Shap-E

Shap-E is OpenAI's open-source code for generating 3D objects from text or images. Unlike Point-E, which outputs point clouds, Shap-E creates implicit 3D functions that render as NeRFs and watertight meshes. An implicit function is a math formula that describes a 3D shape. It ships with notebooks for text-to-3D, image-to-3D, and encoding existing models. It is a research codebase for ML developers, not a hosted app. Code is MIT-licensed on GitHub. Encoding an existing model needs Blender 3.3.1 or newer. Check the repo for install steps.

Key Features

1. Text-to-3D

Sample a 3D model conditioned on a text prompt with the included notebook. Describe the object and get an implicit 3D result.

2. Image-to-3D

Sample a 3D model from a synthetic view image. Removing the image background first gives the best results.

3. Watertight Meshes

Output is implicit 3D functions, not point clouds. The resulting meshes are watertight and ready for downstream use.

4. Encode Existing Models

Load a 3D model, encode it into a latent, and render it back. This path needs Blender 3.3.1+ and the BLENDER_PATH variable.

5. Sample Gallery

The README showcases nine text-conditional samples, like a chair shaped like an avocado. More examples live in samples.md.

6. Easy Install

Installs with pip from the repo in one command. A model card documents the release for transparency.

Frequently Asked Questions (FAQ)

What is the difference between Shap-E and Point-E?

Shap-E outputs implicit 3D functions that yield watertight meshes; Point-E outputs 3D point clouds. Both are OpenAI research releases.

How do I do image-to-3D?

Use the sample_image_to_3d notebook with a view image. Remove the background first for best results.

What does encoding a model require?

Blender 3.3.1 or newer plus the BLENDER_PATH environment variable, per the README.

Can I use it commercially?

Yes. The MIT License permits commercial use, modification, and distribution.

Is there a hosted version?

No. Shap-E is a research codebase you run locally; no web app or API exists.

Conclusion

Shap-E is OpenAI's open-source route from text or images to watertight 3D meshes. Implicit-function output sets it apart from point-cloud methods. It suits ML developers and researchers building on 3D generation. The MIT license keeps it free to adapt. Try the sample notebooks to see its style of outputs. Check the GitHub repo for notebooks, samples, and the model card.