Image to CAD: how to photograph a part so AI can model it

You do not need a scanner, a turntable or special lighting. You need a phone, a few minutes and a pair of calipers or a ruler. Here is how to get a good model on the first try with AI Reverse Engineering.

Taking the photos

1. Clear the background. Put the part on a plain, contrasting surface, such as a dark part on white paper. Clutter gets read as geometry.
2. Use even light. Daylight from a window or two lamps from either side. Avoid a single harsh light that hides edges in shadow.
3. Shoot square-on. Take the front, side and top straight on, then one three-quarter view that shows how the faces connect. One photo works; more angles resolve ambiguity.
4. Step back and zoom. Holding the phone close with the wide lens bends straight edges. Stand back a little and zoom so the part fills most of the frame.
5. Show the holes. Turn the part so holes, slots and pockets face the camera at least once. Features the camera never sees cannot be modelled.

Give it one real measurement

A photo has no scale: a 30 mm bracket and a 300 mm bracket look the same. So the AI reads proportions from the image, and you give it one real dimension. Measure the longest straight edge you can measure accurately, such as a flange length or an overall width, and type it in. Every other dimension scales from it.

Anything the AI inferred rather than read, such as a wall thickness hidden behind a lip, is flagged for review. If a dimension is critical (a bore that takes a bearing, a hole pattern that meets another part), measure it and set it directly.

Say what a photo cannot show

Add a line of notes with the part. The things that most often matter:

• Material and process: "aluminium, machined" or "injection-moulded ABS".
• Thread sizes: "M8 tapped holes", "1/4-20".
• Hole types: through, blind, counterbored or countersunk.
• Symmetry: "the left and right sides are identical".
• Anything internal: "hollow, 2 mm walls".

Check the model

1. Read the feature tree. It lists the part as features in plain words. If it saw a slot where there is a rib, correct it.
2. Look at the flagged dimensions. Check each one against the part.
3. Drag the sliders. Every dimension is a live parameter; move one and the part rebuilds.
4. Describe fixes in words. "The boss should be 2 mm taller and the fillet smaller" works.
5. Export. STEP for CAD, STL for printing, or a Fusion 360 script with the full feature tree.

What image to CAD is good at

Prismatic and turned parts: brackets, plates, flanges, spacers, housings, knobs, caps and clips, anything built from flat faces, cylinders, cones and fillets. These are most of the spares and fixtures a workshop needs.

It is weaker on freeform surfaces, internal features you cannot see, and tolerances tighter than a photo can show. For those, start from an STL scan instead (see STL to CAD), or measure the critical dimensions and set them.

A real run

The part at the top of the AI Reverse Engineering page is a cast part photographed once, on a desk, with an iPhone. It was rebuilt as a valid solid with 14 live parameters. That run was not anchored to a measurement, so its sizes are proportional; one caliper reading would have fixed the scale of all of them.

Going from a photo to a part you can change is the everyday form of vibe-CADing.