Abstract
Recently, probabilistic denoising diffusion models (DDMs) have greatlyadvanced the generative power of neural networks. DDMs, inspired bynon-equilibrium thermodynamics, have not only been used for 2D imagegeneration, but can also readily be applied to 3D point clouds. However,representing 3D shapes as point clouds has a number of drawbacks, most obviousperhaps that they have no notion of topology or connectivity. Here, we explorean alternative route and introduce tetrahedral diffusion models, an extensionof DDMs to tetrahedral partitions of 3D space. The much more structured 3Drepresentation with space-filling tetrahedra makes it possible to guide andregularize the diffusion process and to apply it to colorized assets. Tomanipulate the proposed representation, we develop tetrahedral convolutions,down- and up-sampling kernels. With those operators, 3D shape generationamounts to learning displacement vectors and signed distance values on thetetrahedral grid. Our experiments confirm that Tetrahedral Diffusion yieldsplausible, visually pleasing and diverse 3D shapes, is able to handle surfaceattributes like color, and can be guided at test time to manipulate theresulting shapes.