In recent weeks, users have generated millions of images with artificial intelligence. Fantastic scenes, absurd combinations, extraordinary hits and terrible misses are an essential part of these models, which keep advancing. But the folks at NVIDIA Research are already working on what could be considered their next evolutionary step: GET3D, a generative system for textured 3D objects, trained from a dataset based on 2D images.
We used the prompt: «Landscape of an old forest, autumn season, brown and golden leaves, sunbeams passing through the trees, old worn out road, oil painting over canvas, highly detailed»
We also tried the Spanish version: «Landscape of an old forest, autumn season, brown and golden leaves, sunbeams passing through the trees, old worn out road, oil on canvas, high detail»
After about ten minutes and several revisions, Stable Diffusion presented this:
And it's not even trying hard, because I used a fairly generic prompt. If I had injected a more specific style or the name of an artist, the result would be even better. This is just a sample of what AI image generation can do today. It still has a lot of work ahead, especially when it comes to human anatomy (hands with seven fingers, three arms, etc.), but time and technology are on its side.
In fact, we've already discussed video generation from text, and now it's NVIDIA's turn with its GET3D project, which aims at building 3D objects and characters.
NVIDIA GET3D: From 2D to 3D with AI
One of the most interesting aspects of GET3D is that its training is based on a dataset with 2D images. All generated models adopt a popular format that allows them to be easily imported into major editing applications, either for later rendering or insertion into graphics engines. Of course, every algorithm is only as good as the training it receives, and GET3D is no exception. NVIDIA's brief demo focuses on cars (from pickups to race cars), motorcycles, chairs, people, and animals.
One of the priorities for GET3D is detail. All previous generative models suffer significant limitations in this regard, but GET3D bets on larger and more diverse datasets while optimizing quantity: NVIDIA anticipates that using a single GPU (I imagine something like an A100), GET3D can generate 20 objects per second, and its training with one million images took two days (again with A100 GPUs).
The best part? GET3D is open source. NVIDIA has already created a profile on GitHub, and will share more information in the coming days.
Official announcement: Click here
