How-to · Prompts
Photo to 3D: AI Generation vs Photogrammetry and Scanning
AI image-to-3D invents what the camera did not see; photogrammetry measures what it did. How the two differ, what documented free options exist, and how to choose.
If the object exists and you can photograph it from all sides, photogrammetry reconstructs the real shape from your photos. If you have one image, a drawing or an idea, AI image-to-3D produces a plausible shape that includes surfaces the photo never showed. They answer different questions: photogrammetry asks “what is this thing”, generation asks “what could this look like in 3D”.
How they differ in principle
Photogrammetry software takes many overlapping photos of one object and reconstructs geometry from what several photos jointly show. Missing coverage becomes holes or noise. An AI generator learns shapes from training data and fills in unseen parts. A 3D AI Studio guide makes the same point from the vendor side: single photos mainly fail on hidden back and side surfaces (our paraphrase of a vendor-authored, undated page), which is why it recommends multiple angles. This distinction is general knowledge about the methods, not a result we measured, and no source we read compares the two on the same object.
| Question | Photogrammetry | AI image-to-3D |
|---|---|---|
| What the input is | Many overlapping photos of a real object | One to a few images, or a sketch |
| Unseen surfaces | Not captured; appear as gaps | Invented by the model |
| Fidelity to a real object | The goal | Not guaranteed |
| Works from concept art | No | Yes |
| Typical use | Real props, surfaces, scenes | Invented props and characters |
Documented options, with limits
RealityScan (Epic Games). Epic’s page says it creates models “from images and laser scans,” installs through the Epic Games launcher on desktop, and has a free mobile app. It is free for individuals and small businesses with under $1 million USD annual gross revenue, and for educators; above that a yearly seat is required, with no price on the page. It gives no hardware requirements on that page.
Meshroom (AliceVision). An open-source node-based photogrammetry framework under the MPL-2.0 license per its repository. The page we read does not state GPU requirements or give a release, so check the project’s install notes.
AI image-to-3D. Meshy’s multi-view tutorial accepts one to four photos of an object, with a plain background and even lighting, on Pro and above and Meshy 7 only. 3D AI Studio’s guide recommends three to eight photos. Rodin accepts one to five. These are input limits, not accuracy claims. See preparing images for image-to-3D.
Shooting for photogrammetry: the documented capture rules
The Meshroom manual’s capture page (accessed 2026-11-08) is the most specific official guidance we read for photogrammetry input. Its rules, in procedure form:
- Light. The page asks for a well-lit scene in diffuse or indirect light, such as an overcast day outdoors. Avoid shadows, reflections and transparent objects; the page lists these as things to avoid.
- Camera. Do not use flash. Keep the focal length fixed (no zooming; a fixed-focal-length lens if possible). Avoid shaky, blurry or warped images.
- Overlap. The page asks for “a side overlap of 60% minimum” and “frontal overlap of 80% minimum.”
- Positions. Move to a new position for every shot, or rotate the object; do not take several images from the same spot. Shoot from all angles, and the page suggests re-photographing an area with a different acquisition pattern for better coverage.
- Framing. The object should fill most of the image.
- Background. Avoid moving objects in the scene. For a rotating rig, use a plain single-color background without distinguishable features.
- How many. The page gives no count; it says more images are better and poor or repetitive ones can be filtered out later to reduce processing time.
It gives no guidance for textureless surfaces and does not say which camera or GPU to use. Compare the AI-side rules in preparing images for image-to-3D: Meshy asks for a plain background and one to four views, the opposite end of the effort scale.
Troubleshooting the choice
Causes follow the two methods’ stated rules; none was reproduced on a specific object.
| Symptom | Likely cause | Fix |
|---|---|---|
| Photogrammetry output has holes | Parts of the object were not covered by overlapping photos | Re-shoot with new positions and angles meeting the overlap guidance |
| Photogrammetry fails on a glossy or transparent object | Meshroom’s capture page lists reflections and transparency as things to avoid | Change the object’s surface or use the AI route and accept an invented shape |
| AI model has a plausible but wrong back | Unseen surfaces are invented | Supply more views within the tool’s count limit (Meshy 1 to 4, Rodin 1 to 5) |
| AI model is restyled | Meshy API image_enhancement defaults to true |
Set it false where the model supports it |
| Neither is the right size for the engine | Scale is not set by the photos | See scale and pivot fixes |
What the output needs next
Both routes produce meshes that need checking. Photogrammetry outputs are commonly dense and need decimation and cleanup; AI outputs often need retopology and UV checks. Neither guarantees a game-ready mesh, and neither source set gives polycounts. For printing a figurine from photos, see photo to figurine.
Rights and subject matter
Whose photo, whose object and whose likeness matter more here than the tool. Photographing your own object is a different position from scanning a branded or copyrighted item or a person; this is not legal advice and the sources we read do not cover it, so ask if the subject is not yours. For AI services, plan terms apply: Meshy’s Free output is CC BY 4.0, with Meshy’s Terms (updated September 19, 2026) saying Meshy owns Free users’ output. See commercial use of AI 3D models.
How to choose
- You have the physical object and need fidelity: photogrammetry first.
- You have only an image, a drawing or an idea: AI generation.
- You need real surface scans (a wall, a rock): photogrammetry.
- You need many invented props quickly: AI, with a cleanup budget.
- Both: use a scan as a reference, or feed multi-angle photos to a generator that accepts them, and compare the two meshes against the object.
To pick a generator for photo input, see the image-to-3D shortlist. Our planned same-object comparison has not been run, so this page gives a way to decide, not a ranking of results.
See also the AI 3D generation guides hub.