Shap-E

Shap-E is an open-source model from OpenAI, released under the MIT license, that generates 3D objects from text prompts or images. Rather than producing an implicit function like a neural radiance field only, Shap-E is trained to directly output the parameters of implicit functions that can be rendered as textured meshes or NeRFs, which lets a single trained model produce either representation. The model works in three modes: text-to-3D, which builds a shape from a written prompt (the repository's own example is "a chair that looks like an avocado"); image-to-3D, which converts a 2D image (ideally with the background removed) into a 3D asset; and model encoding, which takes an existing 3D model, renders it from multiple views, and encodes it into the same latent space used for generation. The architecture pairs an encoder that maps 3D assets into implicit-function parameters with a conditional diffusion model trained on those parameters. Shap-E is distributed as a pip-installable Python package with example Jupyter notebooks rather than a hosted app, so it targets developers and researchers comfortable running code locally, typically with a GPU. The optional model-encoding workflow requires a separate installation of Blender 3.3.1 or later.