Point-E

Point-E from OpenAI generates 3D point clouds from text or images in one to two minutes on a single GPU, a sharp contrast to earlier optimization-based text-to-3D methods that required many GPU hours per object. It works in two stages: a text-to-image diffusion model first produces a single synthetic view, then a point cloud diffusion model conditions on that image to output a colored 3D point cloud, with additional SDF regression models available to convert point clouds into meshes. The repository, installable via pip, includes example notebooks for image-to-point-cloud and direct text-to-point-cloud sampling, Blender rendering scripts, and P-FID and P-IS evaluation tooling with the associated datasets. The pure text-conditioned model is smaller and lower quality than the image-conditioned path, and outputs capture coarse shape and color rather than fine detail. Code and weights are open source under the MIT license, and the system serves as a common research baseline in generative 3D work.