Product visualisation

One photograph to an AR asset

No CAD file, no scanning rig, no studio day. One image of a chair, reconstructed into geometry, finished into a product asset, and placed in a real room at true size. Every step ran on one desktop machine.

A cream boucle barrel chair photographed three-quarter view on a plain grey backdrop

The photograph

One object, three-quarter view, plain seamless, even light. Reconstruction quality is decided here more than anywhere downstream.

1024 x 1024 · SDXL Turbo, local

The raw reconstruction rendered untextured, showing the chair and a slab of studio floor beneath it

The reconstruction

Shape only, and it cannot tell the product from the set: the seamless and the contact shadow came back as a slab of geometry under the chair.

500,580 triangles · Hunyuan3D 2.0, local

The finished chair asset, upholstered in camel boucle, floor slab removed

The product asset

Floor cut away, simplified, scaled to a real chair, box-projected UVs, the catalogue bouclé applied with its sheen lobe, MikkTSpace tangents baked.

60,572 triangles · 70 x 78 x 62 cm

The reconstructed chair placed in a room at true size

In the room

glTF for Android and the web, a matching USDZ for iOS AR Quick Look, both at metric scale with resizing locked off so the answer to "will it fit" is the true one.

4.93 MB glTF · 3.44 MB USDZ

Loading the piece

The source photograph, for comparison against the 3D result

The photograph it came from

Drag to turn it. On a phone, "View in your room" places it on your floor at 78 cm tall.

What this technique does well, and where it stops

Single-image reconstruction is very good at silhouette and proportion, and that is most of what a shopper is judging when they ask whether a piece suits the room. It is not a measuring tool and it is not a scanner.

The model returns shape, not colour. The upholstery here is the catalogue's own Camel Bouclé applied afterwards, which is the honest arrangement anyway: the mill owns the fabric, not the reconstruction.
The back of the chair is inferred, not observed. One view cannot see it. For anything where the reverse matters, shoot two or three views and use the multi-view model instead.
Marching cubes leaves faint stair-stepping on surfaces that should be perfectly smooth, and the legs merge into the base where the original photo put them in shadow.
Dimensions come from a stated real-world height, not from the photograph. The chair is right because a lounge chair is about 78 cm tall, not because anything was measured. Trade drawings still win for a spec sheet.
Not yet run on a physical handset. The USDZ passes OpenUSD's validators and is served with the correct content type, which is as far as a desktop can take it.

Where it earns its keep is catalogue depth. A furniture house with four hundred photographed SKUs and no 3D budget can have most of them placeable in a room, at roughly a minute of compute each, without a photographer or a modelling contract.