cv2.erode

Erosion is the mask-editing primitive that people keep reinventing badly. If your seg mask has speckles, if two blobs are fused where they shouldn't be, if a detector's box-shaped mask needs trimming before you paste a re-rendered region back - this is the node. What it is not is a GrowMask substitute, and half of the friction with it is that the kernel socket is not a widget.

What erosion does

For every pixel, take the minimum over the neighbourhood defined by the structuring element. A white pixel survives only if all of its neighbours inside that element are white. So white regions shrink, thin structures disappear, and isolated bright specks get eaten - which is exactly what you want when you're cleaning a mask rather than a picture.

Two details bite. First, the element's shape matters as much as its size: a rectangular element erodes corners differently from an ellipse or a cross, and for masks, a cross or ellipse is usually what you meant. Second, iterations is not the same as one big kernel - repeated 3×3 cross erosions approximate a diamond, not a square, and the tooltip is explicit that the default is OpenCV's 1. Since this pack's wrapper has borderType (with BORDER_WRAP unsupported) and a borderValue literal, edges are yours to control too.

The inputs that matter

src accepts an IMAGE, a MASK or an NPARRAY - and, unusually, a LATENT. The pack treats erosion as channel-agnostic and value-range-agnostic, so a LATENT gets processed in latent space as a float32 array, frame 0, values untouched, no 8-bit round trip. That's a legitimate trick if you're experimenting with latent-space masks; it is also the kind of thing you should try once on purpose rather than discover by accident.

kernel is the one that stops people. It is NPARRAY only - you cannot wire a MASK or an IMAGE into it, and you cannot leave it empty, even though OpenCV itself would happily default to a 3×3 rectangle. Build it from cv2.getStructuringElement (another auto-generated wrapper in the same pack, category image/CV/low-level/cv2 G, with a shape dropdown and ksize_x/ksize_y widgets), or hand-write a small matrix with Parse Matrix for something custom. Then anchor (-1, -1 = centre, a CV_TUPLE you can wire from CV Tuple), iterations, borderType and borderValue are the optional knobs - the last one takes a literal like "255" or "(0, 0, 0)".

The single output, result, echoes the input's format: IMAGE in, IMAGE out; MASK in, MASK out. Masks are the money case here, and they stay masks, so cv2.erode drops straight into a masking chain without a bridge node on either side.

Where it fits

Masks, mostly. Erode to kill specks and separate touching blobs before a connected-components split; erode to trim an over-generous inpaint mask so the re-render doesn't eat the boundary you wanted to keep; erode a copy as the "core" of a mask you're about to distance-transform. In detailing pipelines the pressure goes the other way - masks generally need growing for seam-free pasting, and inpainting.md is blunt that generous overlap plus mask blur is what stops the "pasted on" look. If your goal is to grow, this is the wrong direction; cv2.dilate is the same node shape with the max instead of the min.

Also worth saying: for "grow or shrink this mask by N pixels", core ComfyUI's GrowMask and FeatherMask are simpler and instant. Reach for cv2.erode when you want a shaped element, multiple iterations, an explicit border mode, a raw array, or erosion on a latent.

Installing it

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

ComfyUI Manager: search comfyui_cv, install, restart. Python ≥ 3.12, recent V3-API ComfyUI. The node lives under image/CV/low-level/cv2 E - there are ~470 of these wrappers, so the category subfolder is the fastest way to find it.

Traps

The kernel link refusing to connect is not a bug; it's the NPARRAY socket doing its job. A Mat() empty kernel means "3×3 rectangle" in C++ and has no equivalent here, so build the element explicitly. Odd, non-zero sizes keep you out of trouble. And if erosion seems to do nothing, check what you actually fed it: erosion of a soft 0–1 float mask will shave sub-threshold values you may not see, so preview the mask after it rather than assuming the graph is broken.