MoRE-GS: Motion-Regime Experts for Future Gaussian Splatting

Shima Salehi; Joshua Peeples

cs.CVOct 30, 2026

MoRE-GS: Motion-Regime Experts for Future Gaussian Splatting

Authors: Shima Salehi, Joshua Peeples

Organizations: Department of Electrical and Computer Engineering, Texas A&M University, College Station, Texas, USA

Abstract

Future dynamic novel-view synthesis requires extrapolating multiple motion regimes within the same scene. A single dynamics model may not simultaneously represent stationary regions, coordinated motion, local interactions, and articulated trajectories. We introduce Motion Regime Expert Routing for Gaussian Splatting (MoRE-GS), a causal, render-aware framework for forecasting dynamic 3D Gaussian scenes. MoRE-GS freezes a Gaussian representation reconstructed exclusively from the observed temporal prefix and forecasts its control-keypoint motion using four complementary experts: persistence, a coefficient-space Transformer, a graph convolutional network, and a structured Transformer with rigid and non-rigid motion constraints. A compact routing table assigns expert mixtures across soft motion groups and forecast-horizon bins, while a persistence gate suppresses unreliable extrapolation. Across D-NeRF and HyperNeRF scenes, MoRE-GS achieves improved performance against other methods, and these results demonstrate that renderer-calibrated composition of complementary motion predictors improves causal dynamic Gaussian forecasting across heterogeneous motion regimes. Project page: https://advanced-vision-and-learning-lab.github.io/MoRE-GS-website/

Figures & tables

Figure 1: MoRE-GS forecasts control-keypoint motion using four complementary experts and an observed-only group–horizon router.

	GaussianPrediction			MoGaF			ODE-GS			MoRE-GS
Scene	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
T-Rex	21.08	0.940	0.046	20.60	0.941	0.050	21.51	0.944	0.054	21.13	0.944	0.048
Mutant	27.55	0.953	0.027	28.72	0.955	0.022	31.01	0.967	0.026	33.55	0.976	0.015
Jumping Jacks	20.00	0.912	0.084	20.12	0.922	0.073	22.23	0.920	0.080	21.18	0.925	0.069
Stand Up	27.44	0.948	0.044	30.38	0.954	0.034	27.23	0.944	0.054	30.68	0.958	0.034
Bouncing Balls	27.09	0.972	0.030	26.82	0.973	0.037	25.76	0.964	0.053	28.40	0.977	0.030

Table 1: D-NeRF 80/20 future forecasting. Best values, determined from the unrounded scores, are shown in bold.

Figure 2: Long-horizon forecasts and absolute RGB residuals on Mutant, Stand Up, and Chicken. All residual maps use the same color scale.

	ODE-GS			MoRE-GS
Scene	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Chicken	18.49	0.594	0.467	19.52	0.581	0.368
3D Printer	16.73	0.543	0.485	16.45	0.519	0.423
Cut Lemon	20.46	0.617	0.379	20.97	0.598	0.345
Split Cookie	19.65	0.630	0.264	19.95	0.638	0.259
Mean	18.83	0.596	0.399	19.22	0.584	0.349

Table 2: HyperNeRF future forecasting under the chronological 90/10 protocol. MoRE-GS uses 16 motion groups. Best values are shown in bold.

Figure 3: Near-, middle-, and long-horizon forecasts on D-NeRF Jumping Jacks and HyperNeRF Chicken.

Individual experts
Variant	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Persistence	24.70	0.936	0.053
Coefficient Transformer	26.51	0.941	0.051
GCN	25.22	0.934	0.052
Structured Transformer	25.95	0.932	0.061
Routing design

Table 3: Causal D-NeRF ablations averaged over all eight scenes. Best overall values, determined from unrounded results, are bold; the selected configuration is underlined.

Method	Params	State	Train (s)	FPS	VRAM
GaussianPrediction	14.173M	69.26 MiB	–	155.11	–
ODE-GS	1.093M	13.10 MiB	10,444	39.00	2.43 GiB
MoRE-GS	14.280M	69.75 MiB	1,135	84.83	2.15 GiB

Table 4: Post-canonical efficiency, averaged over eight D-NeRF scenes. State excludes reconstructed scenes.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: Router visualization for $G=16$ : (a) dominant soft-group assignment, with color indicating group and marker size indicating confidence; inference retains full soft assignments. (b) Group–horizon expert weights.

Figure 5: Final-horizon forecasts and absolute RGB residuals on the five D-NeRF scenes not visualized in the main paper.

Scene	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
T-Rex	$20.61\!\pm\!0.64$	$0.934\!\pm\!0.012$	$0.059\!\pm\!0.012$
Mutant	$33.19\!\pm\!0.27$	$0.975\!\pm\!0.001$	$0.015\!\pm\!<0.001$
Jumping Jacks	$21.20\!\pm\!0.10$	$0.922\!\pm\!0.003$	$0.072\!\pm\!0.002$
Stand Up	$30.36\!\pm\!0.23$	$0.956\!\pm\!0.001$	$0.035\!\pm\!0.001$
Bouncing Balls	$27.60\!\pm\!0.57$	$0.972\!\pm\!0.003$	$0.037\!\pm\!0.005$
Hook	$26.50\!\pm\!0.07$	$0.938\!\pm\!<0.001$	$0.041\!\pm\!<0.001$

Appendix

Table 5:

G=16

stability over seeds 17, 42, and 123; values are mean

\pm

standard deviation.

	Persistence			MoRE-GS
Quartile	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Q1	28.31	0.955	0.032	31.43	0.970	0.024
Q2	24.51	0.937	0.050	27.68	0.951	0.040
Q3	23.23	0.929	0.061	26.25	0.941	0.050
Q4	22.39	0.921	0.070	24.12	0.928	0.065

Appendix

Table 6: Quartile forecasts for persistence and MoRE-GS. Best values per row and metric are bold.

	Group-only			Group–horizon
Quartile	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Q1	31.88	0.968	0.024	31.43	0.970	0.024
Q2	27.58	0.950	0.042	27.68	0.951	0.040
Q3	25.77	0.938	0.053	26.25	0.941	0.050
Q4	23.79	0.926	0.067	24.12	0.928	0.065
Mean	27.38	0.946	0.046	27.47	0.948	0.045

Appendix

Table 7: Quartile forecasts for group-only and group–horizon routing. Best values per row and metric are bold.

	DCT-II			Random orthonormal
Scene	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
T-Rex	19.04	0.924	0.067	19.05	0.930	0.063
Mutant	32.70	0.969	0.016	32.31	0.968	0.017
Jumping Jacks	20.61	0.907	0.086	20.82	0.910	0.084
Stand Up	28.92	0.951	0.040	29.37	0.953	0.038
Bouncing Balls	26.47	0.970	0.043	26.99	0.973	0.039

Appendix

Table 8: Matched temporal-basis ablation. Best values per scene and metric are bold.

	GaussianPrediction			MoRE-GS
Scene	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
T-Rex	19.52	0.933	0.063	19.64	0.929	0.067
Mutant	25.31	0.941	0.037	28.97	0.957	0.024
Jumping Jacks	19.32	0.907	0.095	18.53	0.837	0.153
Stand Up	22.92	0.926	0.068	27.60	0.945	0.044
Bouncing Balls	23.17	0.955	0.052	27.73	0.977	0.032

Appendix

Table 9: D-NeRF forecasting under the matched 70/30 observed/future split. Best values per scene and metric are bold.

	GaussianPrediction			MoRE-GS
Scene	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
T-Rex	22.62	0.952	0.039	22.62	0.954	0.039
Mutant	27.90	0.956	0.023	30.39	0.966	0.017
Jumping Jacks	23.00	0.938	0.051	21.87	0.928	0.064
Stand Up	30.75	0.964	0.024	32.22	0.971	0.020
Bouncing Balls	29.90	0.980	0.020	34.04	0.988	0.015

Appendix

Table 10: D-NeRF forecasting under the matched 90/10 observed/future split. Best values per scene and metric are bold.

Dynamic Novel View Synthesis

Dynamic novel view synthesis creates time-consistent images or videos from unseen camera positions in scenes that change over time.

See all 43 papers

Space-Time Forecasting of Dynamic Scenes with Motion-aware Gaussian Grouping
Feb 25, 2026Junmyeong Lee, Hoseung Choi, Minsu ChoSpatiotemporal Forecasting 3D Gaussian Splatting
Pohang University of Science and Technology (POSTECH) · RLWRLD
On the Design of Mixture-of-Experts for Dynamic Gaussian Splatting
Jul 9, 2026In-Hwan Jin, Hyeongju Mun, Joonsoo Kim +23D Gaussian Splatting Dynamic Scene Reconstruction
Department of Electrical and Electronics Engineering, Pusan National University, Busan, Republic of Korea · Immersive Media Research Section, Electronics and Telecommunications Research Institute (ETRI), Daejeon, Republic of Korea · Department of Electrical and Electronics Engineering and JYS AI Nexus Institute, Pusan National University, Busan, Republic of Korea
4DGS-JEPA: Temporally Compositional Joint-Embedding Prediction for Dynamic Gaussian Splatting
Aug 24, 2026Yongchao Huang3D Gaussian Splatting Joint-Embedding Predictive Architecture

Feb 25, 2026J/K move · Enter open · S save

Space-Time Forecasting of Dynamic Scenes with Motion-aware Gaussian Grouping

Junmyeong Lee, Hoseung Choi, Minsu Cho

Pohang University of Science and Technology (POSTECH) · RLWRLD

Spatiotemporal Forecasting 3D Gaussian Splatting 3D Scene Representation Motion Representation Learning

Open paper AskPDF arXiv

Forecasting dynamic scenes remains a fundamental challenge in computer vision, as limited observations make it difficult to capture coherent object-level motion and long-term temporal evolution. We present Motion Group-aware Gaussian Forecasting (MoGaF), a framework for long-term scene extrapolation built upon the 4D Gaussian Splatting representation. MoGaF introduces motion-aware Gaussian grouping and group-wise optimization to enforce physically consistent motion across both rigid and non-rigid regions, yielding spatially coherent dynamic representations. Leveraging this structured space-time representation, a lightweight forecasting module predicts future motion, enabling realistic and temporally stable scene evolution. Experiments on synthetic and real-world datasets demonstrate that MoGaF consistently outperforms existing baselines in rendering quality, motion plausibility, and long-term forecasting stability. Our project page is available at https://slime0519.github.io/mogaf

Abstract

Dynamic Novel View Synthesis

Explore similar work

Space-Time Forecasting of Dynamic Scenes with Motion-aware Gaussian Grouping

On the Design of Mixture-of-Experts for Dynamic Gaussian Splatting

4DGS-JEPA: Temporally Compositional Joint-Embedding Prediction for Dynamic Gaussian Splatting