DreamFusion

DreamFusion is a research method for generating 3D content from text prompts by leveraging pre-trained 2D diffusion models. Developed by Google Research, it optimizes a neural radiance field (NeRF) so that rendered views of a 3D scene align with images implicitly defined by a powerful text-to-image diffusion model, such as Imagen. Instead of requiring 3D training data, DreamFusion uses score distillation sampling to convert the diffusion model’s knowledge into a 3D representation.

The approach produces textured 3D objects that can be viewed from arbitrary camera angles and relit, making it particularly useful for synthetic asset creation in graphics, games, AR/VR, and prototyping. The website presents qualitative results, technical explanations, and supplementary material, including comparisons with prior work like Dream Fields.