CV Draw Plot Frame

Why you'd reach for this

Draw a camera trajectory onto a blank canvas and you get a squiggle. Is that a three-metre wobble from a loose tripod, or a 300-metre walk around a building? The picture is identical. The numbers exist, but they're in an array nobody reads.

This node supplies the missing half of a plot: a border, a grid, ticks labelled in world units, axis names and a scale bar. Put it under your drawn path or point cloud and the squiggle becomes a figure with a size.

Critically, it doesn't measure the data - it's fed the same world extent your mapper was, and reproduces that mapping exactly. That's what stops the amusing failure where the tick marked 50 m lands a centimetre away from the point that's at 50 m.

How it works

The geometry is shared with CV Fit Points To Box - one function used by both nodes, so they cannot disagree about where a world coordinate lands, including which way up it lands. That shared logic also explains the y_axis control's existence: pixel rows grow downward, so mapping a world axis straight onto rows draws your top-down map mirrored, and a left turn comes out looking like a right turn. y down (pixel rows) is the default because points that came through Fit Points To Box are already in pixel space; pick y up (world) only if you flipped them yourself.

Tick values are rounded to 1/2/5 × 10ᵏ so the labels read as round numbers - you ask for roughly six grid lines and get whatever step makes sense across the visible span. Grid lines that fall outside the plot box are skipped rather than clamped. units is a suffix on every tick label and the scale bar. The scale bar is the one annotation that still says how big the picture is after someone crops or rescales it, which is why it's on by default in the bottom-left.

Because nothing here reads the data, it also works as an empty graticule: render the frame, then draw points or a path on top. That ordering is the intended one - draw the frame first.

margin must match the margin you gave CV Fit Points To Box. The tooltip is blunt about it, and it's the one field that silently ruins the alignment if you set it on one node and not the other.

Inputs and outputs that matter

• source - the canvas to draw on, e.g. a constant image the size you want, or the image you're overlaying. Its size defines the plot box.
• extent_from (optional) - Nx2 world points whose bounding box sets the plotted range. Pass the same array you fed Fit Points To Box; if you're overlaying two tracks, pass both concatenated. With it connected, the min/max widgets are ignored.
• x_min/x_max/y_min/y_max - the manual alternative when extent_from is empty.
• margin - must match Fit Points To Box.
• grid_x / grid_y - how many lines to aim for; 0 turns a direction off.
• units, x_label, y_label - m, px, mm; axis names, drawn bottom-right and top-left. For a top-down (XZ) trajectory of a drive, the defaults x/z are already right.
• scale_bar - a labelled bar of known world length, default bottom-left (none / bottom-left / bottom-right). It's the one annotation that still says how big the picture is after someone crops or rescales it, so leave it on.
• color / grid_color - ticks, border and scale bar vs the grid itself. The grid should stay well below the data's contrast.
• image - the only output.

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, current ComfyUI (V3 node API), or install ComfyUI CV from ComfyUI Manager. CPU drawing; the geometry came from upstream whatever produced your world points.

Common issues

• Ticks don't line up with the data. margin differs between this node and Fit Points To Box, or you fed the frame a different point set than you fed the mapper (one track instead of both).
• Everything drawn mirrored / left turns look like right turns. y_axis left on the default while your second coordinate is a world axis. Switch to y up (world). The KB's notes on reconstructing geometry from frames are the same story in a different context: sign and axis conventions are where these pipelines quietly go wrong.
• Grid drowns the data. grid_color too bright, or too many lines. Aim for 4–8.
• Labels unreadable. The frame is sized by its canvas - give it more pixels, or fewer ticks.
• Nothing visible on a MASK source. BGR literal on a one-channel canvas keeps only blue; use a single value, and expect a grey frame.
• Contrib trap. A plain opencv-python install over the contrib wheel empties the shared contrib submodules and contrib-backed nodes disappear. tools/repair_opencv_contrib.py --check then --apply.