Research map / 3D vision / Gaussian splatting
D-MiSo: Editing Dynamic 3D Scenes using Multi-Gaussians Soup
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 D-MiSo, a mesh-inspired dynamic Gaussian Splatting representation (Multi-Gaussians linked to triangle soups) that enables per-object editing of dynamic 3D scenes over time while matching SC-GS-level reconstruction quality.
Problem
Editing dynamic 3D Gaussian Splatting scenes over time is hard: prior editable methods such as SC-GS require manually selecting static points and adjusting control-point centroids, which is hard to reproduce and struggles with scaling and local edits.
Approach
Parameterize Gaussians as flat triangle faces (GaMeS), group them into Core-Gaussians (deformed by a deformation MLP over time) and Sub-Gaussians (attached in the Core's local frame, plus a rotation-deformation network), so that editing the Sub-Triangle Soup or the estimated mesh modifies objects directly, like a mesh.
Key results
- On D-NeRF, D-MiSo averages 42.27 PSNR versus 43.31 for SC-GS (SC-GS remains higher on average) and wins on Jumping Jacks (42.05 vs 41.13).
- On NeRF-DS, D-MiSo averages 24.0 PSNR versus 24.1 for SC-GS, with best per-object results on some scenes.
- On PanopticSports, D-MiSo averages 29.04 PSNR and 0.92 SSIM versus 28.7 PSNR and 0.91 SSIM for Dyn3DGS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use D-MiSo instead of SC-GS | dynamic scene reconstruction · D-NeRF average | PSNR: 43.31 -> 42.27 | Hurt |
| use D-MiSo instead of SC-GS | dynamic scene reconstruction · D-NeRF average | SSIM: .997 -> .993 | Hurt |
| use D-MiSo instead of SC-GS | dynamic scene reconstruction · D-NeRF average | LPIPS: .0063 -> .0065 | Hurt |
| use D-MiSo instead of SC-GS | dynamic scene reconstruction · NeRF-DS average | PSNR: 24.1 -> 24.0 | Hurt |
| use D-MiSo instead of NeRF-DS | dynamic scene reconstruction · NeRF-DS average | SSIM: .898 -> .849 | Hurt |
| use D-MiSo instead of NeRF-DS | dynamic scene reconstruction · NeRF-DS average | LPIPS: .127 -> .198 | Hurt |
| use D-MiSo instead of Dyn3DG | dynamic scene reconstruction · PanopticSports mean | PSNR: 28.7 -> 29.04 | Helped |
| use D-MiSo instead of Dyn3DG | dynamic scene reconstruction · PanopticSports mean | SSIM: 0.91 -> 0.92 | Helped |
| use D-MiSo instead of Dyn3DG | dynamic scene reconstruction · PanopticSports mean | LPIPS: 0.17 -> 0.14 | Helped |
| increase batch size from 8 to 10 (ablation) | D-MiSo · D-NeRF average | PSNR: 42.08 -> 41.58 | Hurt |
| reduce batch size from 8 to 4 (ablation) | D-MiSo · D-NeRF average | PSNR: 42.08 -> 41.16 | Hurt |
| increase batch size from 2 to 8 (ablation) | D-MiSo · NeRF-DS average | PSNR: 23.9 -> 23.6 | Hurt |
Surprising findings
- Reconstruction results are roughly on par with SC-GS, and SC-GS has higher mean PSNR on D-NeRF, so the paper's advantage is editability rather than quality.
- Batch size during training matters more on NeRF-DS than D-NeRF, with smaller batches (1-2 views) sometimes degrading PSNR.
Limitations
- Fine-grained areas not well covered in training (e.g., individual fingers) cannot be edited, because Triangle Soup fidelity depends on the fitted mesh.
- Evaluation is limited to reconstruction quality; editing quality is shown qualitatively only.
- Core-Gaussian count and Sub-Gaussian count are hyperparameters (e.g., 25 Sub-Gaussians per Core) that need tuning.
Experimental setup
Datasets: D-NeRF NeRF-DS PanopticSports
Benchmarks: D-NeRF NeRF-DS PanopticSports
Metrics: PSNR SSIM LPIPS
Baselines: SC-GS D-NeRF TiNeuVox Tensor4D K-Planes FF-NVS 4D-GS DynMF HyperNeRF NeRF-DS Dyn3DGS 3DGS
Builds on: 3D Gaussian Splatting GaMeS SC-GS anchor Gaussians
Closest papers by meaning
- GaussianEditor: Swift and Controllable 3D Editing with Gaussian Splatting (CVPR 2024 · similarity 0.91)
- SC-GS: Sparse-Controlled Gaussian Splatting for Editable Dynamic Scenes (CVPR 2024 · similarity 0.91)
- Efficient Dynamic Scene Editing via 4D Gaussian-based Static-Dynamic Separation (CVPR 2025 · similarity 0.90)
- Motion Decoupled 3D Gaussian Splatting for Dynamic Object Representation (AAAI 2025 · similarity 0.89)
- Dynamic Gaussians Mesh: Consistent Mesh Reconstruction from Dynamic Scenes (ICLR 2025 · similarity 0.89)
- LOD-GS: Achieving Levels of Detail using Scalable Gaussian Soup (CVPR 2025 · similarity 0.89)
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.