CV Find Absent Color (Lab)

Here's the problem this node solves, and it's a more common one than the name suggests: you're drawing a bounding box, a mask outline or a crosshair on an image, and the colour you picked happens to be a colour that's already all over the picture. Green boxes over a forest. Red lines over a tomato. The annotation is technically there and practically invisible. CV Find Absent Color (Lab) hands you a colour that cannot be confused with anything in the frame - the CSS named colour that is furthest, in L*a*b*, from every pixel you gave it.

Why you'd reach for it

It is a visualization helper, and the pack is unusually disciplined about having those as separate nodes - the data nodes never draw, you choose what gets rendered. This one picks the ink. Practical wiring:

• Feed an image, take color_hex (a COLOR socket) and drop it into any colour picker input - CV Color From Hex or the drawing nodes' colour fields.
• Take color_name when you want the choice to be readable in a log or a text overlay ("rebeccapurple" is a lot more useful in a filename than #663399).
• Take color_rgb when the consumer wants three floats - it's a CV Scalar-shaped NPARRAY, exactly the shape CV Scalar emits, so it broadcasts into the low-level cv2.* drawing wrappers (this is the pack's answer to "cv2 needs a Scalar and ComfyUI has no Scalar type").

The secondary use is palette analysis: point it at a stylised render, and the absent colour it returns tells you which region of colour space your image never visits. That's a diagnostic, not a workflow, but it's a fast one.

How it works

The image is converted to OpenCV's L*a*b*, and each of the 148 CSS named colours is measured against it: for every candidate, the distance to its nearest pixel. Candidate colours within threshold of any pixel count as present. Of the rest, the one with the greatest distance wins. If nothing is absent, you still get the furthest colour back, but found flips to false so a downstream switch knows the answer is a judgement call rather than a guarantee.

Distance is selectable: Euclidean (L2) is the classic ΔE-style metric; Manhattan (L1) sums per-channel differences and is easier to reason about channel by channel. Both operate in 8-bit Lab space, so L is 0–255 and a/b are offset like OpenCV's, not the −128…127 lab you might be picturing.

The inputs that matter

• image - the frame(s) to analyse. A batch is fine, and a colour counts as present if it appears in any frame, so one pass over a whole shot covers your whole video.
• metric - Euclidean (L2) by default. Leave it unless you have a reason.
• threshold - how close a pixel has to be before a colour is "present", default 10, range 0–255 in half steps. At 0 only an exact match counts, and given there are 148 candidates you'll likely find one absent and get a found = true you shouldn't fully believe. Raise it - 15–30 is where the answer starts to mean "nothing in the image is remotely like this colour".
• max_pixels (optional, advanced) - per-frame cap on pixels searched, default 200 000, applied as an even stride. It's the reason a 4K frame doesn't take a second. The flip side is in the tooltip and it's real: a colour present in only a handful of pixels can fall between strides and be reported absent. If your subject is small, raise it.

Outputs: color_rgb (NPARRAY), color_name (STRING), color_hex (COLOR), distance (FLOAT), and found (BOOLEAN). distance is the winner's distance to the nearest image pixel - pair it with found in a conditional if you want the graph to react to a borderline result.

Install

ComfyUI Manager → search ComfyUI CV, or clone the repo into ComfyUI/custom_nodes and restart:

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

The pack needs opencv-contrib-python-headless~=5.0.0.93, Python 3.12+ and a ComfyUI new enough for the V3 node API. No models.

Traps

• A tiny threshold plus genuine colour variety → found = false. That's the node telling you the image already contains something close to every CSS colour, which usually means you set the threshold too low or the image is a colour-chart-like scene.
• CSS colours are a coarse palette. They're web colours, not a full Lab gamut. On a monochrome image you'll get something wildly saturated, which is what you want for an overlay and absolutely not what you want if you were hoping for a "subtle" annotation.
• It doesn't know the annotation colour from other overlays. Give it the original image, not one you already drew on, or the previous overlay's colour counts as present.
• Values are returned in RGB order even though the pipeline carries BGR, because the image arrives as RGB in ComfyUI. The pack is explicit about this in the tooltip; if you hand color_rgb to a raw cv2 drawing wrapper that expects BGR, expect swapped channels.