Reference guide
Scan prototype
An experiment that turns phone photos of a product into a model with a real, flat label.
What it proved
From photos alone, the prototype recovered 96% of the words on a label (179 of 187) as flat artwork. That is the piece that matters: the packaging suite supplies the shape, and this method supplies the label from photos.
It lives in prototypes/scan-m0/ on the scan-prototype-m0 branch. It was tested on a synthetic capture of a creatine jar rendered in Blender, 50 photos and 150 walkaround frames, where the true answer was known.
What works
| Step | Method | Result |
|---|---|---|
| Camera positions | Apple's photogrammetry session, then edge refinement | Works. COLMAP failed: it folded the cylindrical jar |
| Shape | Silhouette hull from cut-outs, then a measured profile | 95% of the surface within 0.8 mm of truth |
| Flat label | Unroll on the measured body, blend, align each photo with optical flow | 96% of words recovered |
How it compares
Text reproduced on views held out from the test:
| Approach | Text reproduced |
|---|---|
| Apple scan | 67 to 69% |
| Our texture on Apple's mesh | 82% |
| Our texture on a Meshy mesh | 77% |
Three 3D generators were also tried (Tripo, Meshy, Hunyuan). All of them invent small print. Photogrammetry and generators both emboss label ink into the geometry.
Not done
- Placing the flat label onto a modelled body.
- A full-resolution re-render, paused part way.
run.shis partly out of date.
Why it is parked
The packaging suite replaces scanning as the shape source for common packaging. The flat-label method is the part worth carrying forward, and it is a natural tool for the concierge service. See Pricing and concierge.
Updated 2026-11-14. Built 2026-11-15.