Shap-E

Shap-E, from OpenAI, generates 3D assets conditioned on text prompts or images. Rather than producing point clouds like its predecessor Point-E, it outputs the parameters of implicit functions that can be rendered either as textured meshes or as neural radiance fields, giving a single generation two usable representations. The model produces assets in seconds on a GPU, far faster than per-asset optimization methods, though output fidelity is lower than optimization-based pipelines achieve. The repository is the official implementation of the accompanying research paper and ships model weights plus example notebooks for text-to-3D sampling, image-to-3D sampling, which works best with the background removed from the input image, and encoding custom 3D models, a workflow that requires Blender 3.3.1 or later for multiview rendering. Released under the MIT license, it is used mainly by researchers and hobbyists exploring generative 3D and as a quick way to draft rough assets from short descriptions such as a chair shaped like an avocado.