Skip to content

DreamFusion

Generate relightable 3D objects from text using 2D diffusion

Visit official site

What DreamFusion is

The method is described as a way to turn text into relightable 3D objects using a pretrained 2D diffusion model as a prior.

What you can do with it

  • Generate a 3D object from text
  • Observe the object from any angle
  • Relight it or compose it with a 3D environment

How it works

  • A Neural Radiance Field starts from random initialisation
  • Gradient descent optimises the 3D representation
  • The 2D diffusion model is used without modification

Pros & cons

✓ What we like

  • No 3D training data is required
  • The image diffusion model needs no modification

! What to watch out for

  • Generation starts from a randomly initialised 3D model
  • The approach depends on a pretrained 2D diffusion model

FAQ

Does it need 3D training data?

The material states no 3D training data is needed.

What is optimised?

It states a randomly initialised Neural Radiance Field is optimised by gradient descent.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review