Photogrammetry has existed for years and almost nobody used it, because it wanted fifty photographs, a turntable, controlled light and an afternoon of processing. Single-image reconstruction is worse at accuracy and vastly better at getting used.
Shoot it properly and most problems disappear
- One object, filling the frame. The single most common mistake is standing too far back. If the object occupies a quarter of the picture, three quarters of the reconstruction is guessing about the background.
- Three-quarter angle. Two faces visible gives the model real information about depth. A flat, straight-on shot gives it a silhouette.
- Even light. Hard directional light bakes shadows into the texture, and baked shadows follow the object into every scene you later place it in.
- Plain background. A busy background competes for the model's attention and sometimes joins the object.
- In focus, no motion blur. Detail the camera missed is detail the model invents.
Real-world size, done carefully
The raw output is normalised, so it arrives without a size. To give it one, type the object's longest real side in millimetres. The mesh's bounding box is measured from its own geometry and one uniform scale factor is applied so that longest axis matches your number.
It is deliberately never per-axis. Scaling width, depth and height each to their own measurement would quietly redesign the object — you would get something with the right dimensions and the wrong proportions, which is worse than an unscaled mesh because it looks correct.
The honest limits
- It is a reconstruction, not a survey. The back was never photographed, so hidden faces are the model's best guess.
- Objects, not buildings. A house or a room needs a modeller, not a mesh reconstructor.
- There is no revision. No text box edits a finished mesh. A different result comes from a different photograph — which means a second attempt, at full price.
- AR is Android-only. iPhone AR needs USDZ, which is not exported, and the interface says so rather than failing quietly.
What it is genuinely good for
Getting a specific real object into a scene. A client's existing sofa. A tap the spec is built around. A piece of facade detail you want to show at scale. A product that needs to sit in a render. Those are all jobs where modelling by hand costs more than the object is worth, and where a stock model would be the wrong object.
What it costs
15 credits ($1.31) per object, usually a couple of minutes. The GLB download, the 3D viewer, the share link and Android AR are all free. Note that a new account's 14 free credits does not quite cover one run, so the smallest top-up is the practical starting point.
Common questions
How many photos do I need?
One — which is the appeal and the limitation. The back of the object is inferred rather than observed.
What objects work best?
Furniture, fixtures, products and facade details: one object, clearly lit, filling the frame at a three-quarter angle.
Can I scale it to real-world size?
Yes — give the longest real side in millimetres and one uniform factor is applied. Never per-axis.
What file do I get, and can I put it in a room in AR?
A textured GLB. Room-scale AR works on Android; iPhone needs USDZ, which is not exported.
See the full workflow and price: Image to 3D model.