NoiseGS: Boosting 3D Gaussian Splatting with Positional Noise

3D Gaussian Splatting (3DGS) has proven effective in rapidly rendering 3D spaces by adaptively increasing the density of isotropic Gaussians from initial points. However, the optimization of Gaussian positions remains underexplored. In particular, for larger Gaussians, pixel-wise gradient directions often vary, preventing the Gaussians from converging to appropriate positions. Such unstable Gaussians lead to blurred results in the final rendering, hindering effective reconstruction performance. To address this issue, we propose NoiseGS, a novel method that integrates a noise injection algorithm into the adaptive density control strategy. By introducing random spherical noise to the center positions of Gaussians selected for replication and splitting, NoiseGS provides an opportunity for unstable Gaussians with large gradients to densify into more optimal positions. This approach effectively mitigates the instability caused by large Gaussians and seamlessly integrates with existing heuristic Gaussian optimization techniques. Experimental results demonstrate that NoiseGS achieves remarkable generalization performance across various complex scenes.