DreamFusion

DreamFusion uses Score Distillation Sampling to generate high-fidelity 3D assets from text prompts via 2D diffusion models.

Developed by Google Research researchers in 2022, DreamFusion bypasses the need for massive 3D datasets by leveraging the Imagen 2D diffusion model. It optimizes a Neural Radiance Field (NeRF) using a specialized loss function called Score Distillation Sampling (SDS). This process allows the system to produce 360-degree geometry with consistent textures and lighting from a single text string (e.g., 'a DSLR photo of a peacock made of stained glass'). The resulting models are exportable as meshes, making them ready for integration into standard graphics pipelines.

https://dreamfusion3d.github.io/