Point clouds have increasingly become the preferred representation for various visual and graphical applications. Denoising diffusion probability models (DDPMs) have demonstrated remarkable applications in areas such as 3D point cloud generation. Furthermore, the ability to generate
or reconstruct high-resolution, high-fidelity point clouds stands as a pivotal requirement within this domain. To this end, we propose a geometry aware diffusion model for 3D shape generation. Considering that the global latent shape space overlooks the inherent topology and fine-grained information within the 3D shape itself, we introduce a geometry latent space that operates in a cascaded relationship with the global shape latent space. This latent space combines the global shape latent with the geometry latent space. To generate high-quality point
clouds, we conduct training on the geometry latent space, which yields superior results compared to training solely on the global shape latent space. Specifically, we represent the point cloud as a set of unordered point representations with positional embedding. We further employ a geometry aware attention block to model global geometric relationships and retaining local detail features, capturing the inductive bias of the 3D
geometric structure of the point cloud. We conduct experiments across various ShapeNet benchmarks, and our approach has demonstrated substantial advancements in the generation of point clouds when compared to the baseline.