CV Chart Series

What this is for

Half the arrays in a vision workflow aren't pictures at all - they're a measurement per thing. One similarity score per detected face, a histogram, a metric per method, a score against a parameter sweep. Preview CV Array renders those as a few anonymous pixels and min-max normalization deletes the absolute scale, which is exactly the part you cared about: is this score above the 0.363 threshold your face-recognition model uses, or not?

CV Chart Series draws them as labelled bars (or lines, areas, steps, scatter) with the values readable, in the node. Hover gives exact numbers, box-zoom works, and the toolbar saves a PNG - so the picture you paste into a write-up carries the numbers, not just the shape.

How it works

The full Apache ECharts option object is built in Python - palettes, legend, reference lines, value formatting - and passed to web/cv_charts.js, which does nothing but instantiate it. All the chart logic is server-side and comes back out as the spec STRING, so you can read and diff the plot as JSON. Values are capped at 200,000 points; past that the node tells you you're charting an image by mistake rather than freezing the tab.

The inputs that matter

values is required. A 1-D array is one series of N samples; a 2-D array is samples × series, one column per series. Both (N, 1) and (1, N) count as a single series of N samples, so you don't have to reshape before feeding it.

chart is the second required input and the one that decides honesty of read: bar for independent measurements (one per object, one per method), line/area/step when consecutive samples form a signal (a histogram, score versus a swept parameter), scatter when samples have no order at all.

From the optional set, the three worth wiring:

• x_labels - name per sample, ;/newline/comma or JSON. Missing names fall back to #0, #1, … Feed the same string that labels the corresponding image tiles and bar i provably refers to tile i. That is a real difference from eyeballing a legend.
• reference_lines - value or value=label, separated like the label lists. 0.363=SFace threshold paints your decision threshold into the picture instead of leaving it in a widget, and color_by → above/below first reference line then splits every mark into pass and fail colours with a legend saying which is which. That's the whole point: a chart where you can see the threshold crossing without hovering anything.
• orientation - horizontal when the sample names are file names, person names or anything else longer than three characters. Vertical bars with rotated labels are where charts go to die.

series_names names the columns for the legend. color_by also takes series color (default) or value (gradient) with gradient_palette. value_range defaults to start at zero, which is correct for bars - auto fits the data and exaggerates small differences, manual (with min_value/max_value) is what makes two runs comparable. show_values prints each value next to its mark so the chart survives being saved as a PNG. Plus x_axis_name/y_axis_name (y wants units), palette, theme, decimals, chart_height.

Output is spec. The node is an output node, so it draws on Queue; Preview as Text reads the charted values.

Install

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

Or search ComfyUI CV in ComfyUI Manager (registry publisher bmad4ever). Python ≥ 3.12 and a V3-node-API ComfyUI; opencv-contrib-python-headless~=5.0.0.93 comes from the pack's requirements. ECharts is vendored - no model or library downloads for this node.

Common issues

• Labels one short. Not an error. Missing names get numbered placeholders by design; a chart with a short label list is still a readable chart, and the author decided a label typo shouldn't kill a run.
• Bars all look identical, or wildly different across runs. Check value_range. auto rescales to whatever is in front of it, so small differences shout. Pin it.
• Nothing appears in the node. Reload the page - the chart is rendered by a frontend module, and a tab that was already open when you installed won't have it.
• Contrib-backed nodes vanished pack-wide. A non-contrib OpenCV wheel clobbered the shared site-packages/cv2. python tools/repair_opencv_contrib.py --check then --apply.