Flat grid
Regular spacing and opaque subjects are the detector's strongest case. The extra region shows why preview and removal remain necessary.
Public Benchmark
Icon sheet extraction benchmark
This benchmark records what automatic extraction found, what a human reviewer expected, and where the tested detector failed. It is designed to make IconSplit's strengths and limits inspectable rather than hiding them behind a single success rate.
Case measurements and verified outputs
Detection was measured on October 17, 2026. The October 30 audit checks stored files, not a new run of the current detector. Detected regions may include text or controls; they are not successful asset counts.
| Source | Original regions | Prepared regions | Manual selections |
|---|---|---|---|
| ChatGPT web | 16 | Not cropped | 8 |
| Gemini web | 1 | 18 | 8 |
ChatGPT web
Measured: 2026-10-17. Processing snapshot SHA-256: d75b331f5d2a576dba365545a6ee233958e8806e3330c5906468d3793ea1111a
Pale decorative circles remain; manual selections are not automatic successes.
Inspect output dimensions and transparency
- logo.png: 78 × 69px; 1849 fully transparent pixels, 0 partially transparent pixels.
- hero.png: 618 × 491px; 110027 fully transparent pixels, 0 partially transparent pixels.
- files.png: 166 × 168px; 7137 fully transparent pixels, 31 partially transparent pixels.
- planning.png: 164 × 170px; 6818 fully transparent pixels, 37 partially transparent pixels.
- messages.png: 165 × 169px; 6918 fully transparent pixels, 40 partially transparent pixels.
- updates.png: 166 × 165px; 7460 fully transparent pixels, 0 partially transparent pixels.
- insights.png: 165 × 165px; 7366 fully transparent pixels, 0 partially transparent pixels.
- security.png: 166 × 166px; 7666 fully transparent pixels, 0 partially transparent pixels.
Folder Repair comparison changes 24 pixels. This is not a visual-quality score.
Gemini web
Measured: 2026-10-17. Processing snapshot SHA-256: d75b331f5d2a576dba365545a6ee233958e8806e3330c5906468d3793ea1111a
Original yields one website container; original folder crop has zero transparent pixels. Prepared JPEG assets retain edge/compression limitations.
Inspect output dimensions and transparency
- logo.png: 39 × 45px; 1087 fully transparent pixels, 26 partially transparent pixels.
- hero.png: 292 × 217px; 16371 fully transparent pixels, 0 partially transparent pixels.
- files.png: 64 × 53px; 1295 fully transparent pixels, 69 partially transparent pixels.
- planning.png: 57 × 59px; 1107 fully transparent pixels, 4 partially transparent pixels.
- messages.png: 64 × 56px; 1632 fully transparent pixels, 57 partially transparent pixels.
- updates.png: 54 × 58px; 1774 fully transparent pixels, 21 partially transparent pixels.
- insights.png: 63 × 55px; 2073 fully transparent pixels, 9 partially transparent pixels.
- security.png: 52 × 59px; 1398 fully transparent pixels, 29 partially transparent pixels.
Folder Repair comparison changes 11 pixels. This is not a visual-quality score.
Two samples do not establish platform superiority. Exact image-engine versions are unverified. SVG output traces paths and does not recover original layers. Case artwork is not part of the CC0 icon catalog.
Download full records (JSON) · Output measurements (CSV) · Source and output evidence package
Historical baseline: July 2026
These results were generated from the same browser-side extraction path used by the product and manually reviewed against the visible source icons. Runtime is a local measurement, not a universal speed claim.
| Fixture | Expected | Detected | Runtime | Reviewed outcome |
|---|---|---|---|---|
| Flat grid | 25 | 26 | 660 ms | One extra region |
| Hand-drawn grid | 50 | 49 | 701 ms | One missed icon |
| Translucent glass grid | 9 | Not scored | Timed out | Known failure case |
Baseline recorded July 11, 2026, using frontend-fast-v1. These historical results have not been rerun against the current detector. Display WebP images below are previews, not the original evaluation inputs. A timeout remains unscored; it is not a current failure-rate claim.
Hand-drawn grid
Light strokes, uneven edges, and smaller subjects increase the chance that one icon is merged with background noise or missed.
Translucent glass grid
Low-contrast transparent boundaries are intentionally kept as a visible failure case. Manual selection can crop them, but clean background removal is not yet reliable.
What is measured
Expected visible icons, returned regions, reviewed false positives or misses, and local processing time. Extraction quality is judged before any manual correction.
What is not measured
The benchmark does not claim universal accuracy, compare unreviewed competitors, or treat processing speed from one machine as a service-level guarantee.
How to reproduce it
Upload each fixture through the homepage, keep automatic detection settings at their defaults, and compare the result list with the visible source icons.
Why preview and manual correction still matter
Automatic extraction is a proposal, not an irreversible export step. A dense or translucent source can create an extra crop, miss a small icon, or include a nearby fragment. IconSplit keeps the detected regions visible so users can remove a false result, repair an edge, or switch to manual selection before exporting.