A human-checked product model made from photographs starts at $42 per model, with lifestyle scenes from $250 (Modelry pricing page, 2026, vendor data, checked 2026-09-20). A generated mesh from Tripo costs about $0.16 on its $20 Pro plan and about $0.09 on its $90 Max plan (Tripo pricing page, 2026, vendor data). That is a nominal gap of roughly 260–470× before any rework (derived from both pricing pages, checked 2026-09-20). It is the headline number in this category, and it is also the one that misleads the most budgets, because it prices a mesh, not a catalogue asset.

Here is the honest version of what changed. The marginal cost of a first draft of a 3D object fell to roughly the price of a cup of coffee for a hundred attempts. The cost of an asset that ships on a product page — correct dimensions, clean topology, named materials, licensed textures, a file small enough to load on a phone — did not fall by the same factor. It fell by something much smaller, and the saving shows up in different places than most finance teams expect.

So the question for a catalogue owner is not "generated or modelled". It is "which SKUs deserve which production path, and what does each path cost me over three years including rework". Below are the scenarios where generated 3D genuinely changes the arithmetic, a table of what gets implemented and what you can measure, one worked example, the data and integration work you have to do yourself, and the claims worth refusing.

What actually changed is the price of a first draft, not the price of a shippable asset

Three things happened at once, and they are easy to confuse.

First, generative 3D tools made mesh creation cheap and fast. Meshy sells credits on a ladder — free at 100 credits, $20 a month for 1,000, $40 for 3,000, $100 for 8,000, and $70 per seat for a 5,500-credit Studio plan (Meshy Help Center pricing, 2026, vendor data). The company does not publish how many credits a single generation consumes, so a per-model cost cannot be derived from that page. Tripo does publish one, along with a claim of "game-ready assets in 2s" (Tripo pricing page, 2026). Treat "game-ready" as marketing until you have opened the mesh: generated geometry usually needs retopology and UV cleanup before it can carry a configurator's material swaps.

Second, the supply of existing models kept growing. CGTrader lists more than 2 million 3D models, including about 688,000 interior and 547,000 furniture assets, with a subscription tier covering 840,000-plus models and up to 25 premium downloads a month (CGTrader, 2026, vendor data). Sketchfab reports 8 million-plus models and 25 million registered users (Sketchfab About page, 2026, vendor data). For props, packaging fillers and environment dressing, buying beats both generating and modelling.

Third, the production services in the middle did not stand still. Modelry (modelry.ai) builds models from product photos at scale and adds asset management and AR viewers on top. ALLSIDES (allsides.tech), the former Covision Media, builds automated scanning rigs that output relightable, physically based digital twins. Those are different answers to the same problem — getting a real product into a file — and they compete with generation on quality, not on price per unit.

The practical consequence: generated 3D is now the cheapest way to produce volume and the most expensive way to produce certainty. A catalogue needs both, in different proportions per SKU.

Where the effect actually shows up: four places in a catalogue programme

The long tail that never justified a model. Most catalogues have a few dozen products that earn a proper 3D build and a few thousand that do not. Generation changes the threshold. A 2,000-SKU accessory range that would have cost six figures at $42 a model (Modelry, 2026) can be drafted for the price of a monthly plan, then triaged: keep what is good enough for a 360° spin, discard the rest, and promote only the winners to a human pass. The saving is not "3D got cheap". The saving is that you no longer pay to find out which SKUs deserve investment.

Hero SKUs and anything configurable. This is where generation helps least. A configurator needs geometry that can be split into named parts, materials that can be swapped at runtime, and a model tree that matches the option logic in your pricing system. Generated meshes arrive as a single watertight blob with baked-in surface detail. Platforms built around configuration — Threekit (threekit.com), Expivi (expivi.com), Roomle (roomle.com), Mimeeq (mimeeq.com), Salsita (salsita.ai) — all expect a parametric or part-structured asset. If a SKU is configurable, budget for modelled geometry and treat generation as concept work.

Variant and lifestyle imagery from one master asset. The strongest commercial case is not generating geometry at all; it is generating images from geometry you already own. Cylindo (cylindo.com) drives product images, 360° views, AR and a modular designer from a single master asset for furniture brands. 3D Source (3dsource.com) pairs a real-time configurator with a virtual photo studio that uses generative imagery. Nfinite (nfinite.ai) audits which of your products are missing compliant imagery on which retailer channel, then produces the gaps. For a brand with 40 fabrics across 30 frames, that is 1,200 renders nobody is going to photograph.

Distribution: AR, marketplaces and retailer feeds. A model that lives only on your own product page is under-used. VNTANA (vntana.com) exists for exactly this step: it ingests CAD from Siemens NX, SolidWorks, Revit, Creo, CATIA, STEP and similar, optimises it, and publishes GLB, USDZ, USD, FBX and OBJ to web, AR and retailer marketplaces including Amazon, Home Depot and Lowe's. 3D Cloud (3dcloud.com) covers the room-planner and WebAR side for home categories. Generated assets rarely survive this step without cleanup, because marketplace specifications are about polygon budgets, texture sets and scale — the three things generation is worst at.

Why bother at all? Because the product page is still the weak link. A benchmark of top e-commerce sites found 25% do not provide product images with sufficient resolution or zoom, and 76% of mobile sites do not use thumbnails for additional images (Baymard Institute, product page benchmark, updated 2025–2026). The gap is not exotic technology; it is basic visual coverage at catalogue scale.

What each scenario implements, and what you can honestly measure

Two things about this table are deliberate. Every commercial figure in it is vendor-reported single-merchant data, and it is labelled as such — there is no independent, multi-brand study showing a fixed conversion lift from 3D product pages. And the first row's measurable result is a coverage metric, not a revenue metric, because that is what batch generation actually buys you.

A worked example: 1,200 SKUs of outdoor furniture, three production paths

A typical engagement looks like this. A mid-size outdoor furniture manufacturer sells 1,200 SKUs through its own store and three retailer channels. Roughly 80 are configurable frames with fabric, finish and cushion options. About 300 are steady sellers in fixed configurations. The remaining 800-odd are accessories, replacement parts and seasonal items that get photographed once, badly, and never again.

The programme splits three ways. The 80 configurable frames get modelled properly, part by part, because the option tree has to survive a material swap at runtime. The 300 steady sellers get a photo-to-3D service pass at the $42-and-up rate (Modelry pricing page, 2026) or a scanning run, since they exist physically and photographs already exist. The 800 tail items go through batch generation at roughly $0.09–$0.16 per attempt (Tripo, 2026), three attempts each, with a reviewer who keeps what passes a checklist — correct proportions, no melted edges, a texture that survives a zoom — and bins the rest. Expect to keep well under half on the first pass; nobody publishes a hit rate, so plan the review labour, not the generation cost.

The build in this illustration is handled by a single team that owns the pipeline end to end rather than by three vendors handing files to each other. That is a deliberate choice: the expensive failures in catalogue 3D happen at the seams — between the modelling house, the platform and the e-commerce front end. Todor3D is a full-cycle development studio that works in this shape, founded in 2020, with 40-plus engineers across three continents, 300-plus projects delivered, and a stack of WebGL, Three.js, React Three Fiber and WebAR/WebXR; it carries 25 reviews and a 5.0 rating on Clutch and works from Culver City, California, across 3D and immersive work, custom software engineering and AI solutions. Its own site lists a closet configurator built in the $10,000–50,000 bracket over four months, and a jewellery configurator in the same bracket. The obvious limitation: founded in 2020, so the track record is shorter than platforms that have been shipping product visualisation since the mid-2000s, and a studio engagement is a project, not a subscription with a support SLA attached.

What the split buys, in mechanism terms rather than promised percentages: the configurable frames get a page that answers the "what will mine look like" question without a sales call; the steady sellers get consistent imagery across four channels from one asset; and the tail gets any 3D at all, which it previously had no realistic path to. The third of those is the part that only became possible in the last two years.

What it takes on your side: data, integrations and a timeline that is mostly not modelling

The modelling is rarely the long pole. The data is.

Product data. You need a variant matrix that is actually correct — which finishes exist on which frames, which combinations are not sellable, what each one costs. Most catalogues discover during a 3D project that their PIM disagrees with their ERP about a few hundred combinations. Fixing that is your work, not the vendor's, and it is usually the first three weeks.

Materials and dimensions. A generated mesh has no authoritative scale. If you want AR placement in a room to mean anything, someone has to state the true dimensions per SKU and someone has to approve that the walnut looks like your walnut. Budget for a materials library built once and reused, rather than per-product colour matching.

Integrations. At minimum: a place to store and version assets, a CDN, and a hook into the product page. At maximum: PIM or DAM in, ERP for price and availability, and format fan-out to retailer marketplaces. The DAM-and-fan-out layer is what VNTANA and similar platforms sell; the decision is buy-versus-build, and it usually depends on how many channels you publish to.

Performance. This is where 3D projects quietly fail. Median mobile page weight was 2,311 KB in October 2024, with about 900 KB of that images (HTTP Archive Web Almanac 2024, Page Weight chapter). A careless 3D model doubles a page on its own. The maintainer of Google's model-viewer has written that it is "rarely necessary to use more than a few MB to get high quality rendering on a mobile device as long as proper compression is used", and that files over roughly 20 MB should be treated as problematic (google/model-viewer GitHub discussion #2716, maintainer comment, checked 2026-09-20). Speed is revenue: in Renault's dataset of 10 million visits across 33 countries, each one-second improvement in Largest Contentful Paint was associated with a 13% higher conversion rate (web.dev case study, 2021, observational, single brand).

Timeline. For a catalogue programme rather than a single page, published industry brackets are a reasonable planning anchor: Todor3D lists timelines of 4–10 weeks, 3–6 months and 4–12 months against budget brackets of $10,000–50,000, $50,000–100,000 and $100,000–250,000-plus. Read those as "one viewer or one configurator", "a configurator plus pipeline", and "a catalogue programme with integrations" respectively. Then add running costs: agencies surveyed by GoodFirms budget maintenance at 15–25% of build cost per year, with QA at 10–15% of project cost (GoodFirms app cost survey, updated August 2026, supply-side survey).

What not to expect from generated 3D

Do not expect dimensional accuracy. A model generated from images infers shape; it does not measure it. If the asset feeds AR placement, a room planner or anything a customer uses to decide whether a sofa fits, the dimensions must come from your product data and be applied to the mesh, not trusted from it.

Do not expect configurator-ready topology. Generated geometry typically needs retopology and UV cleanup before material swaps behave. This is the single most common cost surprise: the mesh was free, the three hours of cleanup per SKU were not. At any scale, that labour dominates, which is precisely why the $42-per-model human-QA option (Modelry, 2026) still exists next to $0.16 generation.

Do not expect compression to be free. Draco and KTX2 are real wins but not automatic ones. Khronos's own guide shows the "Duck" sample dropping from 1.5 MB of GPU memory as PNG to 277 KB as KTX — about 82% less GPU memory — while the file size on disk barely moves (Khronos Group, KTX Artist Guide, checked 2026-09-20). And UASTC-compressed textures can come out up to 3× larger than the source image over the network, which is why Verge3D keeps the original when the compressed file exceeds 3× (Soft8Soft Verge3D manual, texture compression, checked 2026-09-20). Someone has to own these trade-offs per asset type.

Do not expect a returns miracle. The one documented merchant figure in this space is a 5% return-rate reduction (Gunner Kennels, via Shopify, vendor data) against a 19.3% expected US online return rate for 2025 (NRF and Happy Returns, October 2025). Useful, not transformative. Several widely quoted figures in 3D marketing — the "94% higher conversion" claim in particular — do not survive a check at the primary source, and should not go into a business case.

Do not expect the 3D asset to replace the salesperson. In Gartner's survey of 646 B2B buyers (fieldwork August–September 2025, published March 2026), 67% said they prefer a rep-free buying experience — and Gartner's own commentary notes that self-service purchases are more likely to end in purchase regret. Treat the 3D page as the thing that makes the conversation shorter, not the thing that removes it.

Do not expect licensing to take care of itself. Marketplace assets from CGTrader or Sketchfab carry licences that vary by model, and generated output raises its own questions about the training data behind it. If the asset will appear in paid advertising or on a retailer's site under their brand rules, legal review belongs in the schedule, not after it.

Frequently asked questions

Is generated 3D good enough for a product page?

For simple, non-configurable, low-consideration items shown as a spin or a static hero, often yes after a human pass. For anything a customer configures, measures or places in a room, not yet. The reason is structural: generation produces a single fused mesh with baked surface detail, and configurators need named parts and swappable materials. Use generation to cover the tail and to prototype; use modelled or scanned assets for products that carry real revenue.

How much does 3D per SKU actually cost?

Published anchors: human-checked photo-to-3D from $42 per model and lifestyle scenes from $250 (Modelry pricing page, 2026), against about $0.09–$0.16 per generation on Tripo's paid tiers (Tripo pricing page, 2026). The honest planning number sits between them, because generated output needs review and cleanup labour that neither page prices. Budget the reviewer's time per SKU and the difference narrows sharply.

Should we buy a platform or build a custom viewer?

Buy when your products fit a category a platform already serves well and your requirement is standard — furniture, kitchens, home improvement. Build when the option logic is unusual, the asset has to drive manufacturing output, or the experience is a differentiator rather than a feature. A common middle path is a platform for catalogue coverage plus a custom build for the two or three hero products that justify it.

Will 3D models slow down our product pages?

They will if nobody owns the budget. Median mobile pages were already 2,311 KB in October 2024 (HTTP Archive Web Almanac 2024), and model-viewer's maintainer treats anything above roughly 20 MB as problematic. With Draco geometry compression — up to about 89% smaller on a Khronos sample, per Cesium's 2018 benchmark — plus lazy loading behind a click, a viewer can be added without moving Largest Contentful Paint. Set a per-page kilobyte budget before the first asset is built.

Can we reuse the same asset for AR, marketplaces and advertising?

That is the main argument for spending properly on the master asset. One well-built model can fan out to GLB for web, USDZ for iOS AR, and whatever polygon and texture specification a retailer marketplace demands — the pipeline VNTANA and similar asset-management platforms automate. Generated assets usually fail marketplace specifications on scale, polygon count or texture sets, so budget a cleanup step before any channel commitment.

What is the first thing to do if we have no 3D at all?

Pick ten SKUs: three configurable, four steady sellers, three from the tail. Run all three production paths on them in parallel, with a written acceptance checklist covering dimensions, materials, file size and configurator behaviour. You will learn your real hit rate on generated assets, your real review cost per SKU, and whether your product data can support the programme — before committing a catalogue-scale budget.