DB-GS: A Diffusion-Based 3D Gaussian Splatting Method for Scene Reconstruction

D reconstruction is widely utilized in modern intelligent systems, with neural networks emerging as a key technique in this domain. However, achieving both high fidelity and efficiency in pixel-level 3D scene reconstruction remains a considerable challenge. Current approaches often rely heavily on traditional Structure-from-Motion (SfM) for scene initialization, followed by deep learning-based optimization. Moreover, using purely image-based inputs frequently results in artifacts and aliasing. To address these limitations, this paper introduces a novel Diffusion-Based 3D Gaussian Splatting (DB-GS) method. This method replaces the SfM initialization with randomly initialized Gaussian ellipsoids of large variance, employs a diffusion-based splitting strategy to refine Gaussian points through backpropagation gradients, and integrates a low-pass filter grounded in the Nyquist-Shannon sampling theorem to effectively reduce artifacts. Experimental results on the official dataset highlight the superior performance of DB-GS, with the proposed TD-NeRF method outperforming several existing 3D Gaussian Splatting (3DGS) methods under random initialization conditions.