What is this?
Shap-E is OpenAI's official code and model release for generating 3D objects conditioned on text prompts or images, producing conditional 3D implicit functions.
Shap-E is OpenAI's official code and model release for generating 3D objects conditioned on text prompts or images, producing conditional 3D implicit functions.Is this a library or a ready-made app?It's a Python library/research codebase meant to be used through example Jupyter notebooks, not a packaged app with a GUI.
It's a Python library/research codebase meant to be used through example Jupyter notebooks, not a packaged app with a GUI.How do I get it running?Install it with pip install -e . from the repo, then run one of the provided notebooks like sampletextto3d.ipynb to generate a model from a text prompt.
Install it with pip install -e . from the repo, then run one of the provided notebooks like sampletextto3d.ipynb to generate a model from a text prompt.Do I need anything beyond Python to use the image-encoding notebook?For the encodemodel.ipynb notebook (encoding 3D models/point clouds into latents), you need Blender 3.3.1 or higher installed and the BLENDER_PATH environment variable set to its executable.
For the encodemodel.ipynb notebook (encoding 3D models/point clouds into latents), you need Blender 3.3.1 or higher installed and the BLENDER_PATH environment variable set to its executable.Does image-to-3D work with any photo I give it?Not out of the box well — for best results on image-conditioned generation you should remove the background from the input image first.
Not out of the box well — for best results on image-conditioned generation you should remove the background from the input image first.What kinds of outputs can the text-conditional model produce?Example samples include things like a chair shaped like an avocado, an airplane shaped like a banana, a spaceship, a birthday cupcake, and a penguin — showing it can blend object categories with creative descriptions.
Example samples include things like a chair shaped like an avocado, an airplane shaped like a banana, a spaceship, a birthday cupcake, and a penguin — showing it can blend object categories with creative descriptions.Can I feed it my own 3D model instead of generating one?Yes, the encodemodel.ipynb notebook lets you load an existing 3D model or trimesh, render multiview images and a point cloud from it, encode that into a latent, and render it back.
Yes, the encodemodel.ipynb notebook lets you load an existing 3D model or trimesh, render multiview images and a point cloud from it, encode that into a latent, and render it back.Is there a way to see more generated examples before trying it myself?Yes, the repo includes a samples.md file with random samples generated on a set of selected prompts.
Yes, the repo includes a samples.md file with random samples generated on a set of selected prompts.What are similar projects?CodeVideoModel, latent-diffusion, DALL-E, generative-models, DALLE2-pytorch, vggt, WorldSculpt, TRELLIS
CodeVideoModel, latent-diffusion, DALL-E, generative-models, DALLE2-pytorch, vggt, WorldSculpt, TRELLISWhich topics is it in?generative, 3d-generation, 3d-reconstruction, computer-graphics, jupyter-notebook, text-to-image, neural-network, 3d, image-generation, generative-ai, computer-vision, pytorch
generative, 3d-generation, 3d-reconstruction, computer-graphics, jupyter-notebook, text-to-image, neural-network, 3d, image-generation, generative-ai, computer-vision, pytorch