This paper presents a method to efficiently train and render 3D Gaussian Splatting (3DGS) models with limited resources. 3DGS is a powerful technique for novel-view synthesis, but its high resource requirements limit its usability, especially on constrained devices. The proposed method addresses these challenges by introducing a guided, purely constructive densification process that steers densification toward Gaussians that improve reconstruction quality. The method also improves training speed by deriving faster, numerically equivalent solutions for gradient computation and attribute updates. These enhancements lead to a robust, scalable solution with reduced training times, lower compute and memory requirements, and high quality. The method achieves competitive quality metrics with 3DGS while reducing model size and training time by 4-5 times. With more generous budgets, the method surpasses 3DGS in quality. The method is particularly effective in constrained environments, such as mobile devices. The method is based on a score-based densification algorithm that uses a flexible, score-based sampling of Gaussian primitives. The method also improves the efficiency of backpropagation by parallelizing it on the projected 2D splats. The method is evaluated on several datasets and shows significant improvements in both quality and efficiency compared to existing methods. The method is particularly effective in scenarios where a fixed number of inputs is required, such as classifier networks. The method is also effective in reducing the memory footprint of 3DGS, making it more suitable for deployment on devices with limited resources. The method is a significant advancement in the field of 3DGS and has the potential to enable novel-view synthesis in a wide range of applications.This paper presents a method to efficiently train and render 3D Gaussian Splatting (3DGS) models with limited resources. 3DGS is a powerful technique for novel-view synthesis, but its high resource requirements limit its usability, especially on constrained devices. The proposed method addresses these challenges by introducing a guided, purely constructive densification process that steers densification toward Gaussians that improve reconstruction quality. The method also improves training speed by deriving faster, numerically equivalent solutions for gradient computation and attribute updates. These enhancements lead to a robust, scalable solution with reduced training times, lower compute and memory requirements, and high quality. The method achieves competitive quality metrics with 3DGS while reducing model size and training time by 4-5 times. With more generous budgets, the method surpasses 3DGS in quality. The method is particularly effective in constrained environments, such as mobile devices. The method is based on a score-based densification algorithm that uses a flexible, score-based sampling of Gaussian primitives. The method also improves the efficiency of backpropagation by parallelizing it on the projected 2D splats. The method is evaluated on several datasets and shows significant improvements in both quality and efficiency compared to existing methods. The method is particularly effective in scenarios where a fixed number of inputs is required, such as classifier networks. The method is also effective in reducing the memory footprint of 3DGS, making it more suitable for deployment on devices with limited resources. The method is a significant advancement in the field of 3DGS and has the potential to enable novel-view synthesis in a wide range of applications.