Photogrammetry builds a 3D model by triangulating features across many overlapping photographs of a real object.
How it works
Take dozens to hundreds of photos covering the subject from all angles with substantial overlap. Software identifies matching features across images, solves for camera positions, builds a dense point cloud, and converts it to a mesh with a projected texture.
What it is good at
- Fidelity to a real object. It measures rather than infers, so proportions and surface detail are genuinely accurate.
- Realistic surface texture. The colour comes from photographs of the real thing.
- Scanning environments and terrain, including from drone footage.
What it struggles with
- Featureless surfaces. Flat white walls have nothing to match between images.
- Reflective and transparent materials. Both violate the assumption that a surface point looks the same from different angles.
- Thin structures. Wires, leaves, fences.
- Moving subjects, unless you have a synchronised multi-camera rig.
- Baked lighting. The captured texture includes the lighting it was shot under, which has to be de-lit for use under different lighting.
Against AI reconstruction
- Photogrammetry needs many photos, real access to the object, and processing time — and gives measurable accuracy.
- Single-image generation needs one picture and gives a plausible interpretation.
If the object exists, you can photograph it thoroughly, and accuracy matters, photogrammetry wins. If you have one photo, or the object does not exist, it is not an option at all.
The shared aftermath
Both produce dense meshes with machine-made topology and projected textures. Retopology, unwrapping and baking apply identically. Photogrammetry meshes are typically far denser, so the reduction is more aggressive.