NVIDIA showed a neural network that turns 2D images into 3D objects
Miscellaneous / / July 21, 2022
NVIDIA, together with experts from Stanford University, has finalized the EG3D neural network, which is capable of generating realistic three-dimensional images based on flat 2D images.
Technology has been demonstrated on the example of portraits of people and muzzles of cats. In all cases, the volumetric image was quite realistic, with correct proportions and no visible artifacts.
To create a three-dimensional model, a neural network generates a series of two-dimensional images using generative adversarial networks, and then creates a 3D object based on them.
The developers note that EG3D works much faster and more accurately than analogues, and does not require colossal computing power - you can use it even on laptops with top gaming video cards. For example, everything worked fine with the GeForce RTX 3090.
The neural network is already available for self-testing - the source code and installation instructions are available on GitHub.
In March, NVIDIA specialists demonstrated a similar neural network Instant NeRF, which glues a series of 2D photos and finishes the picture to a three-dimensional scene. It was called the "Polaroid of the future."
Read also🧐
- Google introduced the Imagen neural network. It generates pictures according to the description
- The neural network has improved the character models of retro games. Pac-Man has become like a monster from nightmares
- Sber introduced the Kandinsky neural network. You can try it on your smartphone app.