A client sent me an image generated by Midjourney this year— a scalloped floral border, petals fading into a mesh ground so fine it looked like breath on glass. Beautiful file. Also: no width, no repeat, no fibre, and a ground finer than any yarn we can knit. My digitizer opened it, looked at me, and asked the only question that mattered. "Which part does she actually care about?"

Nobody had asked her that. She had spent two weeks perfecting the image and three minutes writing the email.

That gap is the whole problem, and it is not a problem with AI. We build lace from concept images constantly — hand sketches, vintage swatches, museum photos, runway screenshots. A render is one more concept image, and a better one in some ways: clean, high-contrast, easy to trace. What breaks the first sample is that a render asserts twenty attributes at once with equal confidence and never says which one you would give up.

An AI render is a strong starting point, but scale, construction and stitch logic turn it into a sample that can actually be produced.

Send the render, then answer three questions

The advice circulating now says AI images aren't manufacturable, so build a proper tech pack first. For lace, that will cost you money.

A tech pack you guessed at is worse than no tech pack. Leave the ground-mesh spec blank and my sample room calls you. Fill it in with a number an AI suggested and we build to it — technically correct, completely wrong. A blank triggers a conversation. A wrong number triggers production.

So send the render, with these three answers in the body of the email:

How wide is it, in millimetres? Renders have no scale, and this is no formality — it changes the construction route, the machine, and which details survive. A motif that reads beautifully at 90 mm may have nothing left at 25 mm.

Where does it sit on the garment? A bra wing, a veil edge, a curtain hem and a cuff differ in stretch, hand feel and edge stability. Same picture, four different products.

If we can only keep one thing, what is it? Scallop silhouette? Petal density? Colour transition? Airiness? Rank them. This one sentence beats a spreadsheet of invented tolerances, because it tells my digitizer what to protect when the machine forces a compromise — and it always forces one.

Add a fibre direction and target price band if you have them. Everything else we propose and you approve.

Width, end use and design priority give the sample room more useful direction than invented technical specifications.

Your render is one of two very different products

Before anyone opens a design file, someone decides which machine builds it. Your render cannot show this, and it constrains everything downstream.

Embroidery offers flexible re-digitizing during development; warp knitting is built around continuous yarn movement and installed machine hardware.

Embroidery routes forgive during development. We re-digitize, re-sew, you see a new sample. That flexibility is why most AI-render projects start there.

Warp-knit is a different negotiation. On a multibar machine, the independent yarn systems available in a motif come from the pattern bars physically installed — hardware, not a software setting, and no digitizing skill lifts that ceiling. When your render shows eight yarn behaviours interacting and the machine supports fewer, something gets merged. Better you choose what merges than we choose for you.

A third case deserves flagging. If your render has that weightless, no-visible-ground quality — motifs floating with nothing behind them — you are describing water-soluble lace, embroidered onto a base that dissolves away afterwards. It produces the airiest real lace there is, and it is least forgiving of a render's ambiguity. With no ground left, every element must connect physically to something, and the stitch paths have to intertwine rather than sit side by side. A design where three petals touch only visually leaves the wash bath in three pieces. Our digitizing and proofing route adds connection points during translation, and they stay visible in the finished lace — so decide together where they go instead of finding them on the sample.

Why the finest detail is the first thing to go

One number governs everything: a satin column should not go below about 1 mm wide. A very good digitizer holds 0.8 mm. Below that we are not being conservative — we are describing a physical impossibility.

A #75/11 embroidery needle runs roughly 0.75 to 0.8 mm thick. Forty-weight thread is about 0.4 mm across. Ask that needle to zigzag inside a 0.5 mm channel and it works in a space narrower than itself. It deflects. The thread shreds from friction heat. Or it punches the same spot until nothing is left to anchor into.

Image models don't know this. They render detail at whatever fineness looks good on screen, finer than any needle resolves. The veining inside your petals, the hairline outline around each scallop, the three-tone gradient in a 2 mm band — those come back changed. Better to know now than be surprised.

Scale is the lever you control. Take the design up 10 to 15% and narrow columns widen naturally, no redrawing needed. Take it down more than 30% and complex elements reliably fall apart, because marginal features cross below the floor all at once. If your render only works at 120 mm and your product needs 40 mm, that is not a digitizing problem for later. That is a design decision for now.

What actually changes between render and sample

Two things get engineered during translation. Neither is a compromise — both are how real lace is built.

The repeat. Your render is one frame. Lace is continuous, so we extract a tile that connects seamlessly on every edge. Repeat alignment is the most common failure in lace development — an eyeballed repeat accumulates error along the run until the join shows. Getting it right sometimes means shifting a motif a few millimetres from where your render put it. On a border, invisible. On a strongly geometric all-over pattern you will notice, so we show you the tile before we commit.

Pull compensation. Thread under tension pulls the fabric inward, so a true circle in the artwork gets drawn deliberately oval and reads round in the finished lace. Every shape in your render gets this treatment. That is why file and fabric never match exactly side by side. The fabric is correct. The file was the instruction.

The first sample proves less than most people expect. It confirms the pattern translated, the scale works, the hand feel is right, the edge behaves. Colour is not on that list — colour approval runs on its own track through lab dips and shade bands, and judging a shade off a sample sewn in whatever white sat on the machine leads people astray. Wash it. Sew a length onto your actual fabric. Then talk about colour.

A better second round

Most AI-render projects need two sampling rounds, and the second is where the design gets good — if you use it properly.

Marking the physical swatch preserves what worked in round one and gives the digitizer precise, usable feedback.

Do not send a new render. Send the physical sample back with notes on it. "This petal edge is right, keep it." "Ground too stiff for a bra wing." "Scallop depth needs to come down." My digitizer acts on a marked-up swatch immediately. A fresh image restarts the translation from zero and discards everything the first round taught us both.

The clients who reach production fastest do not have the best prompts. They treat the render as the opening argument rather than the verdict.

Start here

Before your next handoff, do four things.

1. Put a millimetre width on the render. Every constraint downstream flows from this one number.

2. Write your priority order in one sentence. Silhouette, density, airiness, colour transition — rank them, so we know what to protect.

3. Leave blanks blank. Do not let a tool fill in ground mesh, stitch density or yarn count. A blank gets you a phone call. A wrong number gets a wrong sample.

4. Ask which construction route we propose, and why. If the answer is warp-knit, ask what the machine will merge before we build, not after.

The mistake to avoid is treating your render as a specification you must defend. It describes intent, and intent is exactly what a sample room needs. We have the machines and we know their limits. What we cannot guess is which part of the picture you fell in love with — and that, more than any technical document, decides whether the first sample lands close.