A team of researchers has presented MeshDiffusion, a new score-based generative modeling method to generate 3D meshes. By harnessing the graph structure of meshes, this method employs a 3D diffusion model to create 3D meshes that are characterized by deformable marching tetrahedra. The team claims that MeshDiffusion has the capability to produce a wide range of realistic and diverse sets of 3D meshes, including shapes that are entirely new and not present in the training data. Additionally, it can even reconstruct the complete 3D mesh from a single 2.5D view by filling in the occluded regions.
