DNGaussian is a novel framework for few-shot 3D Gaussian radiance field synthesis, designed to optimize sparse-view 3D scenes with global-local depth normalization. The method addresses the limitations of existing methods, such as high training costs and slow inference speed, by leveraging depth regularization to improve scene geometry and color appearance. Key contributions include: 1. **Hard and Soft Depth Regularization**: These techniques constrain the geometry of 3D Gaussian radiance fields by encouraging the movement of Gaussians, allowing for spatial reshaping while maintaining fine-grained color performance. 2. **Global-Local Depth Normalization**: This approach normalizes depth patches on local scales to focus on small local depth changes, enhancing the reconstruction of detailed geometry. 3. **Efficiency**: DNGaussian achieves a 25× reduction in training time and over 3000× faster rendering speed compared to state-of-the-art methods, while maintaining or improving quality. Experiments on datasets like LLFF, DTU, and Blender demonstrate that DNGaussian outperforms current methods in terms of quality, efficiency, and real-time rendering speed. The method is also robust to different monocular depth estimators and can handle complex scenes, wide-ranging views, and multiple materials.DNGaussian is a novel framework for few-shot 3D Gaussian radiance field synthesis, designed to optimize sparse-view 3D scenes with global-local depth normalization. The method addresses the limitations of existing methods, such as high training costs and slow inference speed, by leveraging depth regularization to improve scene geometry and color appearance. Key contributions include: 1. **Hard and Soft Depth Regularization**: These techniques constrain the geometry of 3D Gaussian radiance fields by encouraging the movement of Gaussians, allowing for spatial reshaping while maintaining fine-grained color performance. 2. **Global-Local Depth Normalization**: This approach normalizes depth patches on local scales to focus on small local depth changes, enhancing the reconstruction of detailed geometry. 3. **Efficiency**: DNGaussian achieves a 25× reduction in training time and over 3000× faster rendering speed compared to state-of-the-art methods, while maintaining or improving quality. Experiments on datasets like LLFF, DTU, and Blender demonstrate that DNGaussian outperforms current methods in terms of quality, efficiency, and real-time rendering speed. The method is also robust to different monocular depth estimators and can handle complex scenes, wide-ranging views, and multiple materials.