cv2.morphologyEx

Every mask you get out of a detector, a segmentation model, a background remover or a threshold is dirty in the same three ways: speckles of noise that shouldn't be there, pinholes that should be filled, and thin bridges connecting things that shouldn't touch. cv2.morphologyEx is the deterministic, millisecond fix for all three - no model, no second sampling pass. When the KB says automatic detailing is "zooming in on a mask, upscaling, and diffusing", this node belongs in the step before: cleaning the region so the re-render happens in the right place.

The operations, in the order you'll care about them

The op dropdown is OpenCV's MorphTypes. The short version:

• MORPH_OPEN - erode then dilate. Kills isolated speckle while leaving surviving shapes roughly their original size. This is the "remove the stray dots" tool.
• MORPH_CLOSE - dilate then erode. Fills pinholes and small gaps, keeps the outer boundary roughly where it was. This is the "heal the mask" tool.
• MORPH_ERODE / MORPH_DILATE - the bare primitives: shrink or grow. Open is these two in sequence.
• MORPH_GRADIENT - dilate minus erode. A clean outline of a mask, without Canny's noise sensitivity.
• MORPH_TOPHAT - input minus its opening. Surfaces bright structures smaller than the kernel; the classic uneven-illumination flattener.
• MORPH_BLACKHAT - closing minus input, the dark-structure twin.
• MORPH_HITMISS - the odd one out. It's template matching for binary shapes, expecting a single-channel 8-bit input and a kernel of 1/0/−1 entries. Ignore it until you specifically need it.

iterations is the how-many-times knob, preset to OpenCV's 1. Two passes of a 3×3 ellipse is roughly one 5×5 but rounder; for organic masks (faces, hair, painted regions) an ellipse kernel respects the shape better than a square.

The input that trips people

kernel is typed NPARRAY and nothing else - the author's own note says it plainly: a data array, not an image. So you don't crop a shape out of a picture; you build a structuring element. The pack's cv2_getStructuringElement does it (shape: MORPH_ELLIPSE / MORPH_RECT / MORPH_CROSS, ksize a (w, h) CV_TUPLE), and the bundled cv2.morphology (simple) subgraph is exactly that pair wired into cv2_morphologyEx.

The rest of the inputs are set-and-forget: anchor defaults to (-1, -1), meaning the kernel center - only move it for deliberately asymmetric effects; borderType is a dropdown (BORDER_WRAP isn't supported); borderValue is a literal like (0, 255, 0) or a bare number that broadcasts across components.

result echoes src's format - IMAGE in comes back IMAGE, MASK comes back MASK, NPARRAY stays NPARRAY. That's deliberate: you can drop this node into a mask chain with no conversion bridges on either side.

One curiosity: morphologyEx is also on the pack's latent-safe list, so a LATENT link is processed in latent space (frame 0, float32, values untouched, no uint8 quantization). It runs. Whether min/max filtering of a latent is useful is a research question you'd be answering yourself.

Install

ComfyUI Manager → search ComfyUI CV → install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/bmad4ever/comfyui_cv
pip install "opencv-contrib-python-headless~=5.0.0.93"

Python ≥ 3.12, a recent ComfyUI on the V3 node API. Core cv2, no contrib submodule, no model files.

When it doesn't do anything

• The kernel won't connect. If you're trying to link an IMAGE into kernel, the socket only accepts NPARRAY. Build the element with cv2_getStructuringElement.
• Holes survive CLOSE. They're bigger than the kernel. Size it against the feature, not against the image - rule of thumb, roughly the width of the gap you want closed, or ~1% of the mask's width if you'd rather guess.
• Everything vanished. MORPH_ERODE with a large kernel eats a thin mask completely, and erosion of a thresholded edge map is usually the culprit. Open/close first, see what you've got, and only then push sizes.