DreamFusion is a pretrained 2D text-to-image diffusion model that generates high-fidelity, controllable 3D objects. It generates 3D objects by using gradient descent to optimize randomly initialized 3D models (Neural Radiance Field), and can be viewed from any angle, re-lit with any lighting, or synthesized with any 3D environment. DreamFusion does not require 3D training data or modification of the image diffusion model, showcasing the effectiveness of using pretrained image diffusion models as priors.