Can AI Generate 3D Models From Photos? We Tried It in 11 Iterations

Every 3D product configurator needs 3D models, and building them is one of the bigger costs in a configurator project, especially for manufacturers with big catalogs. AI models such as OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 gained real 3D capabilities this year. GPT-6 Astra can work directly in Blender, building geometry and setting up materials, and in one demo Claude Opus 5.5 modeled a Shinkansen train from more than 5,000 separate objects. So we wanted to know whether AI for 3D model creation is ready to take on that work.

I ran an experiment to find out. I gave an AI agent photos of an F1 car and asked it to build a 3D model. I then added two reviewer agents that compared each version with the photos and sent feedback, and let the three of them work for 12 hours.

AI can generate 3D models from photos, and our result is a solid starting point. The final AI 3D model looks convincing, and its geometry is reasonably well structured. Up close, some shapes and details are still off, so a 3D team would need to finalize it before it could go into a 3D configurator. In this post, I explain how the experiment was set up, what the AI got right and wrong, and what it could mean for companies that need 3D models of products.

The experiment at a glance

What Did AI Produce in the First 20 Minutes?

In 20 minutes, the AI agent produced a model that was clearly an F1 car but got many of the shapes and details wrong. I took the idea from a BetterStack demo. I gave an AI agent a set of photos of an F1 car and asked it to build a 3D model. The agent got the overall idea of the car right. The proportions were roughly there, but the shapes were awkward and a lot of the detail was missing or misshapen.

That is a lot to get in 20 minutes. If you only need a quick visual, such as a first look at an idea, a rough model like this will already do the job.

But a 3D product configurator asks for much more. Buyers rotate, zoom and inspect the product from angles a photo never shows. At that level of inspection, the 20-minute model was only a starting point.

How Did the Automated Review Loop Work?

Instead of stopping at the first result, I split the job between three agents. A designer agent built and revised the model. Two reviewer agents compared each new set of 3D renders with the original photos, pointed out the differences and gave the version a score. The designer agent then worked from that feedback on the next version, and the cycle repeated until the score reached 95%.

How the AI review loop improved the 3D model:

Each pass through the loop counts as one iteration. It took 11 to get from the first score of 48% to the 95% target.

The point of the setup was to take me out of the middle of the process. Normally a person checks each version, finds what is off and writes the next instruction. Here the reviewer agents did that work. The run took 12 hours but needed very little of my attention until the end, when I looked at the result and judged it myself.

What Does the 95% Score Mean?

The reviewer agents scored each version on how closely its renders matched the reference photos. The target was 95%, and the loop stopped once their score reached it. Nobody checked the model's accuracy independently, and the run did not test whether it was ready for production.

A score like this is useful for running the loop, but it only covers how the model looks in renders. It says nothing about whether the dimensions are correct, how clean the geometry is, how the model performs in a browser or how easily a 3D artist could edit it later.

When I went through the final model myself, it looked impressive, but in the details a lot of small things were off or misshapen.

Is an AI 3D Model Good Enough for a Product Configurator?

Probably, yes. The process is still slow and costly, though, and a person needs to review the final model before it goes into a configurator. The wireframe is encouraging as well: around the cockpit and bodywork, the mesh looks reasonably well structured.

But rendering well and working inside a configurator are different tests. Before we would treat an AI 3D model as a configurator asset, we would check it in four areas:

1. Viewing angles. Customers turn the product and zoom in, so a flaw hidden in one render can show up in another.
2. Materials and options. A configurator swaps finishes, colors and parts, and the model has to be built so those swaps look right.
3. Browser performance. Configurators load 3D models on the web, often in formats such as glTF, and a heavy model can slow the page down.
4. Editing and upkeep. Products change over time, so a 3D specialist needs to be able to open the model and update it without starting again.

The next step is to put this model into a working configurator and get a 3D specialist's view on how practical the geometry is. For the wider list of what a configurator needs, see our guide on what to look for in configurator software.

What Would It Take to Create 30 AI 3D Models?

Our run took 12 hours. Run one after another at the same pace, 30 similar models would take about 15 days, and that is before anyone inspects or cleans them up. They would also use roughly 12 billion tokens. AI providers charge by the token, so the cost of running the AI matters as much as the time.

The runs don't have to happen in sequence, though. Separate models can be worked on at the same time, and I see that as the bigger opportunity: once the workflow works for one model, it can be repeated across several agents. How far you can push it depends on your compute capacity and token budget, and on how many people you have to review the results.

Parallel runs don't make any single model cheaper or quicker to produce. They shorten the wait for the whole batch. That is why I find it more useful to ask how much of this work can run side by side than how fast AI is.

With the agents running side by side, most of the remaining work sits with people. The reviewers can only compare against the photos they get, so the references have to be good. Because the loop improves whatever the reviewers score, the review criteria decide what "better" means for your product. And at the end, a person still decides whether the output is usable.

What This Means for 3D Teams and Configurator Projects

The main lesson from the run is about process. If agents have a repeatable way to check and improve their own work, the job needs far less of your attention, even when it runs for hours. Human review can then go where it helps most: accepting the result and listing what still needs finishing.

For a 3D team, the obvious question is how much work it would take to turn this output into an asset they could ship. We don't have that number yet, and it will differ from product to product.

The run also matches what we saw when we tried to build a whole configurator with AI. In our vibe coding test, the 3D side came together quickly, while product rules and pricing still needed people who knew the product. Here too, the visual work got faster and the final call stayed with a person.

How We Use AI at Salsita

At Salsita, we build 3D configurators for manufacturers with complex and modular products. Our clients usually provide the 3D models of their products, and we build the configurator around them: the product rules, the pricing and the integrations to their other systems. We don't create the models ourselves, but we follow closely how AI is changing the work around them.

That is why we run experiments like this one. We test each new AI model on real configurator work, from building a whole configurator to, in this case, the 3D models every configurator depends on. Knowing where AI holds up and where it falls short means we can give our clients a realistic picture when they ask what it could do for their project.

AI is also part of the product itself. The AI assistant in our configurators guides buyers through their options in plain language and applies changes in real time. It works within the manufacturer's product rules, so everything it suggests is something they can actually build.

Want to see how a 3D configurator could work for your products? Book a demo, bring one product and leave with a clear next step.

FAQ

Can AI generate 3D models from photos?

Yes. In our test, an AI agent built a recognizable 3D model of an F1 car from photos in about 20 minutes. With two AI reviewer agents checking its work over 12 hours, the model got much closer to the photos, although some shapes and details were still off.

How long does it take AI to create a 3D model?

Our first rough model took about 20 minutes. Reaching the reviewers' 95% similarity target took 11 iterations and 12 hours. A different product or a different target would change those numbers.

How accurate is an AI 3D model?

In our run, the final model scored 95% for visual similarity, as judged by two AI reviewer agents. That figure is the agents' own estimate and was not checked independently. Up close, several small details were still off or misshapen.

What is an automated review loop in AI 3D modeling?

One AI agent builds the model while other agents check it. In our case, two reviewer agents compared renders with the reference photos, scored each version and sent feedback. The designer agent revised the model, and the loop repeated until the score reached the target.

Can I use an AI 3D model in a product configurator?

You can use it as a starting point. Before it goes live, check it from every angle, make sure materials and options switch cleanly, test how it performs in a browser and confirm that a 3D specialist can maintain it.

How much does it cost to generate a 3D model with AI?

There is no general price, since it depends on the AI model you use and what the provider charges. For reference, our 12-hour run used about 400 million tokens.

Can AI create 3D models for a whole product catalog?

In principle, yes, if you run several models in parallel. Thirty runs like ours done one after another would take about 15 days. Running them side by side cuts the waiting time, within the limits of your token budget and the people available to review the results.