Stitch: how we turn a logo into an embroidery file

If you run a decoration shop, you know this moment. A customer emails their logo as a PNG and asks for it on fifty polo shirts. Before the embroidery machine can sew a single stitch, someone has to turn that picture into a machine file: a long list of instructions that says where the needle goes in, in what order, and with which thread. That job is called digitizing. It is usually done by hand by a specialist, it takes time, and you pay for it on every new logo.

Stitch does it automatically, inside DesignMint Commerce, in a couple of seconds. You upload the logo, check the preview, and download a file your machine understands.

Why a picture isn’t enough for an embroidery machine

A printer can work from pixels. An embroidery machine can’t. It sews with thread, and thread has rules that pixels don’t. Every stitch has a length, and stitches that are too short snap the thread. Every time the machine moves from one shape to another, it either travels with the thread still attached or stops and cuts it. That cut is called a trim, and every trim costs time on every single garment in the order.

So digitizing isn’t tracing. It’s planning: which stitch type suits each shape, what order the shapes sew in so the machine travels as little as possible, and what foundation goes underneath so the fabric doesn’t pucker. A few terms come up a lot:

How Stitch works, step by step

1. It reads the artwork onto a fine grid: 900 cells on the long side, about 0.11 mm per cell for a 100 mm design. Every decision after this is made in real millimetres, not pixels.
2. It reduces the colours to at most one thread per needle on your machine, so nobody has to re-thread halfway through a job.
3. It cleans up slivers and specks that are too small to sew.
4. It decides the sewing order: what goes first, what sits on top, and how to travel between shapes.
5. It picks a stitch type for each shape. Narrow shapes try satin first. Shapes that are too wide, too tightly curved or too broken up fall back to tatami fill.
6. It adds underlay, graded by how wide each shape is, and routes the travel stitches underneath areas that haven’t been sewn yet, so they end up hidden.
7. It writes the file: Tajima DST (or Melco OFM) in 0.1 mm whole-number units, so the size never drifts from rounding.

Stitch is a rule-based engine, not an AI model. That was a deliberate choice. When a file comes out wrong, we can point to the exact rule that made the decision and fix it, instead of guessing what a model was thinking.

Testing it against people who do this for a living

“It looks right on screen” is a terrible test for embroidery. A design can look perfect and still sew badly. So on the second day of the project we built a benchmark.

We took 294 real production files from customer jobs, all made by professional digitizers. For each one we turned the stitches back into flat artwork, ran that artwork through Stitch at exactly the same size, and compared the two machine files. We measured what a production manager actually cares about: trims per thousand stitches, how many stitches are dangerously short (under 0.3 mm), how dense the stitching is, how wide the satin columns get, and how much of their coverage ours overlaps (1.0 would be a perfect match).

The first result was humbling. Our engine trimmed more than ten times as often as the professionals did.

Each round fixed one specific thing. Round 1 reworked the sewing order and cut trims by more than two thirds. Round 2 added proper underlay. That cost a few trims, but it closed 26% of the gap in stitch count, which matters more for how the finished piece looks and feels. Satin v2 rebuilt satin columns from the centre line of each stroke instead of tracing the outline, and it reduced trims on 195 of the 294 designs.

We’ll be straight about where we are: trims are still about three times the professional rate. Most of what’s left comes from open outlines and how travel stitches follow finished satin edges. That’s what we’re working on next.

The benchmark also caught a bug nobody would have spotted by eye. One large open design took 33.7 seconds to build, and 65 seconds in the worst case, because two search steps slowed down sharply as designs got bigger. Capping those searches and adding a spatial index brought it down to about two seconds.

Matching a real file, stitch for stitch

One university’s circle logo became our main test case, because we had the exact file a production house had sewn for it. Our first attempt was way off: 2,544 stitches against their 6,688, with 326 stitches shorter than half a millimetre. It also came out 46.7 mm wide instead of 51.2 mm.

The size problem was sneaky. We were measuring the design from the edges of the image file, not from the ink inside it. So a logo meant to be 2 inches wide sewed at 1.84 inches. After four rounds of fixes and a new “Production spec” mode, we landed at 6,679 stitches, about 20.5 metres of thread, zero short stitches and the same five trims as the professional file.

When a reviewer said we were wrong, we measured first

We showed our files to an experienced production reviewer. He pointed at the red fill under the white letters and called it “unnecessary stacking”. His suggestion was to cut the red away around every letter so the colours don’t overlap.

We built it. The stitch count went up from 6,798 to 7,950, because every hole in every letter needed its own border. The red broke into 16 separate pieces with three extra trims. It was worse on every number we track.

So we went back to the data. The professional file for the same logo had 9% of its red sitting under the letters. Ours had 10%. Across 243 multi-colour files in the benchmark, the typical design had 11% of its stitches more than a millimetre under a later colour. A little overlap isn’t a mistake; it’s how professionals stop gaps appearing between colours. We reverted the change.

He was right about other things, though. Every colour was filling at the same angle, one logo had a fan-shaped sliver at the bottom, and some tagline lettering was too small to sew. We fixed all three, and “too small to sew” is now a warning with a one-click “leave out” button.

What we learned

• Measure against real work, not your eye. The benchmark changed more decisions than any review did.
• Feedback is valuable, but check it against data before you build it. Experts are right most of the time, not all of the time.
• Size bugs hide in assumptions. “Measure the image” and “measure the ink” sound the same until a 2-inch logo sews at 1.84.
• Rules you can read are rules you can fix.

Common questions

What file formats does Stitch accept and export?

You can upload PNG, JPG or SVG artwork, or open an existing DST file. Stitch exports Tajima DST, the format most commercial embroidery machines read, or Melco OFM.

Is automatic digitizing as good as a professional digitizer?

For most logos it gets close, and it’s getting closer. On our 294-file benchmark its coverage overlaps the professional files closely, but it still trims about three times as often. For complex or very large runs, a quick human review is still a good idea, and the 3D preview makes that review fast.

How much does it cost?

Each new Stitch design uses 35 credits from your DesignMint Commerce wallet.