PointFlow

Why we included this project

PointFlow is the reference PyTorch implementation of a well-cited ICCV 2019 paper that models 3D point clouds as a hierarchy of distributions. If you are working on generative 3D geometry, say synthetic data for robotics, view synthesis, or shape completion, it is a clean baseline to compare against newer diffusion-based point cloud generators. The code implements both levels of a generative flow hierarchy in a variational framework, so you can sample whole shapes and then sample an arbitrary number of points from a given shape. It includes a demo script for visualizing outputs as point sets, plus utilities and pretraining support. It is an academic implementation from 2019, so expect to adapt dependencies and plumbing for current PyTorch versions rather than drop it into a production pipeline as-is, but that trade-off is typical and acceptable for research-grade geometry generation.