Stable-Dreamfusion is an open-source PyTorch implementation of Google’s DreamFusion technique: it generates 3D assets from a text prompt — or from a single image — by combining a NeRF neural 3D representation with a Stable Diffusion model as the guide. The project has about 8,900 stars on GitHub; it is a research codebase whose last updates were in 2023.
DreamFusion showed that a 2D text-to-image diffusion model can supervise the optimization of a 3D scene, so no 3D training data is needed. Stable-Dreamfusion implements that score distillation sampling approach on top of an efficient instant-NGP-style NeRF backbone, letting you type a prompt and get a textured 3D object. It supports text-to-3D generation, image-to-3D reconstruction from a reference picture, and exporting the result as a mesh you can open in standard 3D tools. A graphical interface is included for interactive experimentation. In practice, generating an asset means optimizing the NeRF for a few thousand iterations under the diffusion model’s guidance, which calls for a CUDA-capable GPU and some patience, but needs no 3D training data at all. The codebase is a widely cited reference implementation in the text-to-3D literature, and many later text-to-3D projects — open-source and commercial — trace their approach back to this combination of an efficient NeRF backbone with score distillation sampling.
Stable-Dreamfusion is free and open source under the Apache-2.0 license. You run it on your own GPU.
Stable-Dreamfusion is for researchers, technical artists and 3D experimenters who want to explore text-to-3D generation with a well-known, hackable reference implementation.