Google previously introduced the world to Imagen, a competitor to DALL-E 2 that has some startingly great results. However, the tech giant has now taken Imagen a step further by creating a text-to-3D A.I. that can turn your words into full 3D models.
Ben Poole, a researcher at Google Brain shared the news of the new text-to-3D A.I. on Twitter. Poole says that he worked alongside Ben Mildenhall, Ajay Jain, and Jon Barron to create the AI, which relies heavily on Google’s Imagen image generator.
According to a paper that the researchers have shared on the arXiv servers, the program is a proof-of-concept. The text-to-3D A.I. image generator is currently being referred to as DreamFusion, and it’s an evolution of Dream Fields, a text-to-3D generator that Google revealed back in 2021.
The biggest difference here between the text-to-3D A.I. found in DreamFusion and Dream Fields is that Dream Fields relied on OpenAI’s CLIP tech. OpenAI has created some impressive things in the past, too, like its fake news generator that it was too afraid to release.
Now, though, the researchers are able to create 3D models simply by using a pre-trained text-to-image diffusion model, Poole explained in his tweet. This removes the need for any 3D data.
It’s unclear what kind of overall use a product like DreamFusion, or really any text-to-3D A.I. image generation might have overall. But, Google’s proof of concept is intriguing, and with more research and studying, it could prove worthwhile for expansion. Right now the results we’re seeing are extremely exciting, and will no doubt push other A.I. generators to follow a similar route in the future.
Of course, the chances of this A.I. taking over the world are slim. But it could make some parts of model generation much easier in the future.