MesonGS: Post-training Compression of 3D Gaussians via Efficient Attribute Transformation

TL;DR

MesonGS targets the large storage footprint of 3D Gaussian Splatting by introducing a post-training codec that (i) prunes Gaussians using a joint view-dependent/view-independent importance score, (ii) transforms attributes via quaternion-to-Euler replacement and RAHT, (iii) applies block and vector quantization combined with LZ77 packing, and (iv) uses a finetune stage to restore rendering quality. The approach yields substantial compression with competitive image quality across diverse datasets, without requiring extensive retraining. The key contributions include a universal pruning criterion, RAHT-based attribute transformation, a carefully designed quantization strategy (including 8- and 16-bit settings), and a finetune regime that maintains fidelity. The results demonstrate broad practicality and efficiency, with fast encoding and robust performance across bounded and unbounded scenes, suggesting a viable path for deploying 3D-GS at scale.

Abstract

3D Gaussian Splatting demonstrates excellent quality and speed in novel view synthesis. Nevertheless, the huge file size of the 3D Gaussians presents challenges for transmission and storage. Current works design compact models to replace the substantial volume and attributes of 3D Gaussians, along with intensive training to distill information. These endeavors demand considerable training time, presenting formidable hurdles for practical deployment. To this end, we propose MesonGS, a codec for post-training compression of 3D Gaussians. Initially, we introduce a measurement criterion that considers both view-dependent and view-independent factors to assess the impact of each Gaussian point on the rendering output, enabling the removal of insignificant points. Subsequently, we decrease the entropy of attributes through two transformations that complement subsequent entropy coding techniques to enhance the file compression rate. More specifically, we first replace rotation quaternions with Euler angles; then, we apply region adaptive hierarchical transform to key attributes to reduce entropy. Lastly, we adopt finer-grained quantization to avoid excessive information loss. Moreover, a well-crafted finetune scheme is devised to restore quality. Extensive experiments demonstrate that MesonGS significantly reduces the size of 3D Gaussians while preserving competitive quality.

Figures (8)

• Figure 1: Overview of MesonGS. ① We prune insignificant Gaussians by considering both view-dependent and view-independent features. ② Geometry compression is performed using an octree to generate voxelized coordinates for future transformations. ③ We replace rotation quaternions with Euler angles. ④ Further compression is achieved by applying RAHT and ⑤ block quantization to important attributes. Notably, RAHT is not applied to the scales when the quantization bit is 8 (refer to the dashed arrow). ⑥ To significantly compress the $1^+$D SH coefficients, vector quantization is employed. ⑦ All components are packed by LZ77 gailly2003zlibziv1977universalziv1978compression codec.
• Figure 2: Quantile-quantile curve, which means $\mathbf{x}$% of least important Guasisans contributes to $\mathbf{y}$% percent of total importance. For both kinds of scenes, 40% of the Gaussians contribute over 80% of the importance. The importance refers to the contribution to the final rendering results.
• Figure 3: 2D example of RAHT.
• Figure 4: Effect of $1^+$D spherical harmonics coefficients. "FT" refers to finetune.
• Figure 5: Qualitative comparison. We can hardly observe visual artifacts on the rendering result of the compressed model compared to its original model.