Google and UC Berkley researchers have developed an AI-powered neural rendering network capable of synthesizing 3D objects from text. Called Dream Fields, this AI model generates the geometry and color of a wide range of objects without 3D supervision. Previous methods only generated objects from a handful of categories, such as ShapeNet, but Dream Fields guides generation with image-text models pre-trained on large datasets of captioned images from the web.
This method optimizes a Neural Radiance Field from multiple camera views so that the rendered images score highly with a target caption according to a pre-trained CLIP model. To maximize fidelity and visual quality, the researchers introduce simple geometric priors, including sparsity-inducing transmittance regularization, scene bounds, and new MLP architectures. What they ended up with was realistic, multi-view consistent object geometry and color from a variety of natural language captions. Google Research is also using artificial intelligence to teach robots how to move like real animals.
- Chromebook runs on Chrome OS - An operating system by Google that is built for the way we live today. It comes with built-in virus protection, updates...
- All the Google apps you know and love come standard on every Chromebook, which means you can edit, download, and convert Microsoft Office files in...
- Get access to more than 2 million Android apps from Google Play to learn and do more.
Dream Fields can be trained with diverse captions written by artists or from COCO. Descriptions control the style of generated objects, such as color and context. The compositional nature of language allows users to combine concepts in novel ways and control generation,” said the research team.
