Abstract
We consider the task of generating realistic 3D shapes, which is useful for avariety of applications such as automatic scene generation and physicalsimulation. Compared to other 3D representations like voxels and point clouds,meshes are more desirable in practice, because (1) they enable easy andarbitrary manipulation of shapes for relighting and simulation, and (2) theycan fully leverage the power of modern graphics pipelines which are mostlyoptimized for meshes. Previous scalable methods for generating meshes typicallyrely on sub-optimal post-processing, and they tend to produce overly-smooth ornoisy surfaces without fine-grained geometric details. To overcome theseshortcomings, we take advantage of the graph structure of meshes and use asimple yet very effective generative modeling method to generate 3D meshes.Specifically, we represent meshes with deformable tetrahedral grids, and thentrain a diffusion model on this direct parametrization. We demonstrate theeffectiveness of our model on multiple generative tasks.