The paper introduces Compressed Gaussian Splatting (CompGS), a novel method for efficient 3D scene representation. CompGS addresses the issue of high data volume in Gaussian splatting (3DGS) by using compact Gaussian primitives, significantly reducing the size of 3D scene representations. The method employs a hybrid primitive structure where a limited set of anchor primitives captures predictive relationships with a larger set of coupled primitives, which are represented as compact residual embeddings. This structure ensures that most primitives are efficiently represented in a compact form. Additionally, a rate-constrained optimization scheme is developed to minimize redundancies within the hybrid primitives, optimizing the trade-off between bitrate consumption and representation efficacy. Experimental results on three datasets (Tanks&Templates, Deep Blending, and Mip-NeRF 360) demonstrate that CompGS achieves superior compression ratios and rendering quality compared to existing methods, making it a promising approach for practical 3D scene representation.The paper introduces Compressed Gaussian Splatting (CompGS), a novel method for efficient 3D scene representation. CompGS addresses the issue of high data volume in Gaussian splatting (3DGS) by using compact Gaussian primitives, significantly reducing the size of 3D scene representations. The method employs a hybrid primitive structure where a limited set of anchor primitives captures predictive relationships with a larger set of coupled primitives, which are represented as compact residual embeddings. This structure ensures that most primitives are efficiently represented in a compact form. Additionally, a rate-constrained optimization scheme is developed to minimize redundancies within the hybrid primitives, optimizing the trade-off between bitrate consumption and representation efficacy. Experimental results on three datasets (Tanks&Templates, Deep Blending, and Mip-NeRF 360) demonstrate that CompGS achieves superior compression ratios and rendering quality compared to existing methods, making it a promising approach for practical 3D scene representation.