CV Draw Path (ordered)

Why you'd reach for this

A plain polyline shows you the shape of a path - a camera trajectory, a tracked point, a contour - but not its direction. That's usually the one thing you need. Did the drone fly the loop clockwise? Is this track going forward in time or backward? Which end is the start?

This node draws an ordered sequence and makes its direction and its ends readable, with three redundant cues on purpose: a colour (or thickness) ramp from first point to last, a different marker shape at each end, and optional heading arrows along the way. Redundant because greyscale, colour blindness and a busy background each kill one cue, and the pack's author is explicitly designing against that - he says so in several of these nodes' own descriptions.

It's the right node for any ordered 2D sequence. For a metric path, the intended chain is CV Project To Plane → CV Fit Points To Box to get pixels, then this, with CV Draw Plot Frame underneath for the axes.

How it works

Points are drawn in order, segment by segment. progress is the direction cue:

• colormap along path (default) walks a sequential ramp from start to end. VIRIDIS by default, and the tooltip's warning is worth repeating: keep a sequential map (VIRIDIS, MAGMA, PLASMA). A rainbow like JET isn't monotone in brightness, so it invents features that aren't in the data and prints as mush. On a 1- or 2-channel canvas the ramp degrades into a luminance ramp, which is the point - it still reads.
• thicken along path grows the stroke instead, for when the colour is already carrying something else.
• solid leaves the line flat and leans entirely on the end markers.

The heading arrows are more careful than they look. arrow_every = N draws one every N points, but the arrow spans a fixed arc length back along the polyline rather than to the previous sample - because consecutive trajectory samples can be a fraction of a pixel apart, and an arrowhead drawn between two near-identical points is invisible. Fixed arc length gives every arrow the same readable size regardless of speed. Arrows over a stretch that isn't moving are skipped: an arrow there would point nowhere.

gap_mask is the honesty feature. (N,) of 0/1, where 0 means this point was held rather than measured - found_mask from CV Visual Odometry (Sequence) is the canonical source. Held points get a small x marker and their segment is drawn dotted, so a stretch the estimator gave up on can't be mistaken for real data. Note the reasoning in the tooltip: a held point sits exactly on its predecessor, so the segment has no length to hang a dash pattern on - the marker is what carries the information.

line_style (dashed, dotted, dash-dot) is implemented by walking the segment and stamping short solid pieces, since cv2 has no stroke pattern. Two caveats the author flags himself: the pattern restarts at every segment, so on a densely sampled path (one sample per video frame, a pixel apart) a dashed line renders solid - lean on progress and the markers instead.

Inputs and outputs that matter

• points - Nx2 or Nx1x2, in order. NaN points are skipped so a gated sequence keeps its indexing; none/empty is a pass-through.
• progress + progress_colormap - the direction cue and its ramp.
• start_marker / end_marker - default a filled square and an up triangle. Make them different shapes; that's half the "which end is which" answer and it survives greyscale.
• marker_radius - default 6. A single-point path still gets marked.
• arrow_every (optional) - 0 means none. 10–20 is usually plenty.
• label_ends (optional) - text at the two ends with a contrast halo: none, start/end, or the actual indices, which is what you want when the path is a frame sequence.
• gap_mask (optional) - as above.
• image and count. count is points drawn.

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, a ComfyUI recent enough for the V3 node API, or install ComfyUI CV from Manager. No model - pure CPU drawing.

Common issues

• Dashed line looks solid. Dense sampling, as described above: the pattern restarts each segment. Use progress and the end markers.
• Two paths on one canvas are indistinguishable. You varied only the colour. Give the second one a different line_style and different end markers - that's the design intent.
• Arrows absent. arrow_every is 0, or the path is standing still in those stretches.
• Output invisible on a MASK. A BGR tuple keeps only its blue component on a one-channel canvas. Use a single value.
• You want axes and units underneath. That's CV Draw Plot Frame, drawn first, fed the same world points and margin as CV Fit Points To Box.
• Contrib wheels. This pack needs the contrib OpenCV build; a plain opencv-python over it empties the contrib submodules. tools/repair_opencv_contrib.py --check / --apply.