cv2.circle

What it actually is

cv2.circle is OpenCV's "draw a circle on this array" call, exposed as a ComfyUI node with the same arguments. Nothing clever is happening: no model, no GPU, one function call, microseconds. That's the pitch. The deterministic primitive layer is exactly where you should stop burning a diffusion pass - a ring, a vignette, a marker on a detected blob is a two-line graphics operation, not a sampling problem (post-processing.md is the long version of that argument).

It ships in ComfyUI CV (bmad4ever/comfyui_cv), a pack of ~470 auto-generated raw cv2.* wrappers plus a few hundred hand-written curated nodes. This one is in the raw bucket, under image/CV/low-level/cv2 C, which is your cue that the author generated it from the OpenCV type stubs and did not sugar-coat it.

How it works

You hand it a canvas, a centre, a radius and a colour, and it rasterises the ring into a private copy of that array. Circle rasterisation in OpenCV walks the outline, so the cost is proportional to the circumference, not the area - drawing a filled circle is cheap, and drawing 200 thin ones is still cheap.

The canvas socket is polymorphic: a ComfyUI IMAGE, a MASK, or an NPARRAY all go in, and the output echoes whatever you plugged in. Feed it an IMAGE and you get an IMAGE back; feed it a MASK and you get a MASK back, drawn onto. So the two real uses are (a) annotate/visualise, and (b) synthesise geometry - draw a filled disc onto a black canvas and hand the result to an inpaint node as a mask. Circle is on the pack's per-frame-safe list, too, so a 4-frame IMAGE batch comes back as a 4-frame batch with every frame drawn on, not just frame 0.

Inputs and outputs that matter

• source - the canvas. IMAGE / MASK / NPARRAY.
• center - a CV_TUPLE, i.e. one composite (x, y) value. It travels as a whole, so you either type the two numbers into the node or wire it from CV Tuple; you cannot connect just the x. Read the author's tooltip as a hint: build the point once, share it.
• radius - plain INT, pixels.
• color - a literal string, "(0, 255, 0)". BGR, not RGB - that green is green. A bare number broadcasts: "255" means white. Note the default is (0, 0, 0, 0), which is black, and on a black background that is a very convincing "the node is broken".
• thickness (optional, 1) - negative means filled (-1).
• lineType (optional, LINE_AA) - antialiased by default. Switch to LINE_8 if you want a hard, aliased 1px stroke.
• shift (optional, 0) - fractional bits in the centre/radius. shift=1 means your (100, 100) is actually (50, 50). Leave it alone unless you are deliberately working in sub-pixel space.
• result - the drawn canvas, same type as source. Wire it to a preview, or on to cv2.bitwise_and / a mask consumer.

Optional inputs are advanced - collapsed until you hit "show advanced inputs" on the node.

Installing the pack

ComfyUI Manager → search ComfyUI CV. Manually:

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

Restart ComfyUI. That's the whole install - the single dependency is the OpenCV wheel (opencv-contrib-python-headless~=5.0.0.93, per requirements.txt). The two hard gates are worth knowing up front: Python ≥ 3.12, and a ComfyUI recent enough to have the V3 node API. Older builds will just fail to load the pack.

Where people get burned

• Black circle, black background. Covered above, still the #1 report.
• "It needs contrib, not plain opencv." The cv2 namespace is shared by all four OpenCV wheels, so if something else pip-installs opencv-python over your contrib build, the contrib submodules go quiet and a chunk of this pack's nodes disappear. The pack ships tools/repair_opencv_contrib.py with a --check and an --apply for exactly this.
• Old ComfyUI portable installs with a broken cv2. ImportError: DLL load failed while importing cv2 is a real, well-documented failure around OpenCV-based node packs, and it takes every pack that touches cv2 down with it. If the whole pack fails to load rather than just this node, that is the class of problem you are in - not a bug in circle.
• Set your expectations on support. The README is unusually blunt: LLM-assisted development, self-declared test overfitting risk, "not recommended for production", updates not planned. And the pack has essentially no Reddit footprint yet - searching the corpus turns up nothing about it. Read the node, test on your own images, don't file an issue expecting a fix.