Toward a New Gaussian Splatting Compression Scheme via Region Adaptive Hierarchical Transform

3D Gaussian Splatting has recently emerged as a powerful representation for Novel View Synthesis but it comes with significant storage and bandwidth requirements, which limit its usability in practical applications. In this work, we tackle the problem of compressing Gaussian Splatting by interpreting it as a Point Cloud with rich per-point attributes, and we exploit the Region-Adaptive Hierarchical Transform (RAHT) — a lightweight and well-established tool in Point Cloud compression standards — to decorrelate and compact the attribute information. While RAHT has been successfully applied in Point Cloud compression, previous attempts to directly integrate it into Gaussian Splatting pipelines have shown limited efficiency due to the redundant and unstructured nature of Gaussian attributes. To overcome these limitations, we introduce two key contributions. First, we propose a quantization-aware fine-tuning strategy that adapts the Gaussian parameters to the transform domain, promoting sparsity in the transformed coefficients. Second, we design a systematic bit allocation strategy across attribute channels, inspired by quantization matrices in image compression. By allocating bits according to the statistical relevance and energy distribution of each attribute, our method improves rate–distortion trade-offs without requiring changes to the baseline representation. Our framework maintains full compatibility with standard Point Cloud coding tools, while reducing storage requirements. Experimental results show that, despite its simplicity, the proposed method achieves state-of-the-art performance among post-training compression approaches, and in several cases, it matches or outperforms more complex methods. This demonstrates that efficient compression can be achieved without altering the underlying representation, allowing practical deployment of Gaussian Splatting in real-world applications.