Enhancing Diffusion-based Point Cloud Generation with Smoothness Constraint

TL;DR

The paper addresses non-smooth surfaces in diffusion-based point cloud generation by introducing a local smoothness constraint within a three-component diffusion framework. It pairs a learnable latent diffusion prior with a conditional diffusion decoder and leverages Tweedie’s denoising principles alongside a graph Laplacian regularizer during sampling to enforce surface smoothness. The approach is evaluated on ShapeNet against GAN- and diffusion-based baselines, showing improved geometry realism and substantially smoother surfaces, especially when the smoothness constraint is applied. The work provides a practical diffusion-based pipeline for generating high-quality, smooth point clouds with a learnable latent prior, offering potential benefits for robotics and 3D scene understanding.

Abstract

Diffusion models have been popular for point cloud generation tasks. Existing works utilize the forward diffusion process to convert the original point distribution into a noise distribution and then learn the reverse diffusion process to recover the point distribution from the noise distribution. However, the reverse diffusion process can produce samples with non-smooth points on the surface because of the ignorance of the point cloud geometric properties. We propose alleviating the problem by incorporating the local smoothness constraint into the diffusion framework for point cloud generation. Experiments demonstrate the proposed model can generate realistic shapes and smoother point clouds, outperforming multiple state-of-the-art methods.

Figures (3)

• Figure 1: The overall structure of our model, containing three modules: Encoder module $q_{\bphi}(\bz|\bX)$; The latent diffusion learns the latent prior distribution $p_{\btheta}(\bz)$; The decoder diffusion module learns a conditional diffusion model $p_{\boldsymbol{\psi}}{(\bX|\bz)}$ to reconstruct the original shape $\bX$ based on its embedding $\bz$. A smoothness constraint is incorporated during the decoder reverse diffusion process.
• Figure 2: Visualization of some generated samples by our model
• Figure 3: Sensitivity analysis on varying $K$ in KNN graph construction for airplane dataset