From Photo to Technical Drawing: What AI Can Actually Do in 2026

A gear from a machine the manufacturer abandoned in 2009. A bracket someone milled from memory fifteen years ago. A coupling whose only documentation is the part itself. Every workshop has that drawer — and until the part has a drawing, nobody will quote it, machine it, or inspect it.

The part with no drawing is a blocked purchase order

Undocumented parts are not an edge case. In maintenance and defense circles the problem has its own acronym — DMSMS, for parts whose manufacturers and documentation have vanished while the equipment lives on. The civilian version is everyday reality: a broken spare on a production line, an obsolete supplier, a legacy product nobody modeled in CAD.

The part in your hand is the specification. The problem is that no machine shop accepts a part in hand as a specification. They need a document that states dimensions, tolerances, material and finish — because that document is what they quote against and what inspection measures against.

What reverse engineering costs today

The established route is 3D scanning followed by CAD reconstruction and manual drafting. Each step has a price tag:

• Hardware. Entry-level professional handheld scanners run $2,000–12,000. Standalone systems like the Artec Leo sit around $20,000–40,000. Certified metrology systems from Creaform or ZEISS start near $30,000 and pass $100,000 with software and options.
• Time. The industry rule of thumb is 2:1 — two hours of point-cloud cleanup for every hour of scanning, before a CAD engineer even begins rebuilding surfaces into a parametric model.
• Drafting. Turning the model into a dimensioned 2D drawing costs $50–100 per hour freelance, $75–200 per hour at a drafting firm.

For a single legacy bracket, "just get me a drawing" the traditional way lands between a few hundred and a few thousand dollars — with days of turnaround. For a one-off spare part, that overhead frequently exceeds the cost of machining the part itself. So people skip the drawing, email the shop a photo and a prayer, and absorb the risk.

What changed in AI research

Five years ago, generating CAD from images was a niche academic problem. It is now one of the fastest-moving corners of applied computer vision — and the milestones are concrete:

• Img2CAD (2024) was the first method to generate CAD models with editable parameters from a single image. Not a mesh — a sketch-and-extrude sequence a real CAD kernel can replay and a designer can edit.
• Text2CAD (NeurIPS 2024) showed parametric CAD construction sequences can be generated token by token, the way language models generate text.
• ReCAD (2025) layered reinforcement learning on vision-language models and set the current state of the art for both text-to-CAD and image-to-CAD, cutting geometric error substantially.
• On the reading side, fine-tuned vision-language models now extract dimensions and GD&T from existing drawings with F1 scores above 0.9 — work that was strictly manual two years ago.

Investors noticed. Adam— the Y Combinator startup Guillermo Rauch called "the v0 of CAD" — raised a $4.1M seed in late 2025 to build an AI CAD copilot. Backflip, founded by the Markforged founders, raised $30M from NEA and a16z to turn 3D scans into editable parametric CAD automatically. Geometry capture is being automated, and the direction is irreversible.

What a photo can never tell you

Now the part of the story most marketing pages skip — and the reason you should distrust any tool that doesn't mention it.

A single photograph cannot reveal the absolute size of an object. This is not a model limitation that better training will fix; it is the projective geometry of image formation. A 40 mm bracket photographed up close and a 400 mm bracket photographed from farther away can produce identical pixels. Every single-view reconstruction method in the literature recovers shape only up to an unknown scale factor. Resolving it requires outside information — a reference object, a second calibrated view, or a measurement.

And size is only the first blind spot. A photo also cannot see:

• Hidden geometry — internal bores, undercuts, blind features on the far side of the part.
• Design intent— whether that ⌀20 hole is a press fit or a slip fit lives in the designer's head, not in the part's silhouette.
• Threads and finish — an M8×1.25 and an M8×1.0 thread are indistinguishable at photo resolution.

A workflow that respects the physics

TechDraw AI is designed around the limit instead of pretending it away. The division of labor is strict:

• AI does what AI is good at. It reads the part topology from your photo — outlines, holes, symmetry, feature structure — and proposes clean orthographic views with dimension placement that follows drafting conventions.
• You do the one thing only you can. Put a caliper on any feature and enter one measured value. That single reference propagates through the entire drawing and removes the scale ambiguity. Then confirm or correct the dimensions that carry your design intent.
• The output is standards-shaped from the start. Proper views, a complete title block, general tolerances per ISO 2768, callouts where a shop expects them — the full bar described in our manufacturing-ready drawing checklist.

That trade turns hours of drafting into minutes of verification. The human is not removed — the human is moved to the single step where judgment is irreplaceable. We think that is the only honest way to build this product, and not coincidentally the only way the resulting drawings are safe to manufacture from.

The bottom line

The research curve is steep: multi-image capture narrows the scale problem, and models like ReCAD improve geometric fidelity every few months. What will not change is what a manufacturing drawing is — a contract stating what you will accept and reject when the parts come back. Contracts get signed by people. Our job is to make the signing take five minutes instead of five hours.