How plinth works

Walk around it

As you circle the object, plinth keeps a frame each time you have actually turned far enough and records the angle it was taken at — from the phone's 6DoF motion tracking where the browser offers it, otherwise from the compass. For each view it solves a metric point cloud: not a relative depth guess but real distances, so every view already agrees on size. Those views are fused into a single signed-distance surface — one skin, not a double-walled shell — at the true scale of the thing in front of you.

The measured angle is the whole point. Two photographs at an assumed ninety degrees put one unverifiable guess under the entire shape; two dozen at measured angles let the small errors cancel. Where a device reports no orientation the sweep still works, but it has to assume an even walk — and it says so on screen rather than pretending.

It tells you what it is sure of

Every build reports on itself. It tells you how many views it actually used, whether the dimensions are metric — a real measurement in centimetres — or only relative because no view carried a scale, how the angles were known (AR tracking, compass, or an assumed walk), and if there is a side of the object you never walked past. It will not invent a back it did not observe: an unseen region comes out as an open hole, not a confident guess.

A single photograph still works

Give it one image and plinth lifts it into a relief — a true 3D surface with real parallax and occlusion, reconstructed from a single angle with nothing recorded about the back. Turn far enough and you are looking at the back of a solved surface; that is the honest ceiling of one-image reconstruction. A silhouette engine is also on hand, carving a volume from a sweep's outlines — faster, but blind to concavities its outlines cannot reveal, and it says which it is.

Where the work happens

Everything runs inside your browser. Geometry is measured by MoGe‑2, the relief and carve by Depth Anything V2, and the subject is separated from its background by RMBG — all through Transformers.js, on your GPU via WebGPU where the browser supports it, on the CPU through WebAssembly where it does not. The models download once and cache; plinth picks the size your device can actually run and tells you how large it is before the first download.

Nothing is uploaded unless you press Publish. Capturing, measuring, adjusting and exporting all happen on your own machine, offline-capable once the models are cached. The renderer is served from this site and never calls out at all.

Taking it elsewhere

Export GLB — which carries the model's provenance (its unit, dimensions and how its angles were measured) baked into the file — or PLY and OBJ for Blender, a game engine or a slicer. From a single-photo relief you can also export the depth map as a 16-bit PNG, grab a PNG of the current view, or record a turntable as WebM.

Publishing

Publishing a scan uploads its geometry to the public gallery and gives you a permalink. That permalink shows the same honest facts you saw when you built it — the dimensions, how the angles were measured, how much of the walk it covered — and never a claim the build did not make. Only publish objects you are happy to have public.