CV Draw Labels

Why you'd reach for this

A numbered dot cloud is fine until it isn't. The moment you care which dot is the wrist and which is the fingertip - or which detection score belongs to which box, or which QR code said what - you need text next to the point.

That's the gap this fills. ComfyUI core's Draw BBoxes can label COCO-style boxes; it can't caption an arbitrary point set. This node labels N points with N strings, in order.

How it works

It walks the point array and stamps a label next to each one with cv2.putText, offset by offset_x and offset_y (defaults 4 and −4, i.e. up and to the right of the point), scaled by font_scale, in your colour at your stroke width.

Leave labels unconnected and each point is captioned with its own index - 0, 1, 2 - which is honestly the mode you'll use most, because it's how you find out the order of an array you didn't produce. (It's also how you spot that a MediaPipe hand node numbers its 21 keypoints in a specific order rather than the order you assumed.)

Wire labels and you get real text. The important detail: it's an array of strings or numbers, one per point, not a comma-separated string. CV Text To Array types one per line and produces exactly that; CV Numbers To Array does the numeric version; and several detectors already emit usable text arrays - a QR or barcode detector's texts, CV Load Labels' array, a classifier's label output. If the array length doesn't match the point count, the node raises with both counts, which is a much better outcome than silently mislabelling everything.

outline defaults on, and it's not decoration. These labels land on whatever the photograph happens to show, and a flat single-colour label at font scale 0.4 is often illegible over texture. With it on, the node picks a contrasting halo - black behind a light label, white behind a dark one - by looking at your foreground colour. Turn it off if you're compositing text onto a flat synthetic canvas.

Points with a non-finite coordinate are skipped without renumbering the rest: the pack's convention is that NaN means "no position for this item", not "an error". That's what lets you blank entries out of a gated array and keep the labels aligned with the surviving rows.

Inputs and outputs that matter

• points - Nx1x2. None or empty passes the image through, count = 0.
• font_scale - 0.4 by default. Shrink it for a dense set; a labelled 21-point hand wants about that, a labelled 500-point cloud wants almost nothing.
• color, thickness - the text itself.
• labels (optional) - the array described above. The single most useful optional input in this section of the pack.
• offset_x / offset_y - nudge labels off the dots. With a big marker underneath, 4/−4 isn't always enough.
• outline - the halo.
• image and count. The count tracks the point array you fed in rather than the labels that survived, so a blank canvas reporting a non-zero count means the points are off-canvas, stacked, or blanked out - not that the node failed.

Install

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 and a current ComfyUI (V3 node API), or install the pack as ComfyUI CV from Manager. CPU drawing only.

Common issues

• Labels unreadable. Raise font_scale and make sure outline is on; also check color isn't a BGR tuple on a single-channel MASK, where only the blue component survives.
• All labels say "0". You wired a single text value where an array belongs. Use CV Text To Array.
• Count mismatch. Array length ≠ point count. They describe the same set, so they have to be built together.
• Text overflows the frame. Points near the right edge with a positive offset_x. Negative offsets or a smaller font.
• Contrib trap. The pack needs opencv-contrib-python-headless; installing a plain opencv-python over it empties the shared contrib submodules and contrib nodes vanish. tools/repair_opencv_contrib.py --check, then --apply.