SpeeDe3DGS: Speedy Deformable 3D Gaussian Splatting with Temporal Pruning and Motion Grouping

TL;DR

SpeeDe3DGS tackles the efficiency–fidelity gap in dynamic 3D Gaussian Splatting by introducing Temporal Sensitivity Pruning (TSP) and Temporal Sensitivity Sampling (TSS) to prune redundant Gaussians, and GroupFlow to distill motion into grouped SE(3) transformations. TSP computes a temporally aggregated sensitivity score $\tilde{U}_{\mathcal{G}_i} \approx \nabla_{g_i}^2 L_2 \approx \sum_{\phi,t} (\nabla_{g_i} I_{\mathcal{G}_t}(\phi))^2$ to identify low-impact Gaussians, while TSS perturbs motion states with $\mathcal{X}(i)=\mathcal{N}(0,1)\cdot\beta\cdot\Delta t\cdot(1-i/\tau)$ to expose floaters. GroupFlow clusters Gaussians into $J$ groups and learns per-group SE(3) transforms $[R_j^{t}|T_j^{t}]$, applying $\boldsymbol{\mu}_i^{t} = R_j^{t}(\boldsymbol{\mu}_i^0 - h_j^0) + h_j^0 + T_j^{t}$ and $\boldsymbol{r}_i^{t} = \text{quat}(R_j^t\, \text{mat}(\boldsymbol{r}_i^0))$. Across MonoDyGauBench and related datasets, SpeeDe3DGS achieves up to $13.71\times$ faster rendering with $2.53\times$ shorter training and significantly fewer Gaussians, while maintaining or surpassing neural-field fidelity. This approach enables real-time rendering of dynamic scenes with high-quality reconstructions and provides a practical acceleration framework for neural motion representations.

Abstract

Dynamic extensions of 3D Gaussian Splatting (3DGS) achieve high-quality reconstructions through neural motion fields, but per-Gaussian neural inference makes these models computationally expensive. Building on DeformableGS, we introduce Speedy Deformable 3D Gaussian Splatting (SpeeDe3DGS), which bridges this efficiency-fidelity gap through three complementary modules: Temporal Sensitivity Pruning (TSP) removes low-impact Gaussians via temporally aggregated sensitivity analysis, Temporal Sensitivity Sampling (TSS) perturbs timestamps to suppress floaters and improve temporal coherence, and GroupFlow distills the learned deformation field into shared SE(3) transformations for efficient groupwise motion. On the 50 dynamic scenes in MonoDyGauBench, integrating TSP and TSS into DeformableGS accelerates rendering by 6.78$\times$ on average while maintaining neural-field fidelity and using 10$\times$ fewer primitives. Adding GroupFlow culminates in 13.71$\times$ faster rendering and 2.53$\times$ shorter training, surpassing all baselines in speed while preserving superior image quality.

Figures (5)

• Figure 1: Visual comparison of the baseline DeformableGS yang2023deformable3dgs and our SpeeDe3DGS methods. Pruning (TSP +TSS) and GroupFlow deliver vastly faster results. Top: as from NeRF-DS yan2023nerf. Middle: basin from NeRF-DS. Bottom: trex from D-NeRF pumarola2020dnerf.
• Figure 2: Comparison of our pruning methods on the real-world NeRF-DS yan2023nerfbell scene. Our proposed Temporal Sensitivity Pruning (TSP) and Temporal Sensitivity Sampling (TSS) methods achieve higher SSIM than the baseline DeformableGS yang2023deformable3dgs model while using $11\times$ fewer Gaussians. The left regions of the renderings appear visually identical, while the right regions show that combining TSP with TSS significantly reduces temporal flicker and floating artifacts compared to both standard pruning and the unpruned baseline.
• Figure 3: Overview of our GroupFlow method. Given a dynamic Gaussian Splatting model $\mathcal{G}$, we identify a subset of Gaussians as control points and assign each Gaussian to the control point $h_j$ with the most similar motion trajectory. The motion of each group is then estimated via a rigid transformation $[R_j^{t}|T_j^{t}]$ at each timestep, reducing inference from per-Gaussian to per-group deformation.
• Figure 4: Ablation on pruning percentages with our SpeeDe3DGS framework. We sweep soft (densification-stage) and hard (post-densification) pruning ratios in $5\%$ increments for the NeRF-DS yan2023nerf and D-NeRF pumarola2020dnerf datasets using the DeformableGS yang2023deformable3dgs codebase. Each configuration is run three times without TSS or GroupFlow, and results are averaged across all runs. ($0\%$, $0\%$) corresponds to the unpruned baseline, while the first row and column show pruning in isolation. The red dot marks our selected ($60\%$, $30\%$) soft–hard ratio. FPS and Train Time improvements are measured on an RTX A5000 GPU.
• Figure 5: Visual comparison of the baseline DeformableGS yang2023deformable3dgs and our SpeeDe3DGS methods on MonoDyGauBench (MDGB) liang2025monocular. The baseline examples are reproduced through retraining, as the original MonoDyGauBench models are not available. For all other visualizations, we use our standalone SpeeDe3DGS codebase, which produces consistently higher FPS and image quality than the MDGB wrapper. Top: slice-banana from HyperNeRF park2021hypernerf. Middle: curls from Nerfies park2021nerfies. Bottom: creeper from iPhone gao2022dynamic.