TokenFlow: Latent Token Rectified Flow Matching for 3D Point Cloud Generation

Key points

• 3D point clouds generation requires capturing both part-level layout and surface detail while preserving permutation equivariance.1
• Existing latent generative models often encode an entire shape into a single vector, which limits their ability to retain part-level structure.1
• A bottleneck of $M \ll N$ learnable tokens aggregates global context through cross-attention at $O(NM)$ cost.1
• These results show that representing a shape as a set of latent tokens, aggregated at $O(NM)$ cost, provides an effective representation for 3D point cloud generation.1

Sources (1)

• [1]TokenFlow: Latent Token Rectified Flow Matching for 3D Point Cloud Generation
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 12:14 AM

  We propose TokenFlow, a hierarchical flow matching framework that represents each shape as a set of latent tokens and couples a rectified flow over these tokens with a conditional flow over points.

  3D point clouds generation requires capturing both part-level layout and surface detail while preserving permutation equivariance.

Extractive summary: sentences quoted from the sources.