Pixels only. The reference SVG is not passed to the tracer.
Development benchmark
Five logos. Fifteen controlled inputs. Failures included.
This test asks a narrow question: how closely does the Vectorize prototype recover the color regions and topology of clean logo rasterizations at two sizes, plus a lightly shaded variant?
- Run
- Runner snapshot
344ddaf- Cases
- 5 logos · 15 inputs
- Dataset role
- Development, not independent
Method
The tracer sees only pixels.
Reference vectors are used to create and evaluate each test. They are never passed to the reconstruction pipeline.
Connected areas become a compact, automatic color palette.
SVGEach region is rebuilt as scalable geometry and flat fills.
99.34%region IoU Topology matched
The SVG is rendered back to pixels for the same comparison.
Visual results
Resolution is part of the result.
The left panel is the controlled raster input. The two output panels embed the generated SVG files directly—not screenshots of them—and let you download the same vectors. The 128px trace is enlarged so lost geometry remains easy to inspect.
Adidas
Lettering, sharp corners and counters
Good at 512px. At 128px the mark remains recognizable, but corners soften and the lettering loses precision.
Slack
Four colors, separated parts and curves
Good in both clean sizes. The four regions stay separated, with less regular curves in the smaller trace.
Wikimedia
Circles, symmetry and three colors
Good in both sizes. At 128px, circles and hard corners show a small visible approximation.
Starbucks
Thin details and enclosed white regions
The 512px result is similar but needs work in internal details. The 128px trace is not suitable for professional use.
Google 2015
Multicolor lettering and open counters
Good at 512px. The 128px trace stays recognizable, while stroke weight and letter geometry become less precise.
Measured results
Region overlap is a diagnostic, not a grade.
Mean color-region IoU is the intersection divided by the union of each rendered color region, averaged across colors. It is not an overall quality percentage and does not replace visual review.
| Logo | Colors | Region IoU · 512 | Region IoU · 128 | Topology | SVG · 512 |
|---|---|---|---|---|---|
| Adidas | 1 | 98.74% | 94.52% | Matched | 4,940 B |
| Slack | 4 | 99.34% | 97.38% | Matched | 3,763 B |
| Wikimedia | 3 | 99.49% | 98.12% | Matched | 2,269 B |
| Starbucks | 2 | 97.75% | 89.44% | Did not match | 26,179 B |
| Google 2015 | 4 | 96.71% | 88.69% | Matched | 4,424 B |
Geometry and rendering
Across clean 512px inputs, mean region IoU ranges from 96.71% to 99.49%. At 128px the range widens to 88.69%–98.12%. A high silhouette score can still conceal incorrect internal detail.
Topology
Components and holes match the reference for four of five logo families at both clean sizes. Starbucks does not match in either size; features smaller than four pixels in the 4× evaluation are ignored.
Color
Clean-input palettes are recovered exactly in these cases. On shaded inputs, the correct number of colors remains, but several representative colors become darker and white can become gray.
Complexity
The clean 512px SVGs range from 2,269 to 26,179 bytes and from 4 to 36 traced regions. This is descriptive complexity, not a smaller-is-always-better target.
What changed
The failed cases shaped the product.
The first experiment forced two colors, which failed on the one-color Adidas case and distorted multicolor palettes. The revised prototype estimates its palette, seeds disconnected regions, treats flat artwork separately and exports disconnected color shapes without a global base layer.
Starbucks at 128px improved from 71.03% to 89.44% mean region IoU, yet the face, star, hair and trademark detail remain visibly unreliable. The correct product response is to warn and review, not to call that result professional.
The Adidas, Slack, Wikimedia, Starbucks and Google marks remain the property of their respective owners. They appear here only as attributed benchmark samples, not as Vectorize branding or endorsement.