A buyer asks for a 3D model, you run one factory photo through a generator, and ninety seconds later something is rotating on your screen. Then the buyer asks the only question that matters — "what's the seat height?" — and the model has nothing to say.
A 3D product model from a photo is a surface reconstruction: it copies the shape the camera could see into an object you can spin. It carries no measurement, no material and no tolerance, because the photograph never recorded any of them. That is not a flaw in the generator. It is what a photograph is.
Four assumptions about these models keep causing trouble in export sales, and each one has a cheap fix.
Myth 1: the model comes out to scale
A photograph records angles. It does not record distance, and distance is the only thing that turns an angle into a millimetre. Reconstruct a scene from that image and every point in it is known relative to every other point only in an arbitrary unit — the software decides what that unit means.
The practical result: two models generated from two photos of the same sofa can disagree on every dimension by a margin a buyer would notice, and both look completely convincing on screen. Nothing in the file says which one is right, because nothing in the file came from a tape measure.
This is the assumption that breaks first in front of a buyer. They rotate the model, see a plausible dining chair, and read the proportions as fact. If your spec sheet says 450 mm seat height and the model's proportions imply 520 mm, you have created a contradiction the buyer will notice — and they will trust the one that looks more real, which is the model.
The fix is not a better generator. It is to put the real numbers where the buyer can see them, on the photo, measured once:
- one reference dimension taken off the physical unit, in millimetres, with the tape or rule visible in a second shot if the buyer is new to you
- every other dimension derived from that reference, not estimated by eye
- the same numbers repeated in text on the spec sheet, so the model, the photo and the table never disagree
Myth 2: a 3D model replaces the dimension drawing
They answer different questions. A mesh answers "what shape is it?" A drawing answers "will it fit, and how much can it vary?"
| What the buyer needs |
3D model (glTF/GLB) |
Dimension drawing or spec table |
| Overall shape and silhouette |
Yes, from any angle |
Only the views you drew |
| Nominal dimensions in mm |
No, unless you declare a scale |
Yes, by definition |
| Tolerance and what varies |
No |
Yes — that is the point of the drawing |
| Datum and measurement origin |
No |
Yes |
| Units and rounding rules |
No |
Yes, on the sheet |
| Surface colour and finish |
Approximate, depends on texture and lighting |
A finish code or a swatch |
| Revision control |
A file name, if you are disciplined |
A revision block with a date |
The format itself is honest about this. glTF is an interchange format for 3D scenes — Khronos describes it as a royalty-free spec for "the efficient transmission and loading of 3D scenes and models," and glTF 2.0 was published as the ISO/IEC 12113:2022 standard in 2022. Nothing in that standard is a tolerance, a datum or a nominal size. Those live on the drawing, which is why serious buyers ask for both and treat them as one package.
Worth knowing where the boundaries sit: a dimension drawing vs spec diagram vs CAD comparison covers which of the three documents a buyer can actually act on.
Myth 3: whatever the photo didn't show, the model fills in
A reconstruction only knows the surfaces the camera saw. The back panel, the underside, the inside of the drawer box, the joinery where the leg meets the apron — none of that is in a front three-quarter shot. The generator still has to close the object, so it does what any model does when data is missing: it continues the surface in the most plausible way.
That plausible way is often wrong in exactly the details a trade buyer checks:
- a back panel rendered flush when the real unit has a 12 mm reveal and visible fixings
- a leg-to-apron joint smoothed into one continuous form, erasing the bolts a buyer needs to see for flat-pack assembly
- an open shelf closed off, or a hollow moulding rendered solid
- a mirror or glass surface reconstructed as an opaque face
Then the buyer rotates the model to the back — the first thing an experienced importer does — and finds a product that does not exist. The model did not lie about your product. It filled a gap you left open.
Two habits prevent it. Shoot the sides you intend to claim, and say plainly what was not photographed: "rear panel not modelled", "drawer interior not shown". A one-line caption costs nothing. A buyer discovering invented geometry costs the relationship.
If the internals are the actual question — what the frame is made of, how the hinge mounts, what a replacement part looks like — a 3D model is the wrong document. That is a parts diagram job, and it needs your real component list, not a surface.
Myth 4: if it rotates on your laptop, it will load for the buyer
A model that previews instantly on a studio machine can sit as a static poster on a three-year-old Android phone on a hotel wifi connection. The failure is not resolution. It is delivery.
A few numbers that decide whether the file arrives:
- Display needs glTF or GLB. In Google's model-viewer documentation, a USDZ-only file "won't display a model (as that requires a glTF or GLB)" — USDZ is the AR path, glTF is the display path.
- Compression has a floor. Compressing the mesh with the DRACO extension needs a decoder that model-viewer fetches on demand, and the documentation puts it at "more than 100KB" — so compression only pays once the savings exceed the decoder.
- Frame rate is defended by dropping resolution, not by loading less. The same documentation says the renderer "tries to keep the frame rate between 38 and 60 frames per second" by scaling the rendered resolution, down to a minimum scale of 0.5. On a weak device the buyer gets a softer image, and on a heavy file they get a long blank rectangle first.
The test that matters takes ten minutes: open the link on an older phone, on mobile data, not in your office. If you cannot, the file is not ready. Numbers for a workable target live in the 3D product model file size breakdown. Before sending, validate the asset with Khronos's own glTF Asset Auditor rather than your viewer — viewers are forgiving, validators are not.
What a 3D product model from a photo is genuinely good for
None of the above makes the format useless. A 3D product model from a photo is a shape document, and shape is a real buying question:
- Orientation and form — where the handle sits, which face is the front, how the profile curves. Buyers deciding between two similar SKUs read this faster from a rotating model than from eight flat photos.
- Assembly order — a model that can be exploded, paired with numbered parts, explains sequence better than a paragraph.
- Fewer "what does the back look like" emails — if you modelled and photographed the back.
- A consistent look across a range — the same camera angle and lighting on forty SKUs reads as one product family.
What it cannot do is stand in for the numbers. A rotating model can show a buyer the shape of a product; only a measurement can tell them whether it fits. That sentence is the whole article, and it is the one to keep in the email template.
The two-document handoff that answers the second question
The workable pattern is not "send the 3D product model from a photo instead of the sheet." It is one asset pair, sent together, where each answers a different question.
| The buyer's question |
What to send |
Why it works |
| "What does it look like from the side?" |
The 3D view, or four orthographic photos |
Shape needs multiple angles; one hero shot cannot answer it |
| "Will it fit my space?" |
Annotated photo with measured dimensions on the real edges |
The number comes from a tape measure, not from the mesh |
| "How much can the size vary?" |
Spec table with tolerance per dimension |
Only a stated tolerance prevents a dispute |
| "What's inside it?" |
Breakdown or parts diagram with part numbers |
Internal structure is a component list, not a surface |
| "Can I show my client?" |
One link with the model, the numbers and the finish code |
A forwarded file loses its context; a shared page keeps it |
A 3D product view built from the same product photo as the measured labels keeps those two documents from drifting apart — the model shows the form, the labels hold the numbers, and both come from one source file instead of three folders.
Before you send a 3D model
FAQ
Is a 3D product model to scale?
No, not by default. A model generated from a photo inherits an arbitrary unit because a photo records angles, not distances. It is to scale only if a person supplies a reference measurement and sets it — which is why the number on a rotating model should never be treated as the number on the drawing.
Can buyers measure a product from a 3D model?
Not reliably. A viewer measures the silhouette you can see, not the nominal dimension the factory produces, and it has no tolerance attached. If the buyer needs to check fit, send dimensions taken off the physical unit, with the tolerance stated, and let the model answer the shape question.
What file format should I send for a 3D product viewer?
glTF or GLB for display, because that is what browser viewers load. USDZ is the AR path on iOS and will not render in a web viewer on its own. If you send both, say which one is for rotation and which is for AR, so the buyer does not open the wrong file and conclude your model is broken.
How accurate is a 3D model from a photo?
Accurate in silhouette for the surfaces the camera saw, unreliable for anything it did not. Occluded faces, interiors and joints are filled in as plausible continuations, so a model can be visually convincing and still show a back panel that does not exist. Accuracy of the shape is a photography question; accuracy of the size is a measurement question.
Do I still need photos if I have a 3D model?
Yes. Photos carry texture at a resolution a mesh usually does not, they show the real finish under real light, and they are what a buyer forwards in a chat. The practical habit that saves the most back-and-forth is to build the model and the annotated photo from the same capture, then share them on one page with the spec table — that is the pattern behind a product card for buyer handoff, where the model, the measured labels and the contact details travel as one link instead of three attachments.
Sources & References