LION: Latent Point Diffusion Models for 3D Shape Generation

The Latent Point Diffusion Model (LION) is a deep learning model that is trained to generate 3D shapes from point clouds. LION is designed as a variational autoencoder (VAE) with a hierarchical latent space that combines a global shape latent representation with a point-structured latent space.

Key features:

  • LION excels at tasks such as multimodal shape denoising and voxel-conditioned synthesis.
  • It can also be adapted for text- and image-driven 3D generation
  • It can be augmented with modern surface reconstruction techniques to generate smooth 3D meshes.

Developed by Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, NVIDIA, University of Toronto and Vector Institute





Limited Period Offer (till 31st May):

8x A100 80GB GPU server @ USD 9,500/month
Best for LLM training, fine-tuning or large-scale inference deployments
Reserve GPU instance