NVIDIA described the GET3D artificial intelligence model, which will facilitate the work of virtual 3D world creators. The algorithm quickly generates people, buildings, vehicles and many other types of three-dimensional objects. One graphic processor, declared to the company, allows GET3D to generate up to 20 objects per second.
The NVIDIA engineers trained AI on flat projections of three-dimensional objects made from different angles. The training was carried out on A100 boosters; it took only two days to download a million samples into GET3D. The system generates objects in the format of a standard triangular polygon grid with high resolution textures. Technology is aimed at the creators of three-dimensional worlds, so it is possible to export objects in formats compatible with game engines, 3D simulations and visualization for films. This can indeed make it easier for developers to create games and applications of metallurgical, the system will also be useful in robotics and architecture.
The algorithm distinguishes the nuances between similar objects: from the vehicles it can generate saddles, trucks, racing cars and vans, from the animal category creates images of foxes, rhinos, horses and bears, the set expands as training takes place. NVIDIA also described the StyleGAN-NADA support IE model, which allows changes in the style of objects: to make the car burned, to turn the building into a haunted house, or to give tiger color to any animal. The company ' s future versions of the GET3D system are going to be trained in real objects rather than artificial models.