noise. cv2.checkRange is OpenCV's answer - walk the array, is every element inside the bounds you gave, and if not, where is the first one that isn't.

It's a raw wrapper in comfyui_cv (~470 cv2.* nodes), and it's one of the few in the catalog whose natural home is a debug branch rather than the main pipeline.

The defaults are already a NaN detector

This is the part worth reading twice. minVal and maxVal default to blank, which the pack maps to OpenCV's own defaults - the extreme ends of the double range. Left alone, the range check is therefore effectively "is this a number at all": a NaN fails every comparison, so it counts as out of range, and so does ±infinity. You don't have to configure anything to catch the classic failure.

Set real bounds and it becomes a validation node instead - "is this mask in 0–1", "is this magnitude map sane", "did the division at the end of my array maths leave anything above 10". The tooltips are precise about the semantics: minVal is an inclusive lower bound, maxVal is exclusive upper.

Inputs and outputs

• a (required) - the array to audit. Accepts NPARRAY, IMAGE or MASK, converting on the way in, and this conversion is a trap worth knowing before you write an assertion: an IMAGE becomes uint8 BGR (values 0–255) and a MASK becomes uint8 single-channel. Checking a converted image against 0–1 will report the whole frame as out of bounds, because nothing is 0–1 any more. Audit NPARRAYs, or set bounds matching the uint8 range.
• quiet (optional, default true) - what happens when something is out of range. true returns false and carries on. false makes OpenCV raise, which inside a ComfyUI run means the node errors out and the queue stops. Counterintuitive naming, useful behaviour: leave it on for a gate, turn it off when you want the run to stop loudly.
• minVal, maxVal (optional, blank by default) - the bounds, as above.

Two outputs:

• ok - BOOLEAN, all-in-range or not. This is what you gate on: a boolean is a switch selector, so it can route a good array down one path and a bad one to a preview or a halt.
• pos - a CV_TUPLE composite point (x, y), the position of the first offending element. Meaningless when ok is true. Wire it into any cv2 point input - a cv2.circle from the same pack will draw you a marker where it blew up - or into CV Split Tuple if you want the numbers loose.

Install

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

Restart, or Manager → search "comfyui_cv". Python ≥ 3.12, a recent ComfyUI on the V3 node API, and:

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

Nothing to download model-wise - it's a scan over an array.

Common issues

It reports out of range and you're certain the data is fine. Check the bounds against what the socket actually carries. A MASK that arrived as 0/255 uint8 is not in 0–1, and a normalised float array can legitimately contain a small negative value from floating-point error.

The run died instead of returning false. You turned quiet off. That's the documented behaviour, and it's a legitimate way to make a bad value halt the queue instead of flowing on.

pos is garbage but ok is true. Expected - OpenCV only writes the position when it finds something. Don't read it unless ok is false.

It only checks one frame of your batch. Feed it an NPARRAY for whole-array checks; IMAGE and MASK links go through frame 0.