The AI-powered 3D object generator may be a work in progress—but for designers, it’s the writing on the wall.
Don’t waste your time looking for Autodesk’s Project Bernini. You won’t find it online. Only a few AI researchers at Autodesk, and perhaps some privileged customers, will ever use the 3D geometry generator in its current form.
That’s because Project Bernini is still a work in progress. Its output is crude. Like many generative AI tools, there’s a yawning gap between its capabilities and those of a properly trained human. CAD designers could sketch circles around Bernini.
For now.
Despite its limitations, Bernini is the most functional version of a 3D object generator that we’ve yet seen. It’s proof that the same miraculous-seeming technology that can instantly write an essay or paint a picture from a text prompt can do the same for 3D models.
Generative CAD is coming—but designers still have a chance to shape the technology for themselves.
Generative AI for 3D models
If you’ve used ChatGPT or any other generative AI tools, you know how Project Bernini works. Give it something to generate—a prompt—and the AI hands back a 3D model.
Type “water pitcher” and in seconds Bernini spawns a functional 3D model of a water pitcher, watertight and empty on the inside. You can also prompt Bernini with a picture, point cloud or voxel file.
Like other generative AI models, Bernini will serve up as many results as you care to sift through. Here are Bernini’s first four stabs at a chair:
One could have a lot of fun experimenting with Bernini’s response to different search terms. Still, it quickly bumps into limits. The AI’s attempts to generate a graphing calculator ranged from bad to awful.
Project Bernini was trained on 10 million 3D objects, a paltry figure in the world of generative AI. But the model can still find interesting ways to fill the gaps. Its training data almost certainly didn’t include a “tree dog,” but if you type that phrase into the prompt field, Bernini delivers:
“We see it very much as a creative tool,” Daron Green, chief scientist at Autodesk Research, told Engineering.com during a demo of the 3D generator.
Green expects that Bernini will serve as a starting point for designers, a way to gain inspiration and speed up the 3D modeling process.
“Instead of somebody laboring to create a 3D object,” he said, “they’ll be given a whole host of objects that are roughly of the sort that they’re interested in. And then they’re able to refine them.”
At this point, it’s still anybody’s guess how generative 3D will fit into a modeling workflow. But one thing is certain: For Project Bernini to become a useful tool for engineers, engineers will have to get involved.
Generative CAD for engineers
Green and the team at Autodesk are excited about Project Bernini’s potential, but they’re keenly aware of its present limitations. For one thing, the 10 million 3D objects that form Bernini’s training data are all from public sources, and many prohibit commercial use.
To make Bernini a professional tool, Autodesk needs professional data, and lots of it. The generative AI model is industry agnostic, but eventually Autodesk will want to incorporate it within industrial software. To get engineering data for Bernini, the developer needs to collaborate with its engineering customers.
“And therein lies a challenge,” Green said, “because very few companies have aggregated data at the volume that’s needed.”
Another challenge is Bernini’s interface. The version seen in the screenshots above is simplistic, with no memory for prompts. It includes a basic voxel editor, but will designers want to use it? Autodesk has no idea. The question of how generative AI will fit into a CAD workflow remains wide open.
“In parallel to the model development, we have a lot of UX/UI research going on as well,” Green said.
One big advantage of Bernini is that it’s inherently scalable. Hooman Shayani, senior principal AI research scientist and senior manager at Autodesk, compared Bernini to a “JPEG for 3D shapes” in the way it compresses 3D training data.
“That compression allows us to bring down the cost of storing the data, the cost of training the model,” Shayani told Engineering.com. Eventually, this will enable end users to efficiently train Bernini on their own data and create custom AI models without a prohibitively high compute cost.
When will you be able to use Project Bernini?
Project Bernini is not a unique effort. It’s one of several 3D generative AI models that researchers at Autodesk and elsewhere are developing. By the time generative 3D makes its way to CAD software like Autodesk Inventor or Fusion (formerly called Fusion 360), you won’t know or care if it’s based on Bernini or some other model.
But when will that be? Green expects it will take a couple of years for Autodesk to work out the best generative AI model, training set and user experience to integrate 3D geometry generation in its software.
Yet users may get some crumbs of generative CAD much sooner. “We do have product-related AI capabilities that are going to get shipped later in the year,” Green said, declining to offer further details.
Since ChatGPT sparked a tidal wave of interest in generative AI, 3D designers have awaited the day this powerful technology would arrive on their shores. Generative CAD is now taking shape on the horizon. Can you feel the breeze?