Take one photo of an oak armchair on your phone, against a white wall, and upload it to an image-to-3D service. Three minutes later you’ll have a mesh of 38,000 triangles with the wood grain from your photo wrapped around it. Your chair was never in anyone’s training set. Nobody scanned it. The software has never seen its back. And yet there’s the back, plain and perfectly plausible, which is either impressive or faintly suspicious depending on what time of night you’re looking at it.
It’s both. Four things happened in those three minutes, two of them learned from data and two computed with maths that predates the web, and once you know which is which you’ll know exactly what the result is good for (a room plan, a listing, a print) and what it isn’t (a measured drawing, ever).
Stage one: how far away is every pixel?
A photo is flat. So the first job, before anything else, is to work out how far away each pixel is, and the model that does this uses the same cues you use without ever thinking about them. The seat edges narrow toward the back? That’s perspective. The armrests sit at armrest height? That’s the model knowing roughly how big an armrest is. The seat’s brighter than the underside, the front legs hide the back legs, the grain gets tighter toward the back rail. Every one of those is a depth cue, and none of them was programmed. The depth model soaked them up from millions of photo-and-depth-map pairs, and on indoor objects it’s wrong by a few percent, which is nothing on a chair against a wall.
Put the same chair in front of a bookshelf, though, and the shelf edges turn into depth cues of their own. The chair grows a back it never had. This is the first of several places where your photo matters more than their software, and it won’t be the last.
Stage two: the sides the camera never saw
Your photo shows the front and one side. The back and the far side have to be invented, and five years ago that was science fiction. Zero-1-to-3, a paper from 2023, did it with a diffusion model: one image in, a plausible view from a new angle out, slowly. The big reconstruction models of 2024 (TripoSR is the one people have heard of) squeezed that into a single pass, under a second on a data-centre GPU. The multi-view models that came after produce four to six consistent views at once and stitch them.
Whichever one ran on your armchair, the back it produced is the most likely back given the front and everything the model saw in training. That’s a careful way of saying it’s a guess. A good one, because most armchairs with that front have that back. But a carved back panel would also come out plain, since nothing in a front view says “carving”, and this is the one thing worth remembering from the whole article. The hidden side is inferred. Anyone selling it to you as a scan is selling something they haven’t got.
MORE NEWS: Honeywell is building new $500 million manufacturing facility in Tempe
DEEPER DIVE: 10 Arizona restaurants make Yelp’s Top 100 Mexican food spots
Stages three and four: triangles, then the grain goes back on
What you have after stage two is a fuzzy idea of where the chair is and isn’t, and Blender and SketchUp can’t open a fuzzy idea. The last two stages are the unglamorous ones, and they haven’t changed much since the 1980s.
| Stage | What happens | On the armchair |
| Three, meshing | Marching cubes, published 1987, walks a grid, finds the cells the surface passes through and drops a triangle in each | 38,000 triangles, watertight |
| Four, unwrapping | Every triangle gets a home on a flat image | 2,048 by 2,048 px |
| Four, projecting | The photo lands on the triangles that faced the camera | Real grain on the front and one side |
| Four, filling | Triangles that faced away borrow colour from neighbours and from the invented views | Repeating grain on the back |
| Four, baking | One texture file saved beside the mesh | 1.9 MB as a GLB |
Vendors argue about what sits underneath stage three (full-and-empty cells, signed distance fields, radiance fields, Gaussian splats, each has its partisans) and it makes no difference to the triangles that come out. Watertight is worth a word. It means no holes, and a mesh with a hole in it has no inside, so if the chair is heading for a printer or a tape measure the word matters and if it’s heading for a room plan it doesn’t. Spin the finished chair slowly and the seam is there, photographed grain on one side and invented grain on the other, and having seen it once you’ll see it on every model forever.
Why your chair came out well and a wire lamp wouldn’t
Chairs reconstruct well for a boring reason: there are a lot of chairs in the training data. ShapeNetCore, from 2015, is about 51,000 clean CAD models in 55 categories, and chairs and tables are two of the biggest. CO3D, Meta’s 2021 set, is about 19,000 phone videos walked around household objects. Objaverse scraped over 800,000 objects from public repositories in 2022 and its XL version passed 10 million a year later. That’s where the recent models learned what things look like, and the coverage is lumpy in ways you’ll see in your results. Household objects? Well covered. Musical instruments, and anything made of thin wire? Badly. A wire lampshade, uploaded the same afternoon as the armchair out of curiosity, came back with gaps wherever two wires crossed, which is the grid being too coarse for the wire and not a bug anyone will fix.
The rest of the failure list is the same story told five ways, a depth cue misreading something.
| What you upload | What comes back | Why |
| Glass or chrome | Holes, warped surfaces | Reflections |
| Two objects | One merged blob | Trained one object at a time |
| A photo from above | Squashed height | Perspective cue |
| A hard shadow across the object | A dent where the shadow was | Shading cue |
| A whole room | Nonsense | Object models, not scene models |
Every photo rule on every service’s upload page is just this list, rephrased politely.
The file you get
The armchair left as a GLB and as an SKP, and it could have left in four other formats, all of them older and simpler than the model that produced it, which is a pleasant irony.
A furniture-only service such as FurniMesh runs the same four stages on one photo of a chair, a lamp or a cabinet and exports GLB, OBJ, STL, FBX and SKP from one mesh. The general services do the same stages on anything from a trainer to a spaceship. What separates them is the training data behind stage two and the export menu at the end, and knowing that much makes you a far more awkward customer, in a good way.
Questions for the vendor
The papers score these systems with Chamfer distance, F-score and volume IoU, none of which you can check and none of which you need. Rotate the preview, look at the back, and then ask the questions that sort the serious vendors from the rest: is the mesh watertight? Can the triangle count be brought down for a web viewer? How big is the texture? What units does the export use, or is there a scale step on import? Which formats? Is my photo deleted if I ask? A vendor who tells you plainly that the hidden side is inferred understands the product. One who won’t say it either doesn’t understand it or hopes you don’t.
2006 versus now
Twenty years ago the same armchair would have spent an afternoon under a structured-light scanner, or been modelled by hand from a tape and a notebook, and the back would have been real because someone had looked at it.
| 2006 | Now | |
| Method | Scanner, or hand modelling | One photo, four stages |
| Time | An afternoon | Three minutes |
| Hidden side | Scanned | Inferred |
| Fit for a layout, a listing, a print | Yes | Yes |
| Fit for a measured drawing | Yes | No |
The no in the last row is permanent. The information was never in the photo, and no amount of training data can put it there.
That’s what image-to-3D does with your photo. It turns one view into a convincing whole by borrowing from millions of other objects, and for an armchair that needs to be in a room plan by this afternoon, borrowed is exactly enough.
