How we test Sticraft's tools
Every Sticraft tool runs in your browser, so the result depends on code we can measure. Before we call one of the tools below finished, we run it on fixed test files, read back the file it actually downloads and write the numbers down with a date. This page shows those numbers, including where a tool falls short.
Last updated:
The method
- Fix the input. Each test uses the same files every time, most of them kept in the site's automated test suite.
- Run the real tool. We drive the shipped page in a real browser (Chromium) and save what it downloads, not a copy of the algorithm.
- Measure what matters for the job: position error for G-code, file size and selectable text for PDFs, dot coverage for DTF prints, edge accuracy for cut-outs.
- Write it down with the date, settings and verdict, and say on the tool page what the tool cannot do.
- Run it again when the code changes. Several of the numbers below are checked by automated tests, so a change that breaks them fails the test run.
We have not yet put the same files through competing tools side by side, so this page makes no "better than" claims. Those runs are listed at the end as open work.
DXF to G-code: exact arcs
Measured 23 August 2026 · DXF to G-code
Test drawing: one line, a full circle (radius 25 mm), a 90° arc (radius 20 mm) and a closed polyline with a rounded corner, in millimetres. Settings: GRBL laser, feed 1200 mm/min, travel 3000 mm/min, power S450, one pass. We compared it with the route a converter takes when it reads the drawing as short straight lines (12° per step).
| Route | Motion lines | Worst error from the true curve | Arc moves |
|---|---|---|---|
| Sticraft, exact geometry | 13 | 0 (below 0.000000001 mm) | 4 real G2/G3 |
| Same drawing as straight segments | 51 | 0.137 mm | none |
The straight-segment error is about 0.55% of the radius, so it grows with the part: 0.548 mm at 100 mm and 2.74 mm at 500 mm, while the exact route stays at zero. Splines have no exact G-code form; they are split into segments that stay within 0.02 mm of the curve, and the report counts them.
Image vectorizer: compression vs resolution
Measured September 2026 · Image Vectorizer
67 test cases built from ten kinds of artwork (logos, outlined art, line drawings and others), each saved as JPEG at quality 30, 50 and 75, as WebP at quality 60, and shrunk to 96 px and 192 px wide. Every traced result is scored against a clean reference of the same artwork.
- Heavy JPEG compression barely moved the result.
- Size did: a line drawing about 96 px across has strokes close to one pixel wide and cannot be traced reliably. Scan or export it larger.
- The three examples shown on the vectorizer page reach an edge accuracy of 1.00 (outline within 2 pixels of the original).
PDF compression
Measured 13 July 2026 · PDF Compressor
A 5.19 MB PDF with text and two photos, compressed in Chromium:
| Input | Output | Change | Text afterwards |
|---|---|---|---|
| 5,186,094 bytes | 157,618 bytes | -97.0% | still selectable |
The compressor keeps text and vector drawings as they are and recompresses the images, so the result stays searchable. A PDF that is already small, or made only of text, shrinks much less.
PDF editor: replacing a line of text
Measured 16 October 2026 · PDF Editor
Test file: a three-page A4 invoice, with one page rotated 90° like a phone scan, plus test pages at 0°, 90°, 180° and 270° with and without a cropped page box.
- Replaced words land on the old baseline and size within 1.5 pt on every rotation and on the cropped page, as real text that can be selected.
- Turkish characters (ı, ş, ğ) come out as real text in an embedded font; a symbol the font lacks (✓) becomes a small image.
- Known limit: the old words are covered with the paper colour, not removed from the file. For sensitive data use Redact PDF.
DTF converter: print files at 300 DPI
Measured 27 October 2026 · DTF Converter
Inputs: a 1600 × 1200 design on pure black, a 256 × 256 logo on white, a pet photo, and six free space photos (nebulae, planets, astronauts) 2400 px wide. Defaults: front preset (25 cm), halftone at 35 LPI.
| Input | Downloaded file | DPI in file | Half-transparent pixels |
|---|---|---|---|
| Design on black | 3000 × 2256 px | 300 | 0% |
| Logo on white | 3000 × 3000 px | 300 | 0% |
| Pet photo | 3000 × 3390 px | 300 | 0% |
| Six space photos | 10 in wide | 300 | 0% |
- The share of pixels that get ink stays within 5% of the original coverage at 20%, 50% and 80% (automated test).
- Glows removed from a black background, laid back over black, match the original pixel within 2 levels out of 255.
Enlarging small files: photos shrunk to a quarter and scaled back up to full size, compared with the original (PSNR, higher is closer):
| Photo | Browser's own scaling | Sticraft (Lanczos-3) |
|---|---|---|
| Moon | 32.06 dB | 32.58 dB |
| Earth | 27.57 dB | 27.98 dB |
| Nebula | 25.07 dB | 25.47 dB |
Known limits: one round-dot screen at 22.5°; removing a black background also clears very dark shades inside the design; no AI upscaling. The dots have not yet been tested on real presses and fabrics.
Background eraser: where it is weak
Measured 22 August 2026 · Background Eraser
Ten test photos, each 500 px or smaller on the long side. The cut-out runs on your device with a 4.4 MB model; the first use downloads about 18 MB in total, which the browser then keeps.
- 8 of 10 pass our basic check, 4 of 10 the strict one.
- The subject is kept (97 to 100%) on nine of ten. The misses are thin parts such as fork tines, and see-through gaps such as wheel spokes left partly filled.
Edge accuracy (boundary F1, 1.0 is perfect) falls as the photo gets bigger, because the mask is made at 320 × 320 and stretched:
| Photo | 0.2 MP | 3.5 MP | about 12 to 14 MP |
|---|---|---|---|
| Person holding an object | 0.805 | 0.577 | 0.418 |
| Orange kitten | 0.974 | 0.588 | 0.373 |
So for the cleanest edge, give it a 1 to 2 MP copy of the photo. This one never sends the photo anywhere.
Not tested yet
- The same files through competing tools (Vectorizer.AI, remove.bg, TinyPNG, Smallpdf, DTF halftone sites). Those tools need the file uploaded; we will run them with our own test files and publish the results here.
- Image compression quality at equal file size (SSIM).
- DTF dot sizes on real presses and fabrics.
Found a file that a tool handles badly? Send it through the contact page. Hard files are what the test sets are made of. How files stay on your device is shown on the privacy proof page.