OpenCV distanceTransformWithLabels_0

distanceTransformWithLabels_0 is the smarter cousin of the plain distance transform, and it answers a question the plain one can't: not just how far each foreground pixel is from the nearest background, but which background region it's nearest to. It runs cv2.distanceTransformWithLabels and hands you two outputs - the distance field and a labels map - and together they let you do nearest-region assignment on a mask.

Where that earns its keep: any time you need to associate parts of a foreground object with their nearest background islands. Think matting - you have a subject mask floating over a scene, and you want to know which side of the subject sits closest to which background patch, so a re-render or a gradient can fall off in the right direction. Or think of it as a quick Voronoi-style partition of the foreground driven by the background's shape. It's more niche than the plain feathering use case, and honestly most people will never need it - but when the question is "which background is this mask region nearest to," there's no easier way to ask it.

The inputs that matter:

• src - NPARRAY, an 8-bit single-channel binary image. Same rule as the plain transform: a BGR or multi-channel input trips the img.type() == CV_8UC1 assertion from the README - convert with cvtColor (code 6) first.
• distanceType - INT (2 = DIST_L2, 1 = DIST_L1, 3 = DIST_C).
• maskSize - INT, 3 or 5. Small wrinkle: if you set labelType to DIST_LABEL_PIXEL (1), OpenCV expects maskSize 3; keep DIST_LABEL_CCOMP (0) if you want the 5×5 mask.
• labelType - INT. 0 = DIST_LABEL_CCOMP - each connected component of background pixels gets its own label, so all the zero pixels in one blob share an ID. 1 = DIST_LABEL_PIXEL - every background pixel is its own label; more granular, usually overkill.

Optional dst and labels are out-parameters - skip both; the node returns everything on its outputs. Outputs: nparray_0 (the distance field, same as the plain transform) and nparray_1 (the labels - an integer image where the value at each foreground pixel is the label of its nearest background region). Wire the distance output to Nparrays2Image for a soft mask, or pipe the labels somewhere that can use an ID map.

Pack-wide plumbing applies: Image2Nparray in, Nparrays2Image out, batch_size==1 only.

Install: ships in geroldmeisinger/opencv-comfyui - Manager → search "OpenCV", or git clone https://github.com/geroldmeisinger/opencv-comfyui into custom_nodes, restart, with opencv-contrib-python installed. The pack-wide Cannot import name 'guidedFilter' conflict error blocks everything if you have duplicate OpenCV wheels. distanceTransformWithLabels_1 is the identical overload twin.