Research map / 3D vision / Gaussian splatting
1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering
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.
Prunes short-lifespan Gaussians with a spatial-temporal variation score and adds a key-frame temporal mask filter so 4DGS renders at over 1000 FPS with ~41x less storage on N3V.
Problem
4D Gaussian Splatting for dynamic scenes requires millions of Gaussians, causing large storage (about 2GB on N3V) and slow rasterization, due to temporally redundant short-lifespan Gaussians and many inactive Gaussians per frame.
Approach
Score each 4D Gaussian by a spatial contribution term times a temporal term based on the second derivative of its temporal opacity and its 4D volume, prune low scorers globally, then use key-frame visibility masks shared over adjacent frames so rendering only processes active Gaussians.
Key results
- On N3V, 4DGS-1K reaches 805 FPS and 1092 raster FPS with 418 MB storage versus vanilla 4DGS at 90 FPS, 118 raster FPS and 2085 MB, with PSNR 31.88 vs 31.91.
- With post-processing (Ours-PP), storage on N3V drops to 50 MB (about 41x smaller than 4DGS) at 31.87 PSNR.
- On D-NeRF, 4DGS-1K reaches 1462 FPS and 2482 raster FPS with 42 MB storage and 33.34 PSNR, versus 4DGS at 376 FPS, 1232 raster FPS, 278 MB and 32.99 PSNR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply 4DGS-1K pruning and temporal filtering | 4DGS · Neural 3D Video Dataset | Raster FPS: 118 -> 1092 | Helped |
| apply 4DGS-1K pruning and temporal filtering | 4DGS · Neural 3D Video Dataset | PSNR: 32.01 -> 31.88 | No clear effect |
| apply 4DGS-1K pruning and temporal filtering | 4DGS · D-NeRF Dataset | PSNR: 34.09 -> 33.34 | Helped |
| add temporal filtering to 4DGS without fine-tuning (ablation) | 4DGS · N3V | PSNR: 31.91 -> 29.56 | Hurt |
| add spatial-temporal variation score pruning (ablation) | 4DGS · N3V | PSNR: 31.91 -> 31.92 | No clear effect |
| add spatial-temporal variation score pruning (ablation) | 4DGS · N3V | Storage (MB): 2085 -> 417 | Helped |
| add post-processing to filter + pruning (ablation) | 4DGS-1K · N3V | Storage (MB): 418 -> 50 | Helped |
| use only the spatial score instead of the combined score (ablation) | 4DGS pruning · Sear Steak | PSNR: 33.76 -> 33.62 | Hurt |
| use only the spatial score instead of the combined score (ablation) | 4DGS pruning · Flame Salmon | PSNR: 28.90 -> 28.75 | Hurt |
| use only the temporal score instead of the combined score (ablation) | 4DGS pruning · Sear Steak | PSNR: 33.76 -> 33.59 | Hurt |
| replace the second derivative with the first derivative (ablation) | 4DGS pruning · Sear Steak | PSNR: 33.76 -> 33.67 | No clear effect |
| replace temporal opacity variation with Sigma_t (ablation) | 4DGS pruning · Flame Salmon | PSNR: 28.90 -> 28.71 | Hurt |
Surprising findings
- Pruning around 80% of Gaussians slightly improved PSNR on N3V (31.92 vs 31.91 for pruning alone) rather than degrading it.
- Temporal filtering without fine-tuning reduced PSNR substantially (29.56 vs 31.91), with fine-tuning recovering quality.
Limitations
- Pruning ratio is a predefined hyperparameter, and an improper ratio causes a sharp drop in rendering quality.
- Method is tailored specifically to 4DGS rather than general Gaussian-based dynamic models.
- Non-rasterization preparation overhead becomes significant after the filter speeds up rasterization.
Experimental setup
Datasets: Neural 3D Video Dataset D-NeRF Dataset
Benchmarks: Neural 3D Video Dataset D-NeRF Dataset
Metrics: PSNR SSIM LPIPS Storage (MB) FPS Raster FPS number of Gaussians
Baselines: 4DGS MEGA Compact3D 4D-RotorGS Dynamic 3DGS 4DGaussian E-D3DGS STG DyNeRF HyperReel K-Planes Deformable3DGS
Builds on: 4D Gaussian Splatting 3D Gaussian Splatting vector quantization
Closest papers by meaning
- Fully Explicit Dynamic Gaussian Splatting (NeurIPS 2024 · similarity 0.94)
- Temporal Smoothness-Aware Rate-Distortion Optimized 4D Gaussian Splatting (NeurIPS 2025 · similarity 0.93)
- 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering (CVPR 2024 · similarity 0.93)
- Efficient Gaussian Splatting for Monocular Dynamic Scene Rendering via Sparse Time-Variant Attribute Modeling (AAAI 2025 · similarity 0.92)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.92)
- FlashGS: Efficient 3D Gaussian Splatting for Large-scale and High-resolution Rendering (CVPR 2025 · similarity 0.92)
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.