Matérn Noise for Triangulation-Agnostic Flow Matching on Meshes

Researchers extended flow matching to triangle meshes with a triangulation-agnostic noise distribution based on discretized Matérn processes.

They pair this with PoissonNet denoising to generate mesh signals that transfer across different triangulations.

Experiments on elastic rest-state sampling and humanoid pose generation show strong quality and diversity.

The approach scales to very large meshes, including models with more than one million triangles.