These two things both turn photographs into 3D models and they are not substitutes for one another. They fail differently, they cost differently, and choosing the wrong one wastes either an afternoon or a week.
The distinction is simpler than the marketing on either side makes it sound.
What each one is actually doing
Photogrammetry measures. You take fifty to two hundred overlapping photographs of a real object from every angle. Software finds features that appear in several images, works backwards to where the camera must have been, and triangulates a point for each feature. The result is a reconstruction of surfaces that were genuinely photographed. Every millimetre of it corresponds to something that was really there.
Single-image AI generation infers. You give it one photograph. A model that has seen an enormous number of objects produces a plausible three-dimensional shape consistent with that view. The front is derived from what it can see. The back is invented. Competently, from everything similar it has learned, but invented.
That is the whole difference, and everything below follows from it.
What follows from it
Accuracy. Photogrammetry gives you real dimensions, and with a scale reference in the shot it gives you real units. Generation gives you correct proportions at an arbitrary size, because a single photograph contains no scale information at all. If you need to 3D print a replacement part that has to fit something, that is a photogrammetry job. If you need a crate for the background of a level, it is not.
The unseen side. Photogrammetry only reconstructs what you photographed; miss the underside and there is a hole there. Generation always gives you a closed object, but the parts it could not see are a guess. Neither is better. They are opposite failure modes, and which one you can tolerate depends on whether anyone will look at the back.
Time and effort. A photogrammetry capture is a careful half hour with a camera, followed by anywhere from twenty minutes to several hours of processing, and it can still fail. A generation is one upload and a minute or two. The gap in effort is roughly two orders of magnitude.
Access to the object. Photogrammetry requires the thing to exist and to be somewhere you can walk around it. Generation works from a photograph someone else took, a product shot, a concept sketch, or a description. If you do not have the object in the room, there is no choice to make.
What they both hate. Glass, chrome, water, and anything mirror-finished defeat both, for the same reason: what the camera records is the room, not the surface. Photogrammetry also struggles with featureless matte surfaces, because there is nothing to match between frames. Generation handles a blank white mug without difficulty.
The decision, as a short list
Pick photogrammetry when the geometry has to be true: a replacement part, a site or artefact record, anything where a measurement will be taken off the model, restoration, survey work.
Pick generation when the geometry has to be convincing: props and set dressing, background assets, concept and previsualisation, a first pass to sculpt over, product visualisation from an existing photograph, anything where you need forty objects this week rather than one object correctly.
Pick both when there is a hero asset and a hundred supporting ones. That is the case more often than either camp admits. Photogrammetry the thing the camera lingers on; generate the rest.
The part people get wrong
The mistake is treating generation as cheap photogrammetry. It is not a cheaper way to get an accurate model. It is a different product. A generated mesh is a plausible object, and asking whether it is accurate is close to a category error, because there is no ground truth for the half of it the photograph did not contain.
Once you stop expecting measurement from it, it becomes obvious what it is good for, and that turns out to be most of the 3D work in most projects. The number of assets in a game or a visualisation that require dimensional accuracy is small. The number that require looking right and existing by Friday is not.
One practical note on topology
Both produce triangulated, dense, unevenly distributed meshes that are built to be looked at rather than worked on. Neither gives you edge loops. If the asset has to deform at a joint, you will be remeshing or retopologising either way, so that is not a point of difference. It is a shared tax on reconstruction of any kind.
The difference is only ever in whether the surface is measured or inferred, and whether your job needs the measurement.