PointDiffusion: Deconstructing point cloud analysis architectures for point diffusion generation

Recent diffusion models have demonstrated significant potential in point cloud generation. However, most existing methods rely on a point-voxel hybrid representation, which results in prohibitively high costs for both training and sampling. Fortunately, advancements in point cloud analysis networks offer efficient solutions for directly processing point clouds. In this paper, we explore a pure point diffusion paradigm (PointDiffusion) for generation by deconstructing point cloud analysis architectures. First, we divide point cloud analysis architectures into several components and explore best practices for the denoising network in diffusion models. Then, we introduce model scaling by adjusting feature channels and subsampling rates to effectively balance overhead and generation quality. Finally, we propose an α-sigmoid scheduler as the noise schedule tailored for point clouds, enabling the model to fully learn the point cloud geometries across different diffusion timesteps. We validate our PointDiffusion family on three tasks: unconditional point cloud generation, point cloud completion on synthetic datasets, and point cloud completion on real-world datasets. Experimental results demonstrate that PointDiffusion outperforms existing methods in both generation speed and quality, excelling particularly in the completion of sparse and fine-grained point clouds. For example, our PointDiffusion-S model, with just 1.2M parameters and 0.02 GFLOPS, achieves generation performance comparable to PVD, while being 22x faster. Similarly, PointDiffusion-XL, utilizing 9.7M parameters and 0.06 GFLOPS, surpasses the state-of-the-art method TIGER in generation quality, while being 5x faster. We hope our study will inspire further interest in pure point diffusion generation.