PointDiffusion: Diffusion-Based Scene Completion in the Point Cloud Domain

2026-06-14

AI Summary

Existing LiDAR-based dense reconstruction methods struggle to meet the real-time demands of autonomous driving due to unstable global representations, supervision signals corrupted by odometry drift, and high latency from multi-step diffusion inference. To address these limitations, this work proposes a single-step latent diffusion completion model that uniquely integrates anchor-based ICP ground-truth refinement with a multi-token Gaussian variational autoencoder, further enhanced by cross-attention pooling for stable scene-level compression. The approach efficiently generates high-fidelity dense point clouds without requiring guidance, achieving a Chamfer distance of 0.024 m² on SemanticKITTI sequence 08—representing a 16-fold reduction—and outperforming LiDiff and ScoreLiDAR by 10–19% in reconstruction quality while reducing inference latency by 25–143×.

Abstract

Reconstructing dense 3D scenes from sparse LiDAR point clouds is a fundamental challenge in autonomous driving, where latent diffusion models offer a promising solution. However, existing approaches rely on object-level autoencoders that collapse into unstable global representations at outdoor scale and suffer from ground truth data corrupted by odometry drift that systematically degrades supervision quality. Furthermore, multi-step diffusion inference incurs prohibitive latency for real-time deployment. We propose a novel multi-token Gaussian VAE with cross-attention pooling for stable scene-scale LiDAR compression, combined with an anchor-based ICP ground truth refinement pipeline that eliminates drift-induced noise from training supervision. Together, these components enable a scaffold-free single-step diffusion completion model that achieves an approximately 16x reduction in squared Chamfer distance on SemanticKITTI seq. 08 (0.396 m^2 to 0.024 m^2), surpasses LiDiff and ScoreLiDAR by 17-19% and 10-11%, respectively, and operates at 25-143x lower inference latency. Our results demonstrate that data quality dominates model design in this regime and that multi-token latent spaces provide a stable first stage for latent diffusion-based scene completion.

Problem

• scene completion
• point cloud
• LiDAR
• diffusion model
• autonomous driving

Innovation

• multi-token Gaussian VAE
• cross-attention pooling
• anchor-based ICP refinement
• single-step diffusion
• LiDAR scene completion