Line art quality test: how we measured it

We tested the image to line art tool on 2026-11-05: Checked live on 2026-11-05: nine drawings took a median of 11.8 seconds, between 9.2 and 15.8. In a separate local Chromium test of the current default tracing settings on 20 existing drawings, tiny fragments averaged 0.34%, with all 20 drawings below 5%. The test used 20 distinct inputs: 6 pets, 4 portraits, 3 objects, 3 buildings or scenes, and 4 synthetic logos. This is an image transformation rather than a classifier, so a false-positive rate does not apply.

Speed: nine drawings timed on the live site from one cloud machine. Outlines: 20 existing drawings were traced in local Chromium using the current default setting. We counted paths with area below 0.1% of the canvas; 742 DXF outlines were read back closed without audit errors..

Timing: we used the live site and wrote down the time from pressing the button to seeing the drawing, nine times in a row. Outlines: we took 20 drawings the tool had made, traced each one in the browser, and counted how many separate outlines each file contained and how many were tiny. The speed figure is one machine at one moment; the time you see depends on your network, the style you pick and how busy the service is.

The timing is a small sample, not a promise for your network or busy times. At default, the mean ink overlap (IoU) was 0.641; fine details can disappear, and a lower clean-up setting can retain more fragments. Text and logos may change. The files are tracing references; we have not tested them on a cutting machine..

The browser thresholds the generated drawing, closes small gaps, removes small specks and holes, traces outlines and removes tiny paths. At the default setting, 20 existing drawings in a local Chromium test had a mean fragment share of 0.34%, with every drawing below 5%. We read back 742 DXF outlines as closed with no audit errors. The result does not establish that files are ready for a cutting machine.

The timed part is the whole wait on the live site, measured from a cloud test machine, not from a visitor’s browser. Network conditions, the selected style and demand can all change the time you see, and it can be longer when the service is busy.

The examples page pairs outputs from two editing paths for each test input. Look at whether the important shapes are present, whether unwanted elements were invented and whether the line weight is useful for your purpose. Small text and logos are particularly fragile. If you need faithful geometry, use a dedicated vector editor and inspect every exported path. We did not validate these results as ready for any cutter or engraving machine.

Line art has no single right answer. We measure speed, the share of runs that return line art, and how many separate outlines the traced files contain.