NVIDIA GET3D, as this model is called, has been trained on nearly a million 2D images of 3D objects captured from different angles, using the Nvidia A100 Tensor Core GPU, which took the researchers around two days.
From this training the AI model is capable of creating three-dimensional objects “with high-fidelity textures and complex geometric details”, as the company states on its official blog. And generate them, in addition, quickly, since it can produce about 20 forms per second.
All this means that, for example, from the images captured in 2D of cars, the model can generate a collection of sedans, trucks, race cars and vans; or if it is images of animals, creatures such as foxes, rhinos, horses and bears.
These 3D objects can be imported later for use in game engines, 3D modelers, and movie renderers.and edit to apply more realistic lighting effects, for example.
Nvidia has indicated that in the future they hope to incorporate camera pose estimation techniques and universal generation, to expand the capabilities of this tool.