The firm OpenIA, at the origin of the astonishing AI ChatGPT has just published a new scientific paper in which a new AI is described. “Point-E” is optimized for creating 3D objects from any query as text. Unlike ChatGPT, Point-E runs locally on your computer and everything you need is available on GitHub.
Note, however, that you may be lost if you are not used to using the Terminal. But that’s where OpenIA’s exploit lies: while ChatGPT currently requires enormous computing power and therefore runs from OpenIA’s servers, Point-E can run directly on your computer.
OpenIA: Point-E creates objects from text directly on your computer
The model is indeed concise enough to give a result in less than 2 minutes, even if your machine has only one GPU. As evidenced by the examples available in the OpenIA paper, the definition of these objects may leave something to be desired. But these objects can already be used as a basis to speed up the design of complete assets.
In recent weeks, several YouTubers had already managed to generate the code for rather impressive 3D objects by experimenting with ChatGPT – which can then be imported directly into 3D design programs such as Blender. A case that did not seem to be one of the purposes initially anticipated by the designers of ChatGPT.
However, more than creating high-definition objects, the goal of the researchers with Point-E is more, we emphasize, to make this kind of models much lighter in order to run them locally rather than in datacenters. We can imagine, for example, this kind of AI making its way into many professional graphic design software.
In particular those intended for the video game industry for the rapid generation of assets. In detail, Point-E allows you to describe with text any object you want to generate, and a corresponding 3D object is displayed less than two minutes later. The system actually combines two AIs: one capable of generating 2D images from text, and another whose role is to generate 3D objects from 2D images.
The fact that Point-E only had to base its learning on 2D images is the secret of its astonishing lightness. We note in passing that Google offers a similar AI, called DreamFusion. However, this requires a very muscular minimum configuration, and the generation takes much longer than with Point-E.
If the terms “compiling a program for your architecture” don’t scare you, you can test Point-E yourself on your computer. All the necessary instructions and files are available on the project’s GitHub page.
