A novel approach to 3D shape generation using a VAE-like network that leverages layers from the Point Net++ architecture as an encoder network and a DeepSDF-inspired decoder network.

To install all dependencies please activate the conda environment in environment.yaml.

The training loop is implemented in pointnet_sdfvae_simple.ipynb. We already used this notebook to train a model on a larger dataset, and saved its weights in working_model.pth.

To generate new cars and visualize the latent space, see latent-visualizations/