Text-to-3D: Google has developed a neural network that generates 3D models from text descriptions

In Brief

Text-to-3D neural network can generate 3D models from text

DreamFusion optimizes 3D scenes based on Imagen text-to-image

2D diffusion model can be used for text-to-image synthesis

The Trust Project is a worldwide group of news organizations working to establish transparency standards.

Google created a neural network capable of creating 3D models from text descriptions. The best part is that the most difficult aspect didn’t even need to be taught. Imagen was used as the foundation for Text-to-3D.

Text-to-3D: Google has developed a neural network that generates 3D models from text descriptions

What should you know about DreamFusion?

Diffusion models trained on billions of image-text pairs have led to recent advances in text-to-image synthesis. Adapting this approach to 3D synthesis will necessitate large-scale datasets of labeled 3D assets as well as efficient denoising 3D data architectures, neither of which are currently available. In this paper, we overcome these restrictions by performing text-to-3D synthesis with a pretrained 2D text-to-image diffusion model. We present a loss based on probability density distillation that allows a 2D diffusion model to be used as a prior for optimizing a parametric picture generator. Using this loss, we use gradient descent to optimize a randomly initialized 3D model (a Neural Radiance Field or NeRF) so that its 2D renderings from random angles have a minimal loss.

The generated 3D model of the specified text can be viewed from any angle, illuminated with variable lighting, and composited into any 3D environment. Its method requires no 3D training data and no changes to the image diffusion model, illustrating the efficacy of using pretrained image diffusion models as prior.

DreamFusion makes relightable 3D models with high-fidelity appearance, depth, and normals based on a caption. Objects are represented as a Neural Radiance Field, with a pretrained text-to-image diffusion prior like Imagen being used.

Examples of Generated 3D From Text

Prompt: photo of a squirrel wearing a medieval suit of armor playing the saxophone
Prompt: photo of a squirrel wearing an elegant ballgown sitting at a pottery wheel shaping a clay bowl
Prompt: highly detailed metal sculpture of a squirrel wearing a purple hoodie riding a motorcycle
Prompt: intricate wooden carving of a squirrel wearing a medieval suit of armor wielding a katana

Putting objects together to make a scene

How does it work?

DreamFusion optimizes a 3D scene based on a caption using the Imagen text-to-image generative model. It suggests Score Distillation Sampling (SDS), which involves optimizing a loss function to produce samples from a diffusion model. As long as we can map back to images differently, SDS enables us to optimize samples in any parameter space, such as a 3D space. To define this differentiable mapping, it employs a 3D scene parameterization that is akin to Neural Radiance Fields or NeRFs. SDS alone creates a passable scene appearance, but DreamFusion enhances geometry with extra regularizers and optimization techniques. The trained NeRFs that are produced are coherent, have excellent normals, surface geometry, and depth, and can be relit using a Lambertian shading model.

Read related articles:


Any data, text, or other content on this page is provided as general market information and not as investment advice. Past performance is not necessarily an indicator of future results.

Damir Yalalov

Damir is the Editor/SEO/Product Lead at mpost.io. He is most interested in SecureTech, Blockchain, and FinTech startups. Damir earned a bachelor's degree in physics.

Follow Author

More Articles