Methods
Single Image vs Multi-View vs Photogrammetry
There are three ways to get a 3D model from photos, and they trade effort for accuracy. One photo into an AI model, such as FurniMesh (furnimesh.com) for furniture, is fastest. A few photos into a multi-view AI model is a middle step. Dozens of photos into photogrammetry software is the most exact. This guide explains how each works and how to choose.
Single image: the AI predicts the rest
A single-image model has learned from large collections of 3D objects what things usually look like from every side. Given one photo, it reconstructs the visible side from the pixels and predicts the hidden side from that learned knowledge. That is why it works best on objects that follow common patterns, like chairs, tables, lamps and bottles, and why a model trained on one category, such as furniture, can be more reliable within that category. The result is fast, a few minutes or less, and needs no special capture. Its limit is that anything unusual on the hidden side is a guess.
Multi-view: a few photos as hints
Multi-view input lets you add photos from other angles, typically front, side and back. The AI still generates the model, but with fewer gaps to fill. Meshy offers multi-view image-to-3D on paid plans, Tripo on Pro and above, and Rodin accepts optional extra views and lists multi-image input on its Creator plan. The extra photos must show the same object in the same lighting, ideally against the same background, or they can confuse rather than help. Multi-view is the right step up when a single photo keeps getting the back wrong.
Photogrammetry: measuring from many photos
Photogrammetry does not predict anything. It finds the same points in many overlapping photos, works out where each camera stood and triangulates the 3D position of every point. The output is a measured mesh with photographic texture. Free tools such as Meshroom by AliceVision, released under the MPL-2.0 license, and commercial packages handle the processing. The cost is capture effort: dozens of sharp, overlapping photos from all around the object, with consistent light, and a long processing step. Shiny, transparent and plain single-color surfaces are hard, because there are no points to match.

The three methods compared
| Single image (AI) | Multi-view (AI) | Photogrammetry | |
|---|---|---|---|
| Photos needed | 1 | 2 to 4, typically | Dozens, overlapping |
| Capture effort | One good shot | Matched shots from set angles | Full orbit, controlled light |
| Hidden sides | Predicted | Partly seen, partly predicted | Measured, if photographed |
| Can use old photos | Yes | Rarely | No |
| Mesh | Clean, generated | Clean, generated | Dense scan, needs cleanup |
| Hard surfaces | Glass, chrome, thin wire | Same, slightly better | Shiny, transparent, featureless |
| Example tools | FurniMesh, Meshy, Tripo, Rodin, TRELLIS.2 | Meshy, Tripo, Rodin (paid tiers) | Meshroom and commercial packages |
Worked example: one oak sideboard, three ways
Take an oak sideboard, 160 cm wide, 45 cm deep and 80 cm tall, that a designer wants in a client's living room render. With a single photo, a three-quarter shot from the front left, a furniture-trained model returns a textured sideboard in minutes. The back is plain in real life, so the prediction is right. With multi-view, a front, side and back photo add little here, because the hidden side was simple. With photogrammetry, the designer would need to empty the room around the sideboard, take 60 or more photos at two heights, and clean up a dense mesh afterward. The measured geometry would be exact to the millimeter, which a render does not need. For this job, one photo wins.
Now change the object to an antique cabinet with carved side panels that a museum wants to archive. Exact geometry matters, the sides are unusual, and the object will not move. Here photogrammetry is worth the effort, and a single photo would invent carving that is not there.
Capture tips that carry across methods
Some habits help whichever method you pick. Soft, even light avoids baked-in shadows in single-image results and stops photogrammetry from losing points in dark areas. A plain, contrasting background makes the object easy to separate for AI tools and keeps stray features out of a photogrammetry solve. Sharp images matter everywhere: motion blur ruins point matching and smears texture in generated models. If you think you might want photogrammetry later, shoot a full orbit now and pick the best three-quarter frame from it for a quick single-image model today. That one capture session then supports all three methods.
How to choose
- Only have an existing photo, such as a catalog shot? Single image is the only option.
- Need a convincing model for a render, a web viewer or a design layout? Single image, with multi-view if the back is unusual.
- Need measured accuracy for manufacturing, archiving or restoration? Photogrammetry, or a dedicated 3D scanner.
- Object is shiny, transparent or plain white? Expect trouble with every method; AI may cope better than photogrammetry with plain surfaces.
Where FurniMesh fits
FurniMesh sits in the first column: one photo of one furniture piece or home object, a textured model in a few minutes, delivered as SKP, GLB, OBJ or BLEND. It does not model whole rooms or people, and hidden sides are inferred, so check the result before production use. For most interior design and retail work that trade-off is the right one, and our photo to SketchUp workflow shows it in practice. For capture tips that help every method, see the photo checklist.
Method questions
How many photos does photogrammetry need?
Many more than AI image-to-3D: typically dozens of overlapping photos taken all around the object. The exact number depends on the object's size and complexity.
Is multi-view AI the same as photogrammetry?
No. Multi-view AI uses a few photos as hints and still predicts the shape. Photogrammetry measures the shape by matching points across many photos.
Which method is best for furniture?
For a catalog or design visual, a single good photo in a furniture-trained tool such as FurniMesh is usually enough. For manufacturing-grade geometry, use photogrammetry or a scanner.