NVIDIA trains AI to generate 3D objects

Nvidia unveiled the artificial intelligence model GET3D, which generates characters, buildings, vehicles and other three-dimensional objects.

According to the developers, the model can generate about 20 objects per second using a single graphics processor.

GET3D was trained on a dataset of 2D images of three-dimensional objects captured from different angles. Nvidia said it took two days to upload about a million images into the model using A100 GPUs.

From the training set of vehicle photographs GET3D generated sedans, trucks, race cars and vans. It can also generate foxes, rhinos, horses and bears after training on animal images.

“The more extensive and varied the training set loaded into GET3D, the more diverse and detailed the result,” the developers said.

According to Nvidia spokesperson Ishi Salian, the system is capable of generating objects with high-quality textures and intricate geometric details.

“GET3D-generated objects have the form of a triangular-mesh, similar to a papier-mâché model, coated with a textured material,” she added.

Users can quickly import objects into game engines and visualization tools for editing. GET3D creates them in compatible formats, the company said.

In October, Nvidia held the GTC 2022 conference, at which it unveiled new GPUs, tools for AI developers, and the Omniverse cloud platform.

The following day, the company announced the professional graphics processor RTX 6000 for AI engineers, artists and researchers.

Subscribe to ForkLog AI on Telegram: ForkLog AI — all the news from the world of AI!