4D Gaussian Splatting has emerged as a new paradigm for dynamic scene representation, enabling real-time rendering of scenes with complex motions. However, it faces a major challenge of storage overhead, as millions of Gaussians are required for high-fidelity reconstruction. Compressing explicit 4D Gaussians is challenging because they exhibit heterogeneous redundancy across static and dynamic regions, and their temporal attributes make direct quantization unstable. In this work, we present OMG4 (Optimized Minimal 4D Gaussian Splatting), a framework that constructs a compact set of salient Gaussians capable of faithfully representing 4D Gaussian models. Our method progressively reduces Gaussians in three stages: (1) Gaussian Sampling to identify primitives critical to reconstruction fidelity, (2) Gaussian Pruning to remove redundancies, and (3) Gaussian Merging to fuse primitives with similar characteristics. In addition, we integrate implicit appearance compression and extend Sub-Vector Quantization (SVQ) to 4D representations with a staged quantization scheme, further reducing storage while preserving quality. Extensive experiments on standard benchmark datasets demonstrate that OMG4 achieves a favorable rate-distortion trade-off over recent state-of-the-art methods, reducing model sizes by over 60% while maintaining reconstruction quality. These results demonstrate the effectiveness of OMG4 as a practical framework for compact and high-fidelity 4D scene representation. Our source code is available at https://minshirley.github.io/OMG4/.
Figures & tables
Figure 1: The overall OMG4 pipeline and performance comparison. Top: Multi-stage compression applied to Real-Time4DGS (RT-4DGS) on Cook Spinach (N3DV dataset) and FreeTimeGS (FTGS) on Dance1 (Selfcap dataset). Bottom: Rate-distortion comparison across 4DGS methods (larger circles indicate higher FPS).
Figure 2: The overall architecture of the proposed OMG4 .
Figure 3: A comparison of rendered images. (Left) Real-Time4DGS [ 45 ] with 3.6M Gaussians. (Right) 1K optimization after Gaussian Sampling with 0.7M Gaussians.
Figure 4: Illustration of Gaussian Pruning. (Left) Space defined by Static and Dynamic score and Gaussian Sampling Boundary. (Middle) A rendered image with unique Gaussians that satisfy both unique static and dynamic thresholds, and one after 1K optimization. (Right) A rendered image with redundant Gaussians that are not included in a unique area.
Method
PSNR ↑
SSIM ↑
LPIPS (A) ↓
FPS ↑
# G ↓
Storage (MB) ↓
K-Planes [ 9 ]
31.63
–
–
0.3
–
311
HexPlane [ 4 ] ∙
31.71
–
0.075
0.2
–
250
Real-Time4DGS [ 45 ] ∗
31.96
0.946
0.051
57
3,397,510
2087
STG [ 26 ]
32.05
0.946
0.044
140
–
200
4DGS [ 41 ]
31.15
–
0.049
82
–
18.0
MEGA [ 49 ]
31.49
–
0.057
77
–
25.05
Table 1: Quantitative results on N3DV [ 25 ] at 1352×1014 resolution. Baseline results are sourced from the corresponding papers unless marked with ∗ ; ∗ denotes reproduced results. ∙ denotes results excluding the Coffee Martini scene.
Method
PSNR ↑
SSIM ↑
LPIPS (A) ↓
FPS ↑
# G ↓
Storage (MB) ↓
FTGS
32.80
0.958
0.040
532
500,000
61.04
Ours–L
32.71
0.955
0.044
346
146,891
4.06
Ours–M
32.62
0.954
0.046
354
109,191
3.07
Ours–S
32.45
0.951
0.050
374
80,673
2.32
Ours–T
32.28
0.950
0.051
379
65,706
1.93
Table 2: Quantitative results of OMG4 on FTGS (FreeTimeGS) [ 39 ] , evaluated on N3DV [ 25 ] .
Method
PSNR ↑
SSIM ↑
LPIPS (V) ↓
FPS ↑
# G ↓
Storage (MB) ↓
FTGS [ 39 ]
26.96
0.908
0.194
241
1,000,000
122
GIFStream [ 23 ] ∗ –L
25.83
0.876
0.262
94
–
7.09
GIFStream [ 23 ] ∗ –S
25.56
0.866
0.284
104
–
5.52
Ours (FTGS)–L
26.93
0.897
0.224
218
293,402
7.50
Ours (FTGS)–M
26.91
0.897
0.228
235
261,630
6.72
Ours (FTGS)–S
26.88
0.896
0.238
265
201,983
5.24
Table 3: Quantitative results on SelfCap [ 43 ] .
Setting
Method
PSNR ↑
SSIM ↑
LPIPS (V) ↓
FPS ↑
#G ↓
Storage (MB) ↓
N=1,F=1200
Real-Time4DGS [ 45 ]
OOM
FTGS [ 39 ]
24.95
0.686
0.310
317
4,000,000
488
Ours (FTGS)
24.76
0.706
0.280
140
1,999,800
52.29
N=4,F=300
Real-Time4DGS [ 45 ]
24.40
0.689
0.293
66
13,121,660
8059
FTGS [ 39 ]
25.04
0.701
0.287
328
4,000,000
488
GIFStream [ 23 ]
23.20
0.662
0.298
44
–
85.85
Table 4: Quantitative results on ENeRF-Outdoor [ 27 ] . N denotes the number of models, and F denotes the number of frames in each chunk.
Table 5: Ablation study on each module and SVQ schedule. GS , GP , GM , and AC denote Gaussian Sampling, Gaussian Pruning, Gaussian Merging, and Attribute Compression, respectively. Simul. denotes simultaneous 3D/4D SVQ, and Staged denotes the proposed staged 3D-to-4D SVQ.
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Rate-distortion curves on N3DV [ 25 ] (Top) and SelfCap [ 43 ] (Bottom).
Dataset / Variant
Baseline
BD-Rate / PSNR (%) ↓
BD-PSNR (dB) ↑
N3DV / Ours (RT-4DGS)
GIFStream [ 23 ]
−71.65
–
N3DV / Ours (RT-4DGS)
4DGS-CC (4DGS) [ 6 ]
–
+0.44
N3DV / Ours (RT-4DGS)
4DGS-CC (Saro-GS) [ 6 ]
−83.79
–
N3DV / Ours (FTGS)
4DGS-CC (4DGS) [ 6 ]
–
+1.30
SelfCap / Ours (FTGS)
GIFStream [ 23 ]
–
+1.30
Appendix
Table 6: PSNR-based BD-rate and BD-PSNR analysis. Storage size is used as the rate. Negative BD-rate indicates storage savings of OMG4 at the same PSNR, while positive BD-PSNR indicates higher PSNR at the same storage.
Dynamic Score
PSNR ↑
SSIM ↑
LPIPS ↓
Absolute acc.
31.65
0.941
0.060
Signed acc. (Ours)
31.85
0.941
0.060
Appendix
Table 7: Ablation on the Dynamic Score of SD-Score.
Variant
PSNR ↑
SSIM ↑
LPIPS ↓
Ours
31.85
0.941
0.060
+ Covariance-weighted merging
31.72
0.941
0.060
+ Temporal-weighted merging
31.75
0.941
0.060
Appendix
Table 8: Ablation on Gaussian Merging design. The proposed method learns merging weights only for position and appearance and uses position-appearance similarity for clustering.
Setting
PSNR ↑
SSIM ↑
LPIPS ↓
# G ↓
Storage (MB) ↓
p=0.6
33.56
0.956
0.046
231,984
4.78
p=0.7
33.55
0.956
0.046
192,789
4.02
p=0.8
33.45
0.955
0.048
143,880
3.09
p=0.9
32.94
0.949
0.057
82,605
1.89
Appendix
Table 9: Ablation study on the quantile level p of Gaussian Pruning.
Setting
PSNR ↑
SSIM ↑
LPIPS ↓
# G ↓
Storage (MB) ↓
M=1
33.49
0.955
0.047
164,788
3.49
M=2
33.45
0.955
0.048
143,880
3.09
M=3
33.43
0.954
0.049
134,637
2.91
M=4
33.36
0.954
0.049
129,664
2.81
Appendix
Table 10: Ablation study on the number of Gaussian Merging rounds M .
Method
PSNR ↑
SSIM ↑
LPIPS (A) ↓
# G ↓
Storage (MB) ↓
STG [ 26 ]
32.01
0.948
0.046
1,289,577
172
CSTG [ 20 ]
31.69
0.945
0.054
–
15.4
Ours–L
31.90
0.947
0.046
418,085
16.99
Ours–S
31.72
0.945
0.050
264,401
11.17
Appendix
Table 11: Quantitative results of applying OMG4 to STG [ 26 ] on N3DV [ 25 ] .
Method
PSNR ↑
SSIM ↑
LPIPS (A) ↓
FPS ↑
# G ↓
Storage (MB) ↓
Real-Time4DGS [ 45 ] ∗
32.21
0.950
0.040
96
2,770,350
1701
4DGS-1K [ 48 ]
31.87
0.944
0.053
805
666,632
49.50
Light4GS [ 28 ] ∙ –L
31.69
-
0.053
37
-
5.46
Light4GS [ 28 ] ∙ –S
31.48
-
0.064
40
-
3.77
Ours–L
32.21
0.948
0.044
162
231,537
4.77
Ours–M
32.13
0.946
0.045
178
171,965
3.63
Appendix
Table 12: Quantitative results on N3DV [ 25 ] at 1024×768 resolution. Baseline results are sourced from the corresponding papers unless marked with ∗ ; ∗ denotes reproduced results. ∙ denotes results excluding the Coffee Martini scene.
Figure 7: Effect of the sampling criterion under a fixed sampling budget. (Left) Rendered images with the full model. (Middle): Our SD Score selection using top 10% Static score and Dynamic Score. (Right) Σt -based sampling with the same sampling ratio.
Figure 8: Visualization of 2D optical flow-based Dynamic score. (Left) Predicted 2D optical flow. (Middle-Left): Optical flow thresholded with τflow . (Middle-Right) Rendered image with the Gaussians projected on high-optical flow region. (Right): Rendered image with the Gaussians of top 10% Dynamic score.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
# G ↓
Storage (MB) ↓
Optical flow
27.19
0.840
0.156
150,889
3.20
Temporal gradient (Ours)
31.75
0.939
0.062
135,684
2.93
Appendix
Table 13: Ablation on replacing the proposed Dynamic Score with optical flow.