cs.CVAug 28, 2026
SaveNon-Uniform Quantisation for 3DGS Compression
Organizations: Vrije Universiteit Brussel, Brussels, Belgium · Nokia, Antwerp, Belgium
Abstract
3D Gaussian Splatting (3DGS) has emerged as a powerful technique for novel view synthesis, yet its high bitrate requirements pose significant challenges for storage and transmission. To enable practical applications and ensure interoperability within the 3DGS ecosystem, standardised compression formats are essential. In this paper, we propose a novel non-uniform quantisation scheme specifically tailored for 3DGS models. Our approach adapts to the underlying data distribution by applying importance-weighted quantisation and eliminating post-voxelisation redundancy through importance weighted merging. Extensive evaluations on benchmark datasets demonstrate that our method achieves state-of-the-art compression performance. Furthermore, the proposed scheme is compatible with any point-cloud-based representation and is intended as a formal contribution to the upcoming MPEG 3DGS compression standardisation activities.
Figures & tables
Figure 1. Overview of the proposed 3DGS quantisation pipeline. The full workflow begins (Top) with high-speed importance estimation for each primitive. This metric directly guides the subsequent non-uniform quantisation of the Gaussian positions (means). Primitives that are co-located within the resulting voxel grid are collapsed via an importance-weighted geometric merging module, combining all attributes into a single representative Gaussian. The remaining attributes are then also processed by our non-uniform quantiser (Bottom) , which adapts the Lloyd-Max algorithm (Bottom-Right) to minimise importance-weighted Mean Squared Error (WMSE) and recursively compresses the resulting reconstruction values (Bottom-Left) to achieve substantial metadata bitrate savings.
Figure 2. Visualisation of the different merging strategies for equally important source Gaussians.
| Dataset | Codec | Method | Pre-Processing [s] | Encoder [s] | Decoder [s] | Post-Processing [s] |
|---|---|---|---|---|---|---|
| MPEG Scenes | V-PCC | Uniform | 11.77 | 759.82 | 4.95 | 10.78 |
| FlexGaussian | 0.31 | 0.00 | 0.00 | 0.02 | ||
| Adaptive Voxelization | 19.40 | 405.12 | 3.50 | 7.51 | ||
| Proposed method | 116.18 | 351.63 | 4.08 | 9.20 | ||
| G-PCC | Uniform | 11.86 | 33.49 | 19.73 | 13.64 | |
| FlexGaussian | 0.31 | 0.00 | 0.00 | 0.02 |
Table 1. Comparison of the execution speed of the different approaches. The execution times of the entropy codecs are seperated from the time the quantisation and pruning takes, and is 0 for FlexGaussian which is not compatible with the lossy entropy codecs.
Figure 3. Qualitative comparison against the State-of-the-Art. One example scene is randomly picked per dataset, and a red box shows where our method has the lowest LPIPS compared to the other methods. The total bitrate usage is also shown below each image. description
Figure 4. Rate-distortion curves comparing the proposed method against the state-of-the-art, only one rate point is given for FlexGaussian ( Tian et al., 2025 ) , because it doesn’t work with lossy entropy codecs.
| Dataset | Codec | Bits | BD-Rate [%] | ||||
|---|---|---|---|---|---|---|---|
| RGB-PSNR | YUV-PSNR | YUV-SSIM | YUV-IVSSIM | LPIPS | |||
| MPEG Scenes | V-PCC | 18 | -5.63 | -4.44 | -7.93 | -0.66 | -4.98 |
| 16 | -14.44 | -13.55 | -18.51 | -10.65 | -16.26 | ||
| 14 | -28.27 | -28.04 | -27.65 | -31.54 | -30.01 | ||
| 12 | 5.75 | -30.60 | nan | -54.96 | -17.18 | ||
| G-PCC | 18 | -11.78 | 1.91 | -29.31 | -6.33 | -8.25 | |
Table 2. Ablation study on the number of bits used for voxelisation.
Figure 5. Rate-distortion curves showing the impact of the number of bits used for geometry quantisation for the proposed method.
| Dataset | Codec | Weighting | BD-Rate [%] | ||||
|---|---|---|---|---|---|---|---|
| RGB-PSNR | YUV-PSNR | YUV-SSIM | YUV-IVSSIM | LPIPS | |||
| MPEG Scenes | V-PCC | Uniform | -11.42 | -11.24 | -10.93 | -14.96 | -14.50 |
| Rendered | -21.07 | -20.42 | -18.49 | -24.66 | -22.15 | ||
| Approximate | -28.27 | -28.04 | -27.65 | -31.54 | -30.01 | ||
| G-PCC | Uniform | -29.74 | -22.89 | -52.38 | -28.12 | -35.45 | |
| Rendered | -38.14 | -30.14 | -56.08 | -34.36 | -38.83 | ||
Table 3. Ablation study on the weighting used for non-uniform quantisation and merging.
| Dataset | Codec | Method | Pre-Processing [s] | Encoder [s] | Decoder [s] | Post-Processing [s] |
|---|---|---|---|---|---|---|
| MPEG Scenes | V-PCC | Uniform | 72.67 | 479.49 | 4.98 | 11.37 |
| Rendered | 324.26 | 419.56 | 4.51 | 9.97 | ||
| Approximate | 116.18 | 351.63 | 4.08 | 9.20 | ||
| G-PCC | Uniform | 67.02 | 34.84 | 21.25 | 14.98 | |
| Rendered | 330.08 | 31.57 | 19.98 | 13.48 | ||
| Approximate | 112.04 | 30.11 | 18.95 | 11.27 |
Table 4. Execution time comparison for the ablation study on different importance weighting approaches.
| Dataset | Codec | Merging | BD-Rate [%] | ||||
|---|---|---|---|---|---|---|---|
| RGB-PSNR | YUV-PSNR | YUV-SSIM | YUV-IVSSIM | LPIPS | |||
| MPEG Scenes | V-PCC | No Merging | -10.70 | -10.25 | -11.30 | -13.34 | -14.82 |
| Basic Averaging | -9.08 | -11.03 | -17.37 | -20.93 | -16.25 | ||
| Weighted Parameter Averaging | -28.27 | -28.04 | -27.65 | -31.54 | -30.01 | ||
| Weighted Covariance | -6.26 | -8.26 | -14.96 | -21.50 | -2.94 | ||
| Weighted Log-Euclidean | -13.62 | -14.73 | -18.54 | -25.40 | -10.44 | ||
Table 5. Ablation study on the different proposed methods to merge co-located primitives.
| Dataset | Codec | Merging | # Splats | BD-Rate [%] | ||||
|---|---|---|---|---|---|---|---|---|
| RGB-PSNR | YUV-PSNR | YUV-SSIM | YUV-IVSSIM | LPIPS | ||||
| MPEG Scenes | V-PCC | No merging | 495486.5 | -10.70 | -10.25 | -11.30 | -13.34 | -14.82 |
| With Criterion | 376527.5 | -28.27 | -28.04 | -27.65 | -31.54 | -30.01 | ||
| Merging All | 375847.5 | -27.86 | -27.86 | -27.55 | -31.78 | -29.72 | ||
| G-PCC | No merging | 495486.5 | -24.06 | -19.87 | -52.35 | -22.80 | -34.83 | |
| With Criterion | 376527.5 | -48.08 | -40.78 | -59.72 | -44.50 | -44.14 | ||
Table 6. Ablation study on the dissimilarity metric used as a criterion for merging co-located Gaussians.
| Dataset | Codec | bits | BD-Rate [%] | ||||
|---|---|---|---|---|---|---|---|
| RGB-PSNR | YUV-PSNR | YUV-SSIM | YUV-IVSSIM | LPIPS | |||
| MPEG Scenes | V-PCC | 1 | -24.78 | -25.09 | -27.67 | -29.72 | -28.73 |
| 2 | -28.27 | -28.04 | -27.65 | -31.54 | -30.01 | ||
| 3 | -27.27 | -26.70 | -27.31 | -30.15 | -29.60 | ||
| 4 | -27.89 | -27.68 | -27.23 | -31.14 | -29.86 | ||
| G-PCC | 1 | -45.68 | -37.45 | -60.35 | -40.42 | -44.12 | |
Table 7. Ablation study on the effect of the number of quantisation bits for the second quantisation application.
| Dataset | Codec | Attribute | BD-Rate [%] | ||||
|---|---|---|---|---|---|---|---|
| RGB-PSNR | YUV-PSNR | YUV-SSIM | YUV-IVSSIM | LPIPS | |||
| MPEG Scenes | V-PCC | Positions | -22.30 | -22.71 | -23.24 | -29.73 | -25.16 |
| Spherical Harmonics | -6.60 | -6.53 | -0.95 | -3.41 | -4.46 | ||
| Scales | 0.78 | 0.42 | -0.22 | 0.33 | -0.04 | ||
| Rotations | 1.36 | 1.22 | -7.47 | 2.67 | -1.87 | ||
| Opacities | -0.11 | -0.50 | -1.35 | -0.35 | -1.23 | ||
Table 8. Ablation study on performance per attribute, gains are marked in bold