Photogrammetry vs AI image to 3D: which should you use?

Two ways to get a model from photos

Both start with photos and end with a 3D mesh, which is why they get confused. Underneath, they're doing opposite things.

Photogrammetry measures. You walk around a real object taking many overlapping photos. Software finds matching points between photos, triangulates where they are in space, and builds a mesh from them. Tools include Meshroom (open source), RealityCapture, Agisoft Metashape and phone apps such as Polycam.

AI image-to-3D infers. A trained model looks at one image (or a few) and generates the most plausible 3D object that would look like that. It's drawing on everything it learned from training data, not measuring your object. Vi3W's image-to-3D is this kind of tool.

Side by side

• Input: photogrammetry needs dozens to hundreds of photos of a real object; AI needs one or a few images of anything, including drawings and renders.
• Accuracy: photogrammetry is close to the real object's shape when captured well; AI is plausible but unseen areas are guesses.
• Real dimensions: photogrammetry can be scaled to real size with a reference; AI output has no reliable real-world scale.
• Needs the physical object: photogrammetry yes; AI no.
• Works from concept art: photogrammetry no; AI yes.
• Capture effort: photogrammetry high (lighting, coverage, processing); AI low.
• Mesh quality: photogrammetry gives very dense, scan-like meshes that need retopology; AI varies by tool (Vi3W retopologises automatically).
• Textures: photogrammetry captures real colour, often with baked lighting; AI generates textures that can also contain baked shading.
• Hard cases: photogrammetry struggles with shiny, transparent and featureless surfaces; AI struggles with thin parts, glass and the hidden back.

When to use photogrammetry

• You need a faithful copy of a specific real object: a museum piece, a product for a catalogue, a heritage site.
• Accuracy matters more than speed.
• You have the object, time and decent light.
• You're capturing environments or large surfaces like rock faces and building facades.

When to use AI image-to-3D

• You don't have the object: it's concept art, a sketch, a render or an idea.
• You need many props quickly, and "plausible" is enough.
• You have one good photo and no chance to take more.
• You want a clean, light mesh without scan clean-up. (Photogrammetry meshes typically need heavy retopology to be game-ready.)
• You're stylising: AI can reinterpret a photo in a style; photogrammetry captures what's there.

Combining them

• Scan the hero objects with photogrammetry; generate the background props.
• Use a scan as the reference shape and a generated model for a stylised version.
• Use AI to fill a missing side when a scan has gaps (check carefully).

A note on Gaussian splatting

A third approach, Gaussian splatting, also reconstructs from many photos, but produces a cloud of "splats" rather than a mesh. It's excellent for viewing captured scenes and harder to use as game geometry. Gaussian splatting vs meshes explains the trade-offs.

Choosing quickly

Ask two questions:

1. Do I have the real object and need it accurate? Photogrammetry.
2. Do I need a usable model from one image, or of something that doesn't exist? AI image-to-3D.

Vi3W is upfront that it isn't a 3D scanner: on its image-to-3D page it says that for a measured, true-to-life copy you should use a scanning app.

Have one good image rather than a full capture? Turn it into a 3D model in Vi3W.