Workflow · 7 min read

Image-to-CAD workflow

The AI CAD agent is vision-capable: drop in a hand sketch, a photo of an existing part, or a screenshot, and it will read it, ask clarifying questions, and write parametric OpenSCAD that matches. After every generation it sends a 4-view technical sheet of its render back to itself for a self-critique pass — so revisions converge against your reference without you having to compare side-by-side.

What kinds of images work?

SourceWorks well forTips
Hand sketch on graph paper Mechanical brackets, mounts, plates, jigs. Label key dimensions in red. Note the screw size. Indicate the up direction.
Multi-view photo of an existing part Recreating parts you can hold (knobs, brackets, clips). Shoot front + side + top with a ruler in frame. Even lighting; no shadows on the part.
Screenshot of a mating surface PCB outlines, motor spec sheets, extrusion cross-sections. Crop tight. Include the dimension annotations from the source.
3D render or product photo Stylistic / aesthetic intent (fluted, hex, ribbed). State which features you want copied vs ignored.
Marketing PNG of an existing model Re-creating a printable knock-off as a starting point. Tell the agent it's reference-only, not the final design.

Step-by-step

  1. Open the AI CAD Modeler.
  2. Drag your image into the chat composer (or click the paperclip icon).
  3. Add a one-sentence prompt that names the part and the unit: "M3 corner bracket like the sketch. All dimensions in mm."
  4. Send. The agent generates a base part and renders it.
  5. An auto-iterate pass runs (up to 2 rounds): the agent sends a 4-view technical sheet of the render back to itself and refines if the result doesn't match the image.
  6. Review the brief + 3D preview. Use plain English to refine specific features by name — see Iteration prompts.

How auto-iterate works under the hood

After the first render, the agent's render engine produces a single image composed of four orthographic views (front, side, top, isometric). That sheet is appended to the agent's next context, alongside your original reference, with the question "does the render match the reference? If not, what specifically should change?".

The agent will either:

  • Accept the render and stop iterating (no credits charged for the extra pass).
  • Propose a targeted edit ("the back rib in the sketch is twice as tall — increasing rib height") and re-generate.

The loop runs up to twice by default — long enough to converge, short enough not to burn credits.

Sketch best practices

Sketches are the single highest-ROI input. A few habits make them dramatically more useful:

  • Use graph paper. Even if you don't dimension everything, the grid gives the agent a visual unit.
  • Dimension in red ink. The contrast helps the vision model lock onto numbers vs structural lines.
  • Always label the screw size or hole spec. "M3", "Ø3.2", "1/4-20" — saves a clarifying turn.
  • Note the print orientation. Arrows or a "UP" label tell the agent which face must stay flat to the build plate.
  • Sketch the mating part lightly. Outline the V-slot, PCB, or bearing the part has to fit. The agent will use it for clearances.
  • Cross out what you don't want. If you sketched an option and rejected it, X it out — the agent won't try to interpret it.

Photo best practices

  • Plain background. White paper or a single-color cloth. Removes vision noise.
  • Three angles. Front, side, top (or front, side, isometric). One angle isn't enough to recover depth.
  • Ruler in frame. A small metric ruler in every shot gives the agent absolute scale.
  • Even lighting. Soft, frontal light. Shadows can confuse the model into thinking a feature exists where none does.
  • Crop tight. Bigger pixels per feature = better extraction.
  • State material in the prompt. Steel, ABS, PLA, brass — affects how the agent interprets reflections/highlights.

Screenshots of CAD or spec sheets

For mating to standardized hardware, screenshots of the spec sheet (motor bolt circle, PCB outline, extrusion cross-section) work brilliantly. The agent can read the dimension callouts directly and clearance-fit holes without you having to type the numbers.

Example prompt with a NEMA17 spec-sheet screenshot:

NEMA17 motor mount, fits the spec sheet attached.
50 × 50 mm × 4 mm plate, M3 clearance bolts on
the 31 mm bolt circle, 22 mm pilot bore for the
boss. Symmetric.

Combining text + image + sketch

The composer accepts mixed input. Common stacks:

  • Text + sketch — sketch carries the shape, text carries hardware specs and material.
  • Text + photo + sketch — photo for "what it looks like", sketch for "what I want to change", text for constraints.
  • Text + 2 screenshots — one for the mating part, one for the look you're after.

Known limits

Vision models aren't measuring tapes.

The agent can recognise that a hole is roughly central, that a chamfer exists, and that there are four mounting holes — but the exact millimetre values are far more reliable when you put them in the prompt (or write them on the sketch). Treat the image as the shape, your text as the specification.

  • Heavy perspective distortion (close-up wide-angle shots) lowers accuracy. Step back, zoom in.
  • Reflective metal parts photographed under point-source lighting can have phantom edges. Diffuse the lighting.
  • Internal cavities you can't see from outside have to be described in text — the agent can't X-ray a part.

For pure mesh-style image-to-3D (figurines, props, characters), use one of our purpose-built generators instead — they output STL meshes, not parametric code:

Image to 3D model

Mesh-style AI for figurines, props, art, and decor. Outputs STL.

Face to 3D

Photo of a face → printable 3D bust.

Pet to 3D

Pet photo → printable 3D figurine.

Try a sketch right now

Open the modeler, drag any image into the composer, and add a one-line prompt. Auto-iterate handles the convergence.

Open the agent