Short answer: photogrammetry is measurement, generation is invention. That single distinction decides almost every case. If the thing exists, you can photograph it properly, and its real proportions matter, scan it. Otherwise generation is faster, cheaper and does not require the object.
| AI generation | Photogrammetry | |
|---|---|---|
| Requires the object | No | Yes |
| Input | A prompt or one photo | Dozens to hundreds of photos |
| Dimensional accuracy | Not guaranteed | High, with a scale reference |
| Capture time | None | Minutes to hours |
| Processing time | Seconds to minutes | Minutes to hours |
| Handles shiny or transparent | Poorly | Very poorly |
| Handles things that do not exist | Yes | No |
| Output topology | Dense reconstruction | Dense reconstruction |
The one thing they share
Both produce a reconstruction, not a model: dense, irregular triangles with no edge flow and a UV layout that exists only to hold the texture that was baked into it. Neither gives you production topology, and both need the same downstream work — retopology, then unwrapping, before the asset is genuinely usable.
People often expect scanning to avoid that. It does not.
Where photogrammetry wins outright
- The object exists and you can photograph it from every side. That is the whole precondition.
- Real proportions matter. With a scale reference in the scene, a scan gives you measurements. A generation gives you a shape.
- It has to be that exact object. A specific prop, a specific building, a heritage artefact — the case where "something like it" is not the assignment.
- Surface detail is the point. Weathering, wear, real material variation. A scan records what is there; a generator produces something plausible instead.
Where generation wins outright
- The object does not exist. Concept work, fictional props, anything not yet manufactured.
- You cannot get to it. It is in another country, or in a photograph, or it is a drawing.
- You need many. Capture does not amortise; a hundred scans is a hundred capture sessions.
- Speed matters more than fidelity. Blocking out a scene, filling a background.
Where both struggle, for the same reason
Shiny, transparent and featureless surfaces. Photogrammetry needs matched features across images to triangulate a point, so glass, chrome and a plain white wall give it nothing to match — the standard workaround is a dulling spray, which is not always allowed. Generation fails on the same subjects for a related reason: a reflective surface shows the room rather than its own shape, and the reconstruction follows what it sees.
Thin structures. Wires, railings, foliage, hair. Neither technique resolves them well.
Picking, in one line each
- The thing is in front of you, and the dimensions matter → scan it.
- The thing is in front of you, and only the look matters → photograph it once and use [image to 3D](/kb/generate-3d/image-to-3d/), which is far less work than a capture session.
- The thing does not exist → generate it.
- You need forty of them → generate them.
Further reading
The Khronos glTF specification covers the format both routes usually export to, and Blender's decimate and remesh documentation describes the cleanup tooling most people reach for first — and why it produces triangle soup rather than usable topology.