Research map / 3D vision / Gaussian splatting
SC-GS: Sparse-Controlled Gaussian Splatting for Editable Dynamic Scenes
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SC-GS, which drives dense 3D Gaussians with sparse control points whose time-varying 6-DoF transforms come from a deformation MLP, enabling dynamic novel view synthesis and user-driven motion editing.
Problem
Novel view synthesis of dynamic scenes from monocular video with Gaussian splatting, where per-Gaussian motion is noisy, slow to train, and generalizes poorly, and where motion is not easily editable.
Approach
Decompose motion into a small set of sparse control points carrying learned 6-DoF transformations (predicted by an MLP over canonical location and time), interpolated to Gaussians with RBF/KNN linear blend skinning, regularized by an ARAP loss on trajectory-based control graphs, with adaptive control point densification and pruning; the sparse graph also supports ARAP-based motion editing.
Key results
- On D-NeRF, the method reaches 43.31 average PSNR, versus 34.01 for 4D-GS and 38.51 for the per-Gaussian deformation baseline.
- On NeRF-DS, average PSNR is 24.1 versus 23.9 for NeRF-DS and 21.5 for TiNeuVox-B, with the authors noting the gain is modest.
- Ablation on D-NeRF shows removing control points drops PSNR from 43.31 to 38.51, and removing the ARAP loss drops it to 42.62.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use sparse control points to drive Gaussian motion instead of per-Gaussian transformations | SC-GS · D-NeRF | PSNR: 38.51 -> 43.31 | Helped |
| Use sparse control points to drive Gaussian motion instead of per-Gaussian transformations | SC-GS · D-NeRF | SSIM: .992 -> .997 | Helped |
| Use sparse control points to drive Gaussian motion instead of per-Gaussian transformations | SC-GS · D-NeRF | LPIPS (VGG): 0.0162 -> 0.0063 | Helped |
| Use SC-GS rather than NeRF-DS | SC-GS · NeRF-DS | PSNR: 23.9 -> 24.1 | Helped |
| Use SC-GS rather than NeRF-DS | SC-GS · NeRF-DS | MS-SSIM: .898 -> .891 | Hurt |
| Use SC-GS rather than NeRF-DS | SC-GS · NeRF-DS | LPIPS (Alex): .127 -> .140 | Hurt |
| Add sparse control points (ablation) | SC-GS · D-NeRF | PSNR: 38.512 -> 43.307 | Helped |
| Add sparse control points (ablation) | SC-GS · D-NeRF | SSIM: 0.9922 -> 0.9976 | Helped |
| Add sparse control points (ablation) | SC-GS · D-NeRF | LPIPS: 0.0162 -> 0.0063 | Helped |
| Add ARAP loss (ablation) | SC-GS · D-NeRF | PSNR: 42.617 -> 43.307 | Helped |
| Add ARAP loss (ablation) | SC-GS · D-NeRF | SSIM: 0.9963 -> 0.9976 | Helped |
| Add ARAP loss (ablation) | SC-GS · D-NeRF | LPIPS: 0.0067 -> 0.0063 | Helped |
Surprising findings
- On NeRF-DS the method is not best on every scene and the authors note NeRF-DS outperforms it on certain datasets due to specular modeling.
Limitations
- Sensitive to inaccurate camera poses, causing failures on HyperNeRF datasets.
- Limited handling of specular effects, with little gain on NeRF-DS.
- Blurry videos with dynamic objects are not handled.
Experimental setup
Datasets: D-NeRF NeRF-DS
Benchmarks: D-NeRF NeRF-DS
Metrics: PSNR SSIM MS-SSIM LPIPS
Baselines: D-NeRF TiNeuVox Tensor4D K-Planes FF-NVS 4D-GS HyperNeRF NeRF-DS
Builds on: 3D Gaussian Splatting As-Rigid-As-Possible deformation Linear blend skinning 4D Gaussian Splatting
Closest papers by meaning
- MoDec-GS: Global-to-Local Motion Decomposition and Temporal Interval Adjustment for Compact Dynamic 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
- Motion Decoupled 3D Gaussian Splatting for Dynamic Object Representation (AAAI 2025 · similarity 0.92)
- Efficient Gaussian Splatting for Monocular Dynamic Scene Rendering via Sparse Time-Variant Attribute Modeling (AAAI 2025 · similarity 0.92)
- HAIF-GS: Hierarchical and Induced Flow-Guided Gaussian Splatting for Dynamic Scene (NeurIPS 2025 · similarity 0.91)
- Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction (CVPR 2024 · similarity 0.91)
- Spacetime Gaussian Feature Splatting for Real-Time Dynamic View Synthesis (CVPR 2024 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.