On the 13 classic ESICUP strip-packing instances, Zorocut's engine reaches a mean density of 76.94 % in the app's default 3 seconds and 78.67 % in 30 seconds. The best published method, sparrow, reaches 85.42 % with 20 minutes on 3 threads; PackingSolver reaches 81.10 %. We are behind on average, ahead of PackingSolver on two instances — and the time budgets are far from equal.
What we measured
ESICUP's 2D irregular strip-packing instances are the shared yardstick of nesting research: 13 sets of polygons, from ALBANO to TROUSERS, to be packed into a strip of fixed width as short as possible. The data is public (CC0) and we used the common JSON format from OR-Datasets (MIT), converted to our engine's input with a small script.
- Density = the area of the parts divided by the material really consumed.
- Spacing 0 and margin 0, as the instances define them. Every layout is checked for overlaps on the exact part outlines before it counts.
- Our runs: Apple M-series, one thread. 3 s is the budget the app uses by default (mean of 3 seeds); the 30 s column is 1 seed.
- Published runs: sparrow's mean of 100 runs at 20 minutes × 3 threads, and PackingSolver from the same table in Gardeyn, Vanden Berghe & Wauters (arXiv:2509.13329).
The results (density, %)
| Instance | Zorocut 3 s | Zorocut 30 s | PackingSolver | sparrow |
|---|---|---|---|---|
| ALBANO | 82.36 | 83.21 | 84.95 | 89.47 |
| SHAPES2 (BLAZ) | 76.93 | 77.47 | 85.49 | 84.68 |
| DAGLI | 81.47 | 82.91 | 84.17 | 89.26 |
| FU | 84.54 | 87.46 | 90.39 | 92.24 |
| JAKOBS1 | 80.67 | 81.67 | 81.67 | 89.09 |
| JAKOBS2 | 73.21 | 74.23 | 80.42 | 84.77 |
| MAO | 79.03 | 81.26 | 75.94 | 86.14 |
| MARQUES | 85.23 | 87.36 | 85.48 | 90.93 |
| SHAPES0 | 60.06 | 61.51 | 66.50 | 68.60 |
| SHAPES1 | 64.34 | 67.67 | 72.55 | 75.69 |
| SHIRTS | 83.16 | 84.24 | 85.99 | 89.66 |
| SWIM | 65.98 | 68.93 | 71.44 | 78.26 |
| TROUSERS | 83.30 | 84.77 | 89.30 | 91.73 |
| Mean | 76.94 | 78.67 | 81.10 | 85.42 |
How to read it
- Against sparrow: about 8.5 points behind at 3 s, 6.8 points behind at 30 s.
- Against PackingSolver: about 4.2 points behind at 3 s, 2.4 at 30 s. Zorocut is ahead on MAO (at both budgets) and MARQUES (at 30 s), and level on JAKOBS1 at 30 s.
- The budgets are not the same: sparrow's figures use 400 times the app's 3 seconds, on 3 threads. Short-budget sparrow figures are not published, so a fair same-time comparison needs our own run of sparrow (MIT licence) — that is our next measurement.
- Everyone struggles on the same sets: SHAPES0 tops out at 68.60 % even for sparrow, because its pieces simply do not interlock well.
Why the gap exists
Zorocut places each part bottom-left on a fine raster and searches the order in which parts are placed. Sparrow and the strongest methods of 2013–2019 work differently: they move parts continuously and let them overlap while they search, then separate them. That class of method is the next step for our engine if shops show us that the gap matters on their jobs.
What it means for a shop
Most job-shop work is rectangles and simple brackets, where the gap between methods is smaller: on our own rectangles job the engine fills 88–89 %. That “most jobs” is our estimate, not a measured share. The engine also has to respect what a benchmark ignores — a part spacing above zero, an edge margin, rotation rules and holes for part-in-part — and it does that on every layout it returns.
The same engine runs in your browser. Nest your own DXF, see the utilisation and a price per part — free, no sign-up, nothing uploaded.
DXF online schachtelnHäufige Fragen
What is ESICUP?
The EURO Special Interest Group on Cutting and Packing. Its dataset repository holds the classic nesting instances that research papers compare on; it is published under CC0.
Is 77 % density good?
It depends on the shapes: the same engine reaches 85 % on MARQUES and 60 % on SHAPES0. Compare methods on the same instances and the same time budget, which is why the table shows both.
Where does the nesting run?
On your computer: natively in Zorocut Nest, and as WebAssembly in the browser for the online nester. Customer drawings are not uploaded to be nested.
Will you publish the same-budget comparison?
Yes. Running sparrow ourselves at 3 s and 30 s on the same machine is the next measurement; this post will be updated with it.
