3D Gaussian Splatting (3DGS) has shown promising results in novel view synthesis, but it relies heavily on the quality of the initial point cloud, leading to blurring and needle-like artifacts in areas with insufficient initializing points. To address this issue, the authors propose Pixel-GS, a novel approach that considers the number of pixels covered by each Gaussian in each view during the computation of the growth condition. This method dynamically averages the gradients from different views based on the number of covered pixels, effectively promoting the growth of large Gaussians in areas with insufficient initializing points. Additionally, a simple strategy is introduced to scale the gradient field according to the distance to the camera, suppressing the growth of floaters near the camera. Extensive experiments on challenging datasets like Mip-NeRF 360 and Tanks & Temples demonstrate that Pixel-GS achieves state-of-the-art rendering quality while maintaining real-time speeds. The method is also shown to be more robust to the sparsity of the initial point cloud.3D Gaussian Splatting (3DGS) has shown promising results in novel view synthesis, but it relies heavily on the quality of the initial point cloud, leading to blurring and needle-like artifacts in areas with insufficient initializing points. To address this issue, the authors propose Pixel-GS, a novel approach that considers the number of pixels covered by each Gaussian in each view during the computation of the growth condition. This method dynamically averages the gradients from different views based on the number of covered pixels, effectively promoting the growth of large Gaussians in areas with insufficient initializing points. Additionally, a simple strategy is introduced to scale the gradient field according to the distance to the camera, suppressing the growth of floaters near the camera. Extensive experiments on challenging datasets like Mip-NeRF 360 and Tanks & Temples demonstrate that Pixel-GS achieves state-of-the-art rendering quality while maintaining real-time speeds. The method is also shown to be more robust to the sparsity of the initial point cloud.