Autodesk Announces Project Bernini, a New Generative AI Tool for Creating 3D Shapes

Autodesk today announced a new research project called Project Bernini, whose mission is to develop generative AI models that can quickly generate functional 3D shapes, which can then be further refined by modelers.

The following are some of the ways in which the modeler can further improve the model.

The following is an introductory video:

How Bernini works - Bernini can generate multiple functional variations of a 3D shape from a variety of inputs, including a single 2D image, multiple images showing different views of an object, point clouds, voxels, text . For more information, see this blog post on the Autodesk website.

Model Creator- Bernini is an in-house project developed by scientists and experts at the Autodesk AI Lab, a division of Autodesk Research. The research paper contributing to this study was published earlier this year by the AI Lab in collaboration with the Chinese University of Hong Kong.

Model Learning Methods - Autodesk trained its models on 10 million 3D shapes. This is a composite dataset consisting of publicly available data with a mixture of CAD objects and organic shapes. [Autodesk envisions a variety of use cases for the model, which is intended to be used by designers in architecture, product design, and entertainment production. Bernini's main goal is to create "functional 3D structures" that can be built or manufactured in the real world. Autodesk further explains what that means and how Bernini differs from other generative AI models:

A simple example is a jug. Many other 3D generative models might produce a jug-like shape with a texture that improves its superficial appearance in very specific lighting environments. However, the Bernini model generates shape and texture separately and does not confuse or fuse these variables. Thus, the jug generated by our model is hollow in the middle and can actually hold water, just as a real-world jug needs to.

How can we test it? Project Bernini is strictly experimental at this stage and is not open to the public. The company is looking for potential partners to help with the model and improve the performance of the generative model.