Novel-view synthesis of dynamic scenes, crucial for AR/VR applications, remains a challenging problem. Recent methods adapt representations like 3D Gaussian Splatting (3DGS) and Neural Radiance Fields (NeRF) for dynamic scenes by incorporating time as the fourth dimension (4D representations). These 4D representations still suffer from aliasing artifacts, especially when generating novel views from divergent viewpoints (zoom-in/zoom-out operations). While using 3D smoothing filters like those proposed in Mip-Splatting might seem like a possible solution, they fail to account for local motion and also exhibit aliasing. To address this, we propose a motion-aware 3D smoothing filter specifically designed for 4D representations. Our approach adapts the filter strength based on local motion information, effectively mitigating aliasing without compromising rendering quality. This is achieved by estimating the joint density function of time and focal-to-depth ratio using a non-parametric estimation method. During inference, we sample from this joint distribution to determine the appropriate smoothing filter. This flexible strategy can be integrated with various 4D representations. Our evaluations on standard datasets demonstrate superior performance compared to state-of-the-art methods.
Figures & tables
Figure 1: We compare novel view synthesis results from our method (b) with SARO-GS (d), focusing on both dynamic regions (e.g., hand movements, cooking actions) and static regions (e.g., bottles, background). The zoomed-in regions (c) highlight that our method produces sharper details and fewer artifacts, demonstrating superior synthesis quality in motion and static areas.
Figure 2: We trained all models on training views with a resolution of 1352×1014 and rendered outputs by adjusting the focal length. Existing 4D representations exhibit significant artifacts during zoom-in and zoom-out operations. Applying static anti-aliasing filters, such as Mip-Splatting [ 67 ] , is also ineffective as motion in the scene leads to incorrect filter estimation and persistent artifacts.
Figure 3: Consider a 4D signal (represented as a 2D signal in (a)). Its different projections at timestamps t1,t2&t3 , form 3D signals (represented as 1D signals in (b), (c), and (d), respectively). These continuous 3D signals are modeled by 3D Gaussian primitives. When observed by multiple cameras at each timestamp, these primitives are band-limited by different sampling intervals. Hence, it is essential to design a motion-aware low-pass filter for dynamic scenes.
Figure 4: Analysis of the sampling rate f/d over time for the “cook-spinach” scene from the Plenoptic Video dataset [ 32 ] . Static primitives maintain a nearly constant f/d , whereas dynamic primitives vary substantially over time, as motion changes their depth. We detail how static and dynamic primitives are identified in the supplementary material.
Figure 5: Single-Scale Training and Multi-Scale Evaluation on Plenoptic Video [ 32 ] dataset. All baseline methods and our approach were trained at a resolution downscaled by a factor of 4 , while novel views are evaluated at the original. We provide zoomed-in regions highlighted by green boxes. Notice that the fine details of the martini (Row 3) being poured into the glass are not preserved in any baseline method, whereas our method successfully retains this detail.
PSNR ↑
SSIM ↑
LPIPS ↓
1× Res.
2× Res.
4× Res.
Avg.
1× Res.
2× Res.
4× Res.
Avg.
1× Res.
2× Res.
4× Res.
Avg.
SARO-GS [ 65 ]
32.75
29.40
27.40
29.85
0.956
0.913
0.884
0.918
0.086
0.167
0.234
0.160
Spacetime-GS [ 33 ]
32.50
29.20
27.09
29.60
0.957
0.918
0.892
0.922
0.087
0.169
0.232
0.163
4DGS [ 61 ]
31.46
28.30
26.36
28.71
0.946
0.901
0.881
0.909
0.095
0.177
0.230
0.167
4DRotorGS [ 10 ]
32.29
29.60
29.10
30.33
0.951
0.920
0.877
0.916
0.092
0.187
0.212
0.162
Grid4D [ 63 ]
31.03
28.13
26.64
28.60
0.799
0.773
0.765
0.779
0.095
0.177
0.234
0.169
Table 1: Single-Scale Training and Multi-Scale Evaluation on the Plenoptic Video Dataset [ 32 ] . All methods are trained at the smallest scale ( 1× ) that is 676 x 507 and evaluated across multiple scales ( 1× , 2× , and 4× ), where increasing resolution simulates zoom-in effects. While our approach performs comparably to existing methods at the training resolution ( 1× ), it demonstrates superior fidelity across all other scales, significantly outperforming prior work.
PSNR ↑
SSIM ↑
LPIPS ↓
1× Res.
2× Res.
4× Res.
Avg.
1× Res.
2× Res.
4× Res.
Avg.
1× Res.
2× Res.
4× Res.
Avg.
Deformable3DGS [ 66 ]
39.57
30.09
27.30
32.32
0.989
0.957
0.934
0.960
0.008
0.034
0.061
0.034
SARO-GS [ 65 ]
36.61
29.59
27.15
31.12
0.982
0.952
0.931
0.955
0.015
0.040
0.064
0.039
4DGS [ 61 ]
35.65
28.03
25.58
29.75
0.987
0.958
0.937
0.960
0.012
0.035
0.059
0.035
4DRotorGS [ 10 ]
32.17
26.77
24.84
27.92
0.965
0.931
0.919
0.938
0.024
0.050
0.075
0.050
Grid4D [ 63 ]
40.30
30.07
28.54
32.97
0.990
0.956
0.933
0.960
0.007
0.035
0.062
0.035
Table 2: Single-Scale Training and Multi-Scale Evaluation on the Synthetic Dataset [ 48 ] . All methods are trained on the smallest scale ( 1× ) that is 200 x 200 and evaluated across multiple scales ( 1× , 2× , and 4× ), where increasing resolution simulates zoom-in effects. Our approach outperforms baselines across all other scales.
Figure 6: Single-Scale Training and Multi-Scale Evaluation on D-NeRF [ 48 ] dataset. All baseline methods and our method were trained at 200×200 image resolution, while novel views are evaluated at 800×800 . Notice the artifacts in the baselines: the arm in “mutant” (Row 3), the arm in “hook” (Row 2), and the top part in “lego” (Row 1). In contrast, our method renders with high fidelity.
W-Temporal Error ↓
W-LPIPS ↓
W-SSIM ↑
Spacetime-GS [ 33 ]
1.127
0.003803
0.989175
4DGS [ 61 ]
0.780
0.002454
0.991685
4DRotorGS [ 10 ]
0.769
0.002653
0.992745
Grid4D [ 63 ]
0.840
0.002689
0.991504
Ours
0.608
0.001806
0.994581
Table 3: Temporal error analysis. We present a quantitative evaluation of our method’s performance against existing techniques, utilizing warped error, warped LPIPS, and warped SSIM as key metrics. Results demonstrate the superior performance of our approach.
PSNR ↑
SSIM ↑
LPIPS ↓
Deformable3DGS [ 66 ]
38.23
0.985
0.017
SARO-GS [ 65 ]
35.30
0.978
0.024
4DGS [ 61 ]
34.19
0.979
0.026
4DRotorGS [ 10 ]
34.22
0.974
0.024
Grid4D [ 63 ]
39.88
0.987
0.012
Ours
34.82
0.974
0.028
Table 4: Single-scale Training and Same-scale Testing on Synthetic Dataset [ 48 ] . We observe that the baselines perform better but our method excels in Multi-Scale Evaluation.(Tab. 2 )
PSNR ↑
SSIM ↑
LPIPS ↓
SARO-GS [ 65 ]
32.09
0.946
0.138
Spacetime-GS [ 33 ]
31.64
0.948
0.139
4DGS [ 61 ]
31.02
0.935
0.150
4DRotorGS [ 10 ]
31.10
0.937
0.145
Grid4D [ 63 ]
30.09
0.932
0.148
Ours
31.43
0.943
0.140
Table 5: Single-scale Training and Same-scale Testing on Plenoptic Video Dataset [ 32 ] . Our approach is comparable to baseline methods, with only a 0.003 SSIM difference from the base SARO-GS representation.
PSNR ↑
SSIM ↑
LPIPS ↓
SARO-GS + 3D Filter
24.92
0.840
0.270
Ours
31.40
0.941
0.117
Table 6: Quantitative comparison of our method against SARO-GS + 3D Filter . Our approach significantly outperforms the baseline. This highlights the effectiveness of our method in handling dynamic scenes.
PSNR ↑
SSIM ↑
LPIPS ↓
Train (min)
# G
SpeeDe3DGS [ 56 ]
27.83
0.940
0.067
4.45
2508
+ Ours ( † )
32.28
0.956
0.064
6.48
2330
hx
0.1
32.13
0.955
0.065
6.41
2314
0.2
32.20
0.956
0.064
6.31
2312
0.3 †
32.28
0.956
0.064
6.48
2330
ht
0.05 †
32.28
0.956
0.064
6.48
2330
Table 7: Integration with SpeeDe3DGS [ 56 ] on the D-NeRF dataset [ 48 ] (single-scale training, 4× testing). Our motion-aware filter improves SpeeDe3DGS by 4.45 dB while using fewer primitives and adding only ∼ 2 minutes of training. Lower rows ablate the bandwidths hx,ht and grid size T ( † marks the default configuration). # G is the number of Gaussian primitives.
PSNR ↑
SSIM ↑
LPIPS ↓
Eval resolution
SpeeDe3DGS
+Ours
SpeeDe3DGS
+Ours
SpeeDe3DGS
+Ours
1× ( 800×800 )
35.01
35.02
0.974
0.974
0.040
0.041
21× ( 400×400 )
32.89
35.88
0.975
0.981
0.025
0.024
41× ( 200×200 )
28.10
35.87
0.951
0.985
0.029
0.014
Table 8: Zoom-out (sub- 1× ) evaluation with SpeeDe3DGS [ 56 ] on the D-NeRF dataset [ 48 ] (averaged over 8 scenes). All models are trained at 1× ( 800×800 ) and rendered at lower resolutions. Our filter keeps quality nearly flat across scales, whereas the base degrades as the render resolution drops.
Figure 7: Comparison with SARO-GS + static 3D Filter . Incorporating the static 3D filter from Mip-Splatting [ 67 ] into SARO-GS leads to spatial blurring and temporal flickering, especially around the moving human subject. In contrast, our method maintains temporal consistency and preserves high-frequency details.
Figure 8: Qualitative Results on HyperNeRF [ 47 ] dataset. We use the “Single-Scale Training and Multi-Scale Testing” setting, where training is performed at a resolution downscaled by a factor of 4 , and novel views are rendered at the original resolution. Notice the blurriness in the “printer head” and “yellow sticker” (highlighted with a green box) in the baseline methods. In contrast, our method better preserves high-frequency details.
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 9: To segment the dynamic region in the image, we utilize SAM [ 30 ] , an online interactive tool. By providing positive clicks, we identify the relevant masks, and any outliers are removed with negative clicks.
Grid Size ( T )
25
50
100
150
PSNR ↑
1×
27.88
28.44
28.97
28.32
2×
28.93
29.54
30.05
29.41
4×
30.58
31.28
31.72
31.11
LPIPS ↓
1×
0.060
0.058
0.056
0.059
2×
0.049
0.045
0.041
0.046
4×
0.043
0.037
0.029
0.038
Appendix
Table 9: Ablation on the temporal grid size T (SARO-GS, D-NeRF [ 48 ] , single-scale training and multi-scale testing). Our default T=100 (bold) performs best across all scales.
Update Interval
100
200
300
PSNR ↑
1×
28.97
27.97
26.83
2×
30.05
29.01
27.81
4×
31.72
30.64
29.34
Appendix
Table 10: Ablation on the filter update interval (in training iterations; SARO-GS, D-NeRF [ 48 ] ). Updating every 100 iterations (bold, our default) performs best.
1× ( 800×800 )
21× ( 400×400 )
41× ( 200×200 )
Scene
B
O
B
O
B
O
bouncing balls
38.08
38.71
30.64
37.14
23.89
33.06
hell warrior
38.84
38.87
39.19
39.72
37.19
40.94
hook
33.61
33.55
33.13
34.71
28.97
35.84
jumping jacks
34.97
34.84
34.71
36.11
29.74
37.23
lego
24.65
24.66
24.33
25.20
22.37
25.82
Appendix
Table 11: Per-scene PSNR for the zoom-out (sub- 1× ) evaluation on the D-NeRF dataset [ 48 ] . B = SpeeDe3DGS, O = +Ours.
Region (distance from silhouette)
% of error reduction
41×
21×
Silhouette ring ( d=1 px, ∼1.7% of pixels)
50.7
57.8
Inside band ( d=2 – 5 px)
11.8
6.2
Deep interior ( >5 px)
20.3
25.8
Background halo ( ≤2 px)
17.0
9.4
Far background ( ∼77% of pixels)
0.2
0.7
Appendix
Table 12: Where the zoom-out error reduction comes from. Share of the total Base → Ours PSNR error reduction by distance from the object silhouette (D-NeRF, 8 scenes, all frames). The boundary band dominates while covering few pixels; the far background contributes almost nothing.
Figure 10: Zoom-out ( 41× ) error maps for the hook (top) and mutant (bottom) scenes. “Base” is SpeeDe3DGS and “Parzen” is our filter. Left to right: ground truth, base render, our render, and the two squared-error maps (shared scale). The base and our renders are visually near-identical, yet the base’s error is concentrated along object silhouettes and edges (bright), which our filter removes (near-dark error map). This explains why the large PSNR gains (e.g., +12.3 dB on mutant ) are perceptually subtle: they arise almost entirely from a thin ring of aliased edge pixels.
PSNR ↑
SSIM ↑
LPIPS ↓
1× Res.
2× Res.
4× Res.
Avg.
1× Res.
2× Res.
4× Res.
Avg.
1× Res.
2× Res.
4× Res.
Avg.
SARO-GS [ 65 ]
34.01
30.46
28.34
30.93
0.962
0.927
0.902
0.930
0.084
0.156
0.222
0.154
Spacetime-GS [ 33 ]
34.24
30.0
27.56
30.6
0.964
0.928
0.903
0.931
0.085
0.165
0.224
0.158
4DGS [ 61 ]
33.61
29.75
27.46
30.27
0.955
0.916
0.896
0.922
0.095
0.177
0.230
0.167
4DRotorGS [ 10 ]
34.26
31.19
30.79
32.08
0.966
0.933
0.940
0.946
0.085
0.187
0.226
0.166
Grid4D [ 63 ]
33.02
29.34
27.66
30.0
0.955
0.912
0.892
0.919
0.091
0.174
0.228
0.164
Appendix
Table 13: Single-Scale Training and Multi-Scale Evaluation on cook spinach scene of Plenoptic Video Dataset [ 32 ] . All methods are trained at the smallest scale ( 1× ) that is 676 x 507 and evaluated across multiple scales ( 1× , 2× , and 4× ), where increasing resolution simulates zoom-in effects. While our approach performs comparably to existing methods at the training resolution( 1× ), it demonstrates superior fidelity across all other scales, significantly outperforming prior work.
PSNR ↑
SSIM ↑
LPIPS ↓
1× Res.
2× Res.
4× Res.
Avg.
1× Res.
2× Res.
4× Res.
Avg.
1× Res.
2× Res.
4× Res.
Avg.
SARO-GS [ 65 ]
29.81
26.68
24.66
27.05
0.936
0.879
0.839
0.884
0.088
0.179
0.266
0.178
Spacetime-GS [ 33 ]
28.53
26.22
24.43
26.39
0.933
0.881
0.846
0.886
0.091
0.186
0.267
0.181
4DGS [ 61 ]
28.94
25.88
24.0
26.28
0.928
0.869
0.844
0.880
0.097
0.194
0.261
0.184
4DRotorGS [ 10 ]
29.64
27.26
27.03
27.97
0.939
0.885
0.899
0.907
0.082
0.189
0.207
0.159
Grid4D [ 63 ]
28.88
26.03
24.37
26.42
0.915
0.856
0.832
0.867
0.108
0.207
0.276
0.196
Appendix
Table 14: Single-Scale Training and Multi-Scale Evaluation on coffee martini scene of Plenoptic Video Dataset [ 32 ] . All methods are trained at the smallest scale ( 1× ) that is 676 x 507 and evaluated across multiple scales ( 1× , 2× , and 4× ), where increasing resolution simulates zoom-in effects. While our approach performs comparably to existing methods at the training resolution( 1× ), it demonstrates superior fidelity across all other scales, significantly outperforming prior work.
PSNR ↑
SSIM ↑
LPIPS ↓
1× Res.
2× Res.
4× Res.
Avg.
1× Res.
2× Res.
4× Res.
Avg.
1× Res.
2× Res.
4× Res.
Avg.
SARO-GS [ 65 ]
34.19
30.34
28.20
30.91
0.964
0.928
0.903
0.931
0.084
0.157
0.224
0.155
Spacetime-GS [ 33 ]
34.0
29.97
27.75
30.57
0.966
0.931
0.909
0.935
0.088
0.164
0.224
0.158
4DGS [ 61 ]
31.74
29.03
27.09
29.28
0.944
0.908
0.890
0.914
0.103
0.178
0.223
0.168
4DRotorGS [ 10 ]
33.15
30.44
30.17
31.25
0.949
0.914
0.917
0.926
0.109
0.184
0.207
0.167
Grid4D [ 63 ]
33.64
29.75
28.01
30.46
0.957
0.915
0.896
0.922
0.087
0.170
0.223
0.160
Appendix
Table 15: Single-Scale Training and Multi-Scale Evaluation on cut roasted beef scene of Plenoptic Video Dataset [ 32 ] . All methods are trained at the smallest scale ( 1× ) that is 676 x 507 and evaluated across multiple scales ( 1× , 2× , and 4× ), where increasing resolution simulates zoom-in effects. While our approach performs comparably to existing methods at the training resolution( 1× ), it demonstrates superior fidelity across all other scales, significantly outperforming prior work.
PSNR ↑
SSIM ↑
LPIPS ↓
1× Res.
2× Res.
4× Res.
Avg.
1× Res.
2× Res.
4× Res.
Avg.
1× Res.
2× Res.
4× Res.
Avg.
SARO-GS [ 65 ]
29.78
26.64
24.68
27.03
0.936
0.877
0.836
0.883
0.082
0.175
0.264
0.173
Spacetime-GS [ 33 ]
29.84
26.91
25.01
27.25
0.937
0.885
0.851
0.891
0.083
0.177
0.259
0.173
4DGS [ 61 ]
30.38
26.35
24.26
26.99
0.932
0.870
0.844
0.882
0.089
0.188
0.256
0.177
4DRotorGS [ 10 ]
29.90
26.54
26.60
27.68
0.931
0.867
0.805
0.867
0.091
0.193
0.263
0.182
Grid4D [ 63 ]
29.91
26.43
24.65
26.99
0.928
0.864
0.837
0.876
0.093
0.193
0.264
0.183
Appendix
Table 16: Single-Scale Training and Multi-Scale Evaluation on flame salmon scene of Plenoptic Video Dataset [ 32 ] . All methods are trained at the smallest scale ( 1× ) that is 676 x 507 and evaluated across multiple scales ( 1× , 2× , and 4× ), where increasing resolution simulates zoom-in effects. While our approach performs comparably to existing methods at the training resolution( 1× ), it demonstrates superior fidelity across all other scales, significantly outperforming prior work.
PSNR ↑
SSIM ↑
LPIPS ↓
1× Res.
2× Res.
4× Res.
Avg.
1× Res.
2× Res.
4× Res.
Avg.
1× Res.
2× Res.
4× Res.
Avg.
SARO-GS [ 65 ]
34.28
31.17
29.26
31.57
0.968
0.933
0.911
0.937
0.086
0.148
0.210
0.148
Spacetime-GS [ 33 ]
33.79
30.78
28.69
31.08
0.971
0.940
0.919
0.943
0.087
0.155
0.207
0.149
4DGS [ 61 ]
30.73
28.66
27.41
28.93
0.954
0.919
0.902
0.925
0.096
0.164
0.209
0.156
4DRotorGS [ 10 ]
33.51
31.12
29.95
31.52
0.969
0.930
0.907
0.935
0.071
0.171
0.202
0.148
Grid4D [ 63 ]
29.89
28.21
27.21
28.43
0.950
0.914
0.899
0.921
0.082
0.158
0.205
0.151
Appendix
Table 17: Single-Scale Training and Multi-Scale Evaluation on flame steak scene of Plenoptic Video Dataset [ 32 ] . All methods are trained at the smallest scale ( 1× ) that is 676 x 507 and evaluated across multiple scales ( 1× , 2× , and 4× ), where increasing resolution simulates zoom-in effects. While our approach performs comparably to existing methods at the training resolution( 1× ), it demonstrates superior fidelity across all other scales, significantly outperforming prior work.
PSNR ↑
SSIM ↑
LPIPS ↓
1× Res.
2× Res.
4× Res.
Avg.
1× Res.
2× Res.
4× Res.
Avg.
1× Res.
2× Res.
4× Res.
Avg.
SARO-GS [ 65 ]
34.45
31.12
29.25
31.60
0.967
0.935
0.914
0.938
0.091
0.154
0.217
0.154
Spacetime-GS [ 33 ]
34.59
31.31
29.10
31.66
0.971
0.941
0.921
0.944
0.088
0.157
0.212
0.152
4DGS [ 61 ]
33.33
30.14
27.95
30.47
0.962
0.926
0.908
0.932
0.092
0.162
0.207
0.153
4DRotorGS [ 10 ]
33.28
31.09
30.06
31.47
0.959
0.935
0.918
0.937
0.094
0.164
0.219
0.159
Grid4D [ 63 ]
30.81
28.99
27.91
29.23
0.956
0.922
0.909
0.929
0.099
0.160
0.207
0.155
Appendix
Table 18: Single-Scale Training and Multi-Scale Evaluation on sear steak scene of Plenoptic Video Dataset [ 32 ] . All methods are trained at the smallest scale ( 1× ) that is 676 x 507 and evaluated across multiple scales ( 1× , 2× , and 4× ), where increasing resolution simulates zoom-in effects. While our approach performs comparably to existing methods at the training resolution( 1× ), it demonstrates superior fidelity across all other scales, significantly outperforming prior work.
W-Temporal Error ↓
W-LPIPS ↓
W-SSIM ↑
1× Res.
2× Res.
4× Res.
Avg.
1× Res.
2× Res.
4× Res.
Avg.
1× Res.
2× Res.
4× Res.
Avg.
Spacetime-GS [ 33 ]
0.902
0.903
0.848
0.885
0.002952
0.002609
0.002557
0.002706
0.992751
0.992130
0.991626
0.992169
4DGS [ 61 ]
0.622
0.573
0.518
0.571
0.002215
0.001695
0.001372
0.001761
0.994050
0.994380
0.994196
0.994209
4DRotorGS [ 10 ]
0.687
0.632
0.569
0.629
0.002014
0.001372
0.002215
0.001695
0.994316
0.994404
0.994534
0.994418
Grid4D [ 63 ]
0.700
0.696
0.646
0.681
0.002480
0.001954
0.001673
0.002036
0.993439
0.993699
0.993482
0.993540
Ours
0.529
0.497
0.448
0.491
0.001685
0.001213
0.000978
0.001292
0.995878
0.995528
0.995704
0.995703
Appendix
Table 19: Temporal Stability Evaluation on sear steak scene of Plenoptic Video Dataset [ 32 ] . All methods are trained at the smallest scale ( 1× ) that is 676 x 507 and evaluated across multiple scales ( 1× , 2× , and 4× ), where increasing resolution simulates zoom-in effects.
W-Temporal Error ↓
W-LPIPS ↓
W-SSIM ↑
1× Res.
2× Res.
4× Res.
Avg.
1× Res.
2× Res.
4× Res.
Avg.
1× Res.
2× Res.
4× Res.
Avg.
Spacetime-GS [ 33 ]
1.198
1.184
1.120
1.167
0.004824
0.004476
0.004407
0.004569
0.990170
0.989600
0.989086
0.989619
4DGS [ 61 ]
1.083
1.033
1.005
1.041
0.003779
0.003297
0.003217
0.003431
0.991249
0.990875
0.990234
0.990786
4DRotorGS [ 10 ]
0.966
0.877
0.800
0.881
0.003855
0.003247
0.002917
0.003340
0.991956
0.992280
0.992499
0.992245
Grid4D [ 63 ]
0.862
0.823
0.778
0.821
0.003471
0.002891
0.002631
0.002998
0.992322
0.992116
0.992019
0.992152
Ours
0.876
0.832
0.776
0.828
0.003680
0.003232
0.003001
0.003304
0.992189
0.992245
0.992340
0.992258
Appendix
Table 20: Temporal Stability Evaluation on flame steak scene of Plenoptic Video Dataset [ 32 ] . All methods are trained at the smallest scale ( 1× ) that is 676 x 507 and evaluated across multiple scales ( 1× , 2× , and 4× ), where increasing resolution simulates zoom-in effects.
W-Temporal Error ↓
W-LPIPS ↓
W-SSIM ↑
1× Res.
2× Res.
4× Res.
Avg.
1× Res.
2× Res.
4× Res.
Avg.
1× Res.
2× Res.
4× Res.
Avg.
Spacetime-GS [ 33 ]
1.265
1.281
1.220
1.255
0.004824
0.004476
0.004407
0.004569
0.991684
0.991092
0.990413
0.991063
4DGS [ 61 ]
0.853
0.791
0.707
0.784
0.003779
0.003297
0.003217
0.003431
0.993425
0.993351
0.993226
0.993334
4DRotorGS [ 10 ]
0.929
0.871
0.803
0.868
0.003855
0.003247
0.002917
0.003340
0.993411
0.993640
0.993664
0.993572
Grid4D [ 63 ]
0.941
0.954
0.913
0.936
0.003471
0.002891
0.002631
0.002998
0.993060
0.992848
0.992499
0.992802
Ours
0.606
0.607
0.574
0.596
0.003680
0.003232
0.003001
0.003305
0.995794
0.995932
0.995966
0.995897
Appendix
Table 21: Temporal Stability Evaluation on flame salmon scene of Plenoptic Video Dataset [ 32 ] . All methods are trained at the smallest scale ( 1× ) that is 676 x 507 and evaluated across multiple scales ( 1× , 2× , and 4× ), where increasing resolution simulates zoom-in effects.
Figure 11: More Qualitative comparisons for coffee martini scene of plenoptic dataset. The differences can be observed in the flow of the coffee from the steel glass to the martini glass for dynamic region & for the static regions the things kept on the left & right side of the table.
Figure 12: More Qualitative comparisons for flame steak scene of plenoptic dataset. The differences can be observed at tip of the torch gun & the flame coming out of it for dynamic region & for static region the blinds & the things kept on the left & right side of the table.
Figure 13: More qualitative results on synthetic dataset
Figure 14: Here we show temporal plot of warping error for cook spinach scene which is trained at the smallest scale ( 1× ) that is 676 x 507 and evaluated across ( 4× ) rendered novel-views. The proposed method consistently exhibits a lower error than the baseline methods, indicating superior temporal stability.
4D Gaussian Splatting (4DGS) excels in dynamic 3D reconstruction and real-time novel view synthesis via efficient 4D Gaussian representations and parallelizable rendering. However, existing 4DGS approaches rely on a single polynomial to model motion, which limits performance in complex dynamic scenes where high-frequency motion components are prevalent, and fails to ensure long-term stability due to cumulative trajectory drift. To address these issues, we propose a Fourier Motion Modeling module: this paradigm decomposes motion into frequency-based sinusoidal components, capturing both low-frequency global trajectories and high-frequency local details to model complex motion patterns accurately. It retains the real-time rendering capability of 4DGS while improving complex motion fitting and long-term coherence. Additionally, we integrate a motion-aware regularization strategy into the loss function: it uses frequency-dependent weights to suppress high-frequency jitter while preserving low-frequency motion coherence. Extensive experiments on N3V and Google Immersive datasets from multiple scenarios demonstrate the effectiveness of our method.
Zhengyang Zhang, Ziyu Lu, PengCheng Li +6
1Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · Central Media Technology Institute, Huawei
4D Gaussian Splatting (4DGS) can render dynamic scenes photorealistically. However, with limited viewpoint coverage, some spatiotemporal regions remain sparsely observed, leading to artifacts, particularly in scenes with large motion. Existing approaches leveraging generative models rely on heuristic virtual-viewpoint selection before refining rendered views. As a result, they cannot actively explore such sparsely observed regions. To address this issue, we propose a pipeline that actively selects spatiotemporal virtual viewpoints to improve 4DGS reconstruction. Our method selects virtual viewpoints for generative enhancement based on the rendering sensitivity and motion-aware observation density of 4D Gaussians, prioritizing views that alleviate observation sparsity. In the refined images, we filter out regions that conflict with captured observations or are likely to contain generative artifacts and then fine-tune 4DGS using only the reliable regions. We evaluate our method on multi-view video benchmarks using new train/test splits designed to induce observation gaps. Results show consistent improvements over prior viewpoint selection strategies and fine-tuning methods in both qualitative and quantitative evaluations, while reducing artifacts.
4D scene synthesis from monocular videos has made significant progress in recent years. However, existing methods are typically constrained by view interpolation. As a result, they struggle to infer unseen regions beyond the observed views. In this paper, we reformulate the task as 4D scene synthesis with unseen regions, which extends beyond traditional interpolation settings. Camera-conditioned video generation enables unseen region synthesis by guiding generation along specified cameras. However, these methods lack explicit 3D priors and are optimized with random camera trajectories. This design leads to severe inconsistencies under large trajectory deviations. To address this limitation, we build a unified training and inference framework with panoramic trajectory guidance. While this design improves cross-view consistency, the panoramic representation alone fails to model dynamic content effectively. Object motion in panoramic space introduces scale and shape distortions. To address this, we propose PanoGaussian, a unified Panoramic-Gaussian representation that distills the panoramic representation into an explicit dynamic Gaussian representation to capture dynamic physical priors of the 4D scene. Experiments demonstrate that PanoGaussian achieves consistent 4D scene synthesis even under large viewpoint variations.
Yuankun Yang, Yi Wei, Wenyang Zhou +1
School of Data Science, Fudan University · 2Central Media Technology Institute, Huawei