Connected Components With Stats 🐦

Connected components is the computer-vision answer to "how many separate things are in this binary image, and what are their sizes and positions?" After you threshold or edge-detect, every region of touching non-zero pixels becomes one component. This node runs OpenCV's connectedComponentsWithStats and hands you a count plus per-component geometry - which is the foundation for counting objects, sizing regions, or filtering by area.

What it gives you

Four outputs, which is unusually generous:

• num_labels (INT): how many components were found (including background, which is usually label 0).
• labels (NPARRAY): an image where every pixel is stamped with its component's ID - 0 for background, 1..N for each blob. Segmenting by this is how you'd mask a single detected object.
• stats (any): the per-component statistics table - bounding box x/y, width, height, and area for every label. This is the node's superpower: filter blobs by area, find the biggest one, drop everything smaller than a threshold.
• centroids (any): per-component centroid coordinates (x, y), useful for measuring or for drawing markers.

The inputs

• Np_bin: a binary NPARRAY - non-zero pixels are "1". The tooltip says it plainly: make it with compare, threshold, inRange, Canny, etc. from this pack's own OpenCV set.
• connectivity (4 or 8, default 8): whether diagonal neighbors count as connected. 8 connects more aggressively; 4 only cardinal directions.
• removeSurrounding (default true): if a region touches three-plus corners of the frame, it's treated as the image's surrounding background and removed from the output. It's a nice touch - it stops your detector from counting "the whole photo's border" as an object.

The one big constraint

This node is single-batch only. Feed it a batch of images and it raises a ValueError; the source literally checks Np_bin.shape[0] != 1. Same story as the pack's Find/Draw Contours nodes. If you get that error, that's what it means - run one image at a time.

Also: if num_labels <= 1 (i.e. nothing was found), it raises "No objects found in the binary image." Not a warning - an exception. If your image has no blobs, this node fails loudly. That's arguably correct behavior for a detector, but expect it when you feed a blank threshold result.

The workflow it belongs to

Image → To Nparray → threshold/Canny → Connected Components With Stats → use stats/centroids to filter or measure, and labels to mask individual components. In this pack's world you'd then feed centroids into Draw Circles to mark what you found.

Install

ComfyUI Manager → search ComfyUI-ArchiGraph, or:

cd ComfyUI/custom_nodes
git clone https://github.com/vincentfs/ComfyUI-ArchiGraph

Restart and run the pack's install script once. OpenCV is the only real dependency; no models.

Verdict

The stats output is what makes this worth reaching for over a plain label map - area filtering turns a naive blob count into "the three biggest regions, their bounding boxes, and their centers." Just remember it's single-image and it errors when it finds nothing.