This paper presents a novel approach to scene completion using diffusion models for 3D LiDAR data. The authors propose a point-level denoising diffusion probabilistic model (DDPM) to complete sparse LiDAR scans, aiming to generate more detailed and complete 3D scenes compared to existing methods. The key contributions include: 1. **Novel Diffusion Scheme**: The authors reformulate the DDPM to operate at the point level, allowing the model to learn detailed structural information from the input LiDAR scan. 2. **Noise Prediction Regularization**: A regularization loss is introduced to stabilize the noise prediction during the denoising process, approximating the predicted noise distribution closer to the expected Gaussian distribution. 3. **Refinement Network**: An additional network is trained to refine the completed scene while upsampling it, improving the quality and resolution of the generated points. The method is evaluated on the SemanticKITTI dataset and compared with state-of-the-art methods, demonstrating superior performance in terms of Chamfer distance and Jensen-Shannon divergence. The authors also show that their method can achieve scene completion on different datasets without fine-tuning, highlighting its generalizability. The proposed approach opens up new avenues for research in 3D diffusion models for scene-scale point cloud data.This paper presents a novel approach to scene completion using diffusion models for 3D LiDAR data. The authors propose a point-level denoising diffusion probabilistic model (DDPM) to complete sparse LiDAR scans, aiming to generate more detailed and complete 3D scenes compared to existing methods. The key contributions include: 1. **Novel Diffusion Scheme**: The authors reformulate the DDPM to operate at the point level, allowing the model to learn detailed structural information from the input LiDAR scan. 2. **Noise Prediction Regularization**: A regularization loss is introduced to stabilize the noise prediction during the denoising process, approximating the predicted noise distribution closer to the expected Gaussian distribution. 3. **Refinement Network**: An additional network is trained to refine the completed scene while upsampling it, improving the quality and resolution of the generated points. The method is evaluated on the SemanticKITTI dataset and compared with state-of-the-art methods, demonstrating superior performance in terms of Chamfer distance and Jensen-Shannon divergence. The authors also show that their method can achieve scene completion on different datasets without fine-tuning, highlighting its generalizability. The proposed approach opens up new avenues for research in 3D diffusion models for scene-scale point cloud data.