This paper proposes a novel diffusion-based method for 3D scene completion using a single LiDAR scan. The approach extends diffusion models (DDPMs) to operate directly on 3D point clouds, enabling scene completion at a large scale. Unlike previous methods that rely on image-based diffusion or voxel grids, this work directly operates on points, reformulating the noise and denoising process to efficiently work at scene scale. A regularization loss is introduced to stabilize the denoising process, approximating the predicted noise distribution to the real data. The method is evaluated on multiple datasets, including SemanticKITTI and KITTI-360, showing superior performance compared to state-of-the-art scene completion methods. The proposed approach achieves more detailed scene completions and is capable of generating scene-scale 3D data without relying on a fixed grid or projection. The method is also compared with non-diffusion approaches, demonstrating competitive performance. The key contributions include a novel scene-scale diffusion scheme for 3D data, a regularization technique to stabilize the denoising process, and the ability to generate more fine-grained details compared to previous methods. The results show that the proposed diffusion formulation can support further research in 3D diffusion generation.This paper proposes a novel diffusion-based method for 3D scene completion using a single LiDAR scan. The approach extends diffusion models (DDPMs) to operate directly on 3D point clouds, enabling scene completion at a large scale. Unlike previous methods that rely on image-based diffusion or voxel grids, this work directly operates on points, reformulating the noise and denoising process to efficiently work at scene scale. A regularization loss is introduced to stabilize the denoising process, approximating the predicted noise distribution to the real data. The method is evaluated on multiple datasets, including SemanticKITTI and KITTI-360, showing superior performance compared to state-of-the-art scene completion methods. The proposed approach achieves more detailed scene completions and is capable of generating scene-scale 3D data without relying on a fixed grid or projection. The method is also compared with non-diffusion approaches, demonstrating competitive performance. The key contributions include a novel scene-scale diffusion scheme for 3D data, a regularization technique to stabilize the denoising process, and the ability to generate more fine-grained details compared to previous methods. The results show that the proposed diffusion formulation can support further research in 3D diffusion generation.