May 8, 2024 - Technology

Exclusive: Autodesk's AI turns text or still images into 3D models

A screenshot of Autodesk's "Project Bernini" a generative AI model for 3D content

A screenshot of Autodesk's Project Bernini, a generative AI model for 3D content. Image: Autodesk

Autodesk, the venerable 3D software company whose tools are used by designers and architects around the globe, is showing off its most ambitious generative AI project to date: a text-to-3D tool called Project Bernini, demonstrated first to Axios.

Why it matters: 3D design and modeling — an essential part of the modern creative process in fields ranging from architecture and product development to drug research and robotics — is one more realm that's about to face the tidal wave of AI disruption.

  • For software makers who can harness it effectively, generative AI can both democratize technology that was once limited to a few and improve productivity for professionals.
  • By the same token, incumbents who don't get it right risk having their long-held franchises toppled by an upstart.

How it works: Project Bernini can create 3D images from a variety of input types including a text prompt, a single 2D image, multiple images and point clouds.

  • The AI model that underlies Project Bernini is designed to have an understanding of how real-world objects tend to exist. For example, it can see an image of a water pitcher and understand that the resulting 3D image should be hollow and water-tight.
  • Typing in the word "chair," meanwhile, will bring up a variety of 3D images of chairs, from which a designer can further refine the design to reflect what they have in mind, using either text or another image.
  • "The main thing is it actually produces good geometry, good shapes, because the designer wants good smooth surfaces, correct topology," Autodesk senior VP and research chief Mike Haley told Axios.

Yes, but: Autodesk is only previewing the tool, which it says is experimental, and it is not making it publicly available yet.

The big picture: There are other text-to-3D tools out there, but Autodesk is banking that its offering will be the most successful at creating images that can fit into a professional workflow.

  • Haley said Autodesk's model is designed to perform even better once companies train it on their own 3D content.
  • "The reality is the internet is not populated with tons of great 3D content," Haley said.
  • Companies could keep their tuned models to themselves, but Haley said there also could be an opportunity for those with substantial 3D libraries to license some of their content to help train models used by others, akin to what Adobe has done with Firefly.

Between the lines: Autodesk isn't the first to devise a 3D generative model. Nvidia released a tool last year, and startups like Luma and Meshy are already in the space.

  • But Autodesk has long owned the 3D corner of the professional design software market, and it's claiming other advantages.
  • Haley says one ability that helps Project Bernini stand out from rivals is that it starts by separating texture from the outline of the object itself.
  • Many models, he said, get tripped up by mistaking the texture of a 3D object from its contours.

Fun fact: The project name is a reference to Gian Lorenzo Bernini, a 16th-century Italian artist and architect credited with creating the Baroque style of sculpture.

What's next: Haley said Autodesk is still considering different ways to turn the model behind Bernini into something customers can use.

  • Among the options is offering a free web-based experience a la ChatGPT. "We're working through that right now," Haley said.
  • Autodesk says it is working on additional models for different industries, including architecture, product design and entertainment.
  • "What you're going to see from us this year is a sort of a gradual rollout of a succession of these generative AI models," Haley said.
Go deeper