Getting 3D assets for AR without blocking the rollout
Most AR projects are not stopped by the technology. They are stopped six weeks in, when someone works out that covering the catalogue means producing two thousand 3D models and there is no plan for who does that. Asset creation is the dominant cost, the dominant timeline risk and the thing that determines whether you get past a pilot, and it is the part least likely to be discussed in a pitch.
Where do product assets come from?
Four routes, and the one available to you depends entirely on what your business already has rather than on what the AR supplier prefers.
If your products are manufactured, you probably have CAD, and that is the best starting point by a wide margin. It is dimensionally accurate, it exists already, and converting it is a known process rather than a creative one. The conversion is not free, because engineering CAD carries far more geometry than a phone can render and no material information a customer would recognise, but it is predictable work.
If they are not, you are choosing between photographing them into a model, modelling them by hand, or generating them. Each has a shape of failure worth knowing before you commit a catalogue to it.
| Route | Works well for | Fails on | Per-item effort |
|---|---|---|---|
| CAD conversion | Anything manufactured: machinery, furniture, appliances, fittings | Nothing much, but needs geometry reduction and real materials adding | Low and predictable |
| Photogrammetry | Matte, opaque, textured objects with surface detail | Glass, chrome, gloss, fur, anything thin or transparent | Low per item, high setup |
| Manual modelling | Hero products, anything needing exactness | Catalogue scale, because it is skilled hours per item | High |
| AI generation | Filling a gallery, concepting, non-critical background objects | Dimensional accuracy, brand-critical detail, anything a customer measures against | Very low, with a review cost |
Can AI generate product models yet?
For some purposes, and not for the purpose most people are asking about. Generated 3D has become genuinely useful for concepting, for populating a scene with objects nobody will scrutinise, and for producing a rough starting point a modeller then corrects.
It is not yet trustworthy where dimensional accuracy matters, and product AR is mostly a dimensional-accuracy problem. The entire commercial argument for placing a sofa in a room is that the sofa is the right size; a model that looks correct and is eight percent wrong produces confident buyers who return the product, which is worse than no feature at all.
The honest current position is that generation shifts the cost of the first draft, not the cost of correctness. Anyone claiming it removes the asset problem for a product catalogue is describing a future state.
Which formats do you actually need?
Two, and they are not interchangeable. glTF, usually in its binary GLB form, is the web and Android format and the one your experience will run on. USDZ is Apple's, and it is what lets an iPhone open a model in its native viewer straight from a link.
Supporting both is normal and not especially painful, because they can be produced from the same source. What matters is that the source is kept. A pipeline that produces only the two delivery formats has thrown away the master, and every future change means starting again.
Ask any supplier what the source format is, where it is stored, and whether you have it. The answer to that question is worth more than their answer about rendering quality.
How small does an asset need to be?
Small enough to load in a couple of seconds on a mid-range phone over mobile data, which in practice is far smaller than a first export. Engineering CAD and film-quality models carry orders of magnitude more detail than a five-inch screen can resolve, and shipping that detail is the single most common reason a demo gets abandoned before it renders.
The work is geometry reduction plus baking surface detail into textures rather than modelling it, and it is a craft skill rather than a button. Done well, the difference is invisible to a customer and decisive for whether they wait for it.
A useful acceptance test: open the experience on the oldest mid-range Android phone in the office, on mobile data, and count. If nobody in the process has done that, the assets are almost certainly too heavy.
Who adds the next hundred products?
This is the question that decides whether the rollout is bounded, and it is worth resolving before signing anything. There are three answers and they have very different long-run costs.
The agency does it, at their rate, forever. Workable for a small hero range, ruinous for a catalogue that changes seasonally, and it makes your release schedule dependent on their capacity.
Your team does it, using a documented process and tooling they are trained on. Slower to start and dramatically cheaper past the first hundred items. This is the right answer for anyone with an evolving catalogue, and it requires asking for it explicitly, because it is not the default any supplier proposes.
Or it is automated from a source you already maintain, which is the strongest position and only available when the products have CAD or a consistent photography standard. If your PLM or product database already holds the geometry, new products can become assets as part of the existing product-launch process rather than as a separate project.
Who owns the assets?
You should, in the source format, with no dependency on the supplier's platform to use them. This is the real lock-in in this industry and it is invisible until the moment you want to change supplier or bring the work in-house.
The pattern to watch for is assets that live inside a proprietary system and are only exportable as low-resolution delivery files. Formally you own them; practically you own a copy that cannot be edited, which means switching supplier means rebuilding the library. For a two-thousand-item catalogue that is a cost large enough to prevent the switch, which is the point of the arrangement.
Put it in the contract in plain terms: source files, delivered to us, in an open format, on request. A supplier confident in their work has no reason to object, and the reaction to being asked is informative in itself.
Common questions
- Where do 3D models for AR product experiences come from?
- Four routes: converting existing CAD, which is the best option and available to most manufacturers; photogrammetry, which photographs a real object into a model; manual modelling by a 3D artist; and AI generation. Which is right depends on what your business already holds rather than on supplier preference, and CAD conversion is dramatically cheaper and more predictable than the alternatives.
- Can photogrammetry scan any product?
- No. It works well on matte, opaque objects with visible surface texture, and fails on glass, chrome, gloss, fur and anything thin or transparent, because those surfaces do not present consistent detail for the reconstruction to match between photographs. Jewellery, glassware and polished metal usually need modelling instead.
- Can AI generate 3D models for product AR?
- It is useful for concepting, for background objects and for producing a first draft that a modeller corrects, but it is not yet reliable where dimensional accuracy matters, and product AR is mostly a dimensional-accuracy problem. A model that looks right and is several percent wrong creates confident buyers who then return the item, which is worse than having no AR at all. Generation currently shifts the cost of the first draft, not the cost of correctness.
- What 3D file formats does AR need?
- glTF, usually as binary GLB, for web and Android delivery, and USDZ for Apple's native viewer. Both can be produced from one source, so supporting both is routine. The important thing is that the editable source file is retained and delivered to you, because a pipeline that keeps only the two delivery formats has discarded the master and every future change starts over.
- Who should create 3D assets for new products after launch?
- Ideally your own team, using a documented process, or an automated pipeline fed from CAD or a consistent photography standard you already maintain. Routing every new product back through the agency at their day rate makes the running cost open-ended and ties your release schedule to their capacity. Ask for the self-service option explicitly, because no supplier proposes it by default.
- Who owns the 3D assets an agency creates?
- You should own the editable source files in an open format, not just low-resolution delivery exports locked inside the supplier's platform. This is the genuine lock-in in AR work and it only becomes visible when you try to change supplier, because rebuilding a large asset library can cost more than staying. Specify source-file delivery on request in the contract.
