/point-e

Point cloud diffusion for 3D model synthesis

Primary LanguagePythonMIT LicenseMIT

Point·E

Animation of four 3D point clouds rotating

This is the official code and model release for Point-E: A System for Generating 3D Point Clouds from Complex Prompts.

Usage

Build

  1. Download repository to local workstation:
cd $HOME
git clone https://github.com/cardboardcode/point-e.git
cd ~/point-e
  1. Create a python3 virtual environment:
cd ~/point-e
virtualenv -p python3 venv
source venv/bin/activate
  1. Install with pip install -e ..
cd $HOME
pip install -e .
pip install notebook

To get started with examples, see the following notebooks:

  • image2pointcloud.ipynb - sample a point cloud, conditioned on some example synthetic view images.
  • text2pointcloud.ipynb - use our small, worse quality pure text-to-3D model to produce 3D point clouds directly from text descriptions. This model's capabilities are limited, but it does understand some simple categories and colors.
  • pointcloud2mesh.ipynb - try our SDF regression model for producing meshes from point clouds.
#image2pointcloud.ipynb
jupyter notebook point_e/examples/image2pointcloud.ipynb
#text2pointcloud.ipynb
jupyter notebook point_e/examples/text2pointcloud.ipynb
#pointcloud2mesh.ipynb
jupyter notebook point_e/examples/pointcloud2mesh.ipynb

Make sure to enable "Trust" in the Jupyter notebook interface.

For our P-FID and P-IS evaluation scripts, see:

For our Blender rendering code, see blender_script.py

Samples

You can download the seed images and point clouds corresponding to the paper banner images here.

You can download the seed images used for COCO CLIP R-Precision evaluations here.