Deformation-based 4D Gaussian splatting (4DGS) enables dynamic scene rendering. However, evaluating deformation networks for all Gaussians at every frame is compute-intensive on edge devices. We find that Gaussian trajectories are often locally predictable, yet prediction reliability varies across Gaussians and time. Thus, we propose CueSplat, a post-training framework that exploits local temporal regularity while keeping the pretrained backbone frozen and retaining its Gaussian set. Specifically, CueSplat predicts Gaussian attributes from recent history and uses a lightweight learned router to select exact deformation updates within a scene-specific budget calibrated offline. Selection accounts for prediction reliability and rendering contribution. Selected indices are reused across nearby frames to reduce routing overhead, while selected Gaussians still receive exact updates each frame. To control accumulated prediction error, we use drift signals to schedule three consecutive full-deformation frames that rebuild the prediction history. Evaluated across four mainstream 4DGS algorithms, CueSplat achieves up to 6.74× and 8.15× speedups on Jetson AGX Orin and Orin NX, respectively, with a negligible quality drop (at most 0.16 dB in dataset-average PSNR).
Figures & tables
Figure 2: Five-frame positional regularity of E-D3DGS on N3DV and Technicolor. Bars partition Gaussian–window pairs into mutually exclusive static, additional linear, additional quadratic, and outside-tolerance cases. Fits use a maximum positional-error tolerance of 0.1 times the geometric-mean canonical scale.
Figure 3: Projected Gaussian center trajectories in sear_steak [ 24 ] over frames 111–130, overlaid on the frame-130 rendering from a fixed viewpoint. The right panel magnifies the boxed region. Green marks locally linear trajectories; red marks trajectories exceeding a geometric straightness threshold.
Figure 4: Overview of CueSplat . (1) Calibration and training determine the scene budget Ks and router Rθ through oracle and router quality gates. (2) Selective inference predicts all attributes from accepted history, selects and reuses indices, and replaces selected predictions with exact outputs from frozen Dϕ . The merged state is rendered and stored in history. (3) Adaptive refresh uses drift signals to schedule three consecutive full-deformation frames that rebuild history, with partial precomputation after target timestamps are fixed.
Dataset
Method
Rendering Quality
Temporal Quality
Efficiency
PSNR ↑
SSIM ↑
LPIPS ↓
tOF ↓
tLP ×100↓
FPS ↑
Speedup ↑
FPS P5 ↑
#Inst. (106)↓
Technicolor [ 36 ]
E-D3DGS [ 1 ]
34.07
0.937
0.0783
0.127
0.557
21.97
–
21.87
2077.66
E-D3DGS + CueSplat
33.91
0.936
0.0789
0.128
0.555
72.78
3.31 ×
64.09
672.37
4DGaussians [ 43 ]
32.32
0.906
0.1062
0.172
0.596
21.68
–
21.64
1564.12
4DGaussians + CueSplat
32.27
0.905
0.1063
0.176
0.571
59.43
2.74 ×
46.44
556.57
GaGS [ 28 ]
32.99
0.927
0.0957
0.156
0.425
10.65
–
10.63
8206.42
Table 1: Dataset-average quality and efficiency of four frozen backbones with and without CueSplat . FPS and FPS P5 are measured on Jetson AGX Orin. #Inst. denotes millions of GPU warp instructions per frame from separate server-side profiling.
Figure 5: Dataset-average throughput on (a) A6000 and (b) Orin NX. Gray bars show the frozen-backbone baselines, and green bars show CueSplat . All twelve backbone–dataset pairs benefit, with peak speedups of 5.04 × on A6000 and 8.15 × on Orin NX.
Backbone
Configuration
PSNR ↑
FPS ↑
E-D3DGS [ 1 ]
Baseline
34.07
21.97
DASH-style
33.54
72.97
Ours
33.91
72.78
4DGaussians [ 43 ]
Baseline
32.32
21.68
DASH-style
32.01
59.69
Ours
32.27
59.43
Table 2: Comparison with DASH [ 4 ] on Technicolor [ 36 ] : five-scene means at 1024×544 . DASH-style is our fixed-partition control; DASH is reproduced with official code.
Variant
PSNR ↑
SSIM ↑
LPIPS ↓
FPS ↑
Baseline
32.58
0.918
0.0909
15.50
Motion-only selection
31.06
0.906
0.1018
62.50
Hold-last prediction
32.29
0.916
0.0919
60.84
Fixed Ks=20%
31.29
0.908
0.0969
67.57
Ours
32.50
0.917
0.0914
60.47
Table 3: Module ablations on 50-frame Technicolor [ 36 ] sequences, averaged equally over five scenes and four backbones. FPS is measured on Jetson AGX Orin.
Variant
PSNR ↑
SSIM ↑
LPIPS ↓
FPS ↑
FPS P5 ↑
Baseline
25.74
0.908
0.1763
17.67
17.62
No precomp.
25.71
0.908
0.1767
40.95
17.68
No refresh
25.50
0.905
0.1803
46.93
40.44
Ours
25.71
0.908
0.1767
40.81
26.20
Table 4: History refresh and precomputation ablations on 300-frame N3DV [ 24 ] sequences, averaged equally over nine scenes and four backbones. No precomp. preserves the refresh schedule and target frames. No refresh disables both refresh and precomputation. FPS and FPS P5 are measured on Jetson AGX Orin.
Figure 6: Sensitivity of E-D3DGS on N3DV. (a) Mean PSNR across prediction coefficients, with the orange outline marking the default. (b) PSNR (left axis) and AGX Orin FPS (right axis) versus the selection-reuse period N . PSNR labels are truncated to two decimal places.
Refresh frames
PSNR (dB) ↑
SSIM ↑
LPIPS ↓
FPS ↑
1
27.50
0.922
0.1184
41.83
3
27.88
0.927
0.1144
38.36
Table 5: Effect of subsequent refresh length with initialization fixed to three frames, evaluated with E-D3DGS and averaged over nine N3DV scenes. FPS is measured on AGX Orin.
Recent 4D Gaussian Splatting (4DGS) methods often fail under fast motion with large inter-frame displacements, where Gaussian attributes are poorly learned during training, and fast-moving objects are often lost from the reconstruction. In this work, we introduce Spatiotemporal Position Implicit Network for 4DGS, coined SPIN-4DGS, which learns Gaussian attributes from explicitly collected spatiotemporal positions rather than modeling temporal displacements, thereby enabling more faithful splatting under fast motions with large inter-frame displacements. To avoid the heavy memory overhead of explicitly optimizing attributes across all spatiotemporal positions, we instead predict them with a lightweight feed-forward network trained under a rasterization-based reconstruction loss. Consequently, SPIN-4DGS learns shared representations across Gaussians, effectively capturing spatiotemporal consistency and enabling stable high-quality Gaussian splatting even under challenging motions. Across extensive experiments, SPIN-4DGS consistently achieves higher fidelity under large displacements, with clear improvements in PSNR and SSIM on challenging sports scenes from the CMU Panoptic dataset. For example, SPIN-4DGS notably outperforms the strongest baseline, D3DGS, by achieving +1.83 higher PSNR on the Basketball scene.
Seung-gyeom Kim, Areum Kim, Yongjae Yoo +1
Department of Applied Artificial Intelligence, Hanyang University · Hanyang University ERICA
We present GaussianSpline, a dynamic 4D Gaussian Splatting method that models per-Gaussian motion using adaptive, non-uniform piecewise cubic Bézier splines. Unlike approaches based on global deformations, fixed-order polynomials, or uniformly parameterized splines, GaussianSpline jointly learns the Bézier control points and the temporal partition of each Gaussian's trajectory. This adaptive representation allocates greater temporal capacity to highly dynamic regions while maintaining compact trajectories for near-static regions, enabling high-fidelity reconstruction of complex and rapidly changing motion. To jointly achieve globally smooth motion and local adaptivity, we enforce G1 continuity between consecutive Bézier segments using a reflection-based constraint on their control points. This produces smooth, fully differentiable trajectories while preserving the flexibility of non-uniform temporal parameterization. Experiments on the Neural 3D Video benchmark, the Google Immersive dataset, as well as on synthetic scenes demonstrate that GaussianSpline substantially improves reconstruction quality in scenes with highly non-uniform and challenging dynamics, while maintaining competitive performance on general dynamic scenes compared with state-of-the-art 4D Gaussian Splatting methods.
Recent 4D Gaussian Splatting (4DGS) methods achieve impressive dynamic scene reconstruction but often rely on piecewise linear velocity approximations and short temporal windows. This disjointed modeling leads to severe temporal fragmentation, forcing primitives to be repeatedly eliminated and regenerated to track complex nonlinear dynamics. This makeshift approximation eliminates the long-term temporal identity of objects and causes an inevitable proliferation of Gaussians, hindering scalability to extended video sequences. To address this, we propose TRiGS, a novel 4D representation that utilizes unified, continuous geometric transformations. By integrating SE(3) transformations, hierarchical Bezier residuals, and learnable local anchors, TRiGS models geometrically consistent rigid motions for individual primitives. This continuous formulation preserves temporal identity and effectively mitigates unbounded memory growth. Extensive experiments demonstrate that TRiGS achieves high fidelity rendering on standard benchmarks while uniquely scaling to extended video sequences (e.g., 600 to 1200 frames) without severe memory bottlenecks, significantly outperforming prior works in temporal stability.
Suwoong Yeom, Joonsik Nam, Seunggyu Choi +7
Sogang University · POSTECH · Seoul National University +2