Tokenflow improves 3d point cloud generation with latent token sets
TokenFlow: Latent Token Rectified Flow Matching for 3D Point Cloud Generation
Computer Vision and Pattern Recognition
Summary
Generating 3D objects made of points is hard because keeping details and overall shape at the same time is tricky. The authors propose TokenFlow, which represents a 3D shape as many smaller pieces called latent tokens instead of one big summary. This lets the system better capture parts of the shape and how they fit together. Their method also uses clever math to keep the computation manageable. Tests show TokenFlow creates better 3D point clouds than other methods on popular datasets.
What this means in practice
- •For 3d graphics developers: Generate detailed and structurally consistent 3D point clouds for virtual objects using latent token sets to improve quality and scalability.$Commercial implications: Enables creation of high-quality 3D assets for games and VR with improved part-level control, supporting commercial content pipelines.
- •For robotics perception engineers: Use improved 3D point cloud generation to simulate diverse and realistic object shapes for robot training and testing scenarios.
Authors
Linhao Wang, Dongjin Huang
Abstract
3D point clouds generation requires capturing both part-level layout and surface detail while preserving permutation equivariance. Existing latent generative models often encode an entire shape into a single vector, which limits their ability to retain part-level structure. 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. A bottleneck of $M \ll N$ learnable tokens aggregates global context through cross-attention at $O(NM)$ cost. On ShapeNet, TokenFlow achieves the best MMD and coverage across airplane, car, and chair, while obtaining 1-NNA comparable to the strongest baselines. On ModelNet, its learned latent set reaches the best reported linear-evaluation accuracy on ModelNet10. 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.