lzzcd001/MeshDiffusion

Encodes 3D meshes as deformable tetrahedral grids (DMTet) and applies diffusion-based score matching for generation, enabling both unconditional mesh synthesis and single-view completion. Integrates PyTorch3D with nvdiffrec for differentiable mesh fitting and rendering, supporting inference at 64×64×64 and 128×128×128 resolutions with pretrained models across multiple ShapeNet categories. Includes texture generation via TEXTurePaper integration and provides full training pipelines from mesh datasets through cubic grid preprocessing to diffusion model optimization.

828 stars. No commits in the last 6 months.