Research map / 3D vision / Gaussian splatting
Temporal Smoothness-Aware Rate-Distortion Optimized 4D Gaussian Splatting
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 an end-to-end rate-distortion optimized compression framework for dynamic 4D Gaussian Splatting built on Ex4DGS, using pointwise Haar wavelet compression of dynamic position trajectories plus pruning, ECVQ and selective opacity quantization.
Problem
Dynamic 4D Gaussian Splatting models for volumetric video require large storage due to many Gaussians, SH parameters, and temporal redundancy, and lack an entropy-aware, bit-level rate-distortion compression framework.
Approach
Applies a single-level Haar wavelet transform to each dynamic Gaussian's position trajectory and discards detail coefficients to exploit trajectory smoothness, combined with 3DGS-style learnable masking and entropy-constrained vector quantization on top of Ex4DGS, with user-controllable RD trade-off via pruning weights.
Key results
- Achieves up to 91x compression relative to the original Ex4DGS model while keeping reasonable visual fidelity; average model size reduced to about 1.1% of Ex4DGS at the most aggressive level.
- At compression Level 1 on N3V, average size 1.26 MB with up to 163 FPS; at Level 6, 11.06 MB with 29.66 dB PSNR versus 4DGaussians at 34 MB and 28.63 dB.
- Wavelet transform ablation: at Level 1 it yields about 19% storage reduction with +0.19 dB PSNR gain; at Level 6, about 26% savings with +0.09 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a single-level Haar wavelet transform and discard high-frequency trajectory details (ablation) | RD4DGS dynamic positions · N3V, compression Level 1 | storage size: 19% storage reduction | Helped |
| Use a single-level Haar wavelet transform and discard high-frequency trajectory details (ablation) | RD4DGS dynamic positions · N3V, compression Level 1 | PSNR: +0.19 dB | Helped |
| Use a single-level Haar wavelet transform and discard high-frequency trajectory details (ablation) | RD4DGS dynamic positions · N3V, compression Level 6 | storage size: 26% storage savings | Helped |
| Use a single-level Haar wavelet transform and discard high-frequency trajectory details (ablation) | RD4DGS dynamic positions · N3V, compression Level 6 | PSNR: +0.09 dB | Helped |
| Increase Haar wavelet decomposition from one level to two or three levels (ablation) | RD4DGS dynamic positions · compression Level 1 | storage size: 1.63 MB to 1.16 MB | Helped |
| Increase Haar wavelet decomposition from one level to two or three levels (ablation) | RD4DGS dynamic positions · tested compression levels | rendering quality | Hurt |
| Quantize additional opacity-related parameters (ablation) | RD4DGS opacity · opacity quantization ablation | storage efficiency | Helped |
| Quantize variance parameters (ablation) | RD4DGS opacity · opacity quantization ablation | PSNR | Hurt |
| Reduce the rate-loss weight from the Level 6 setting (ablation) | RD4DGS · N3V, higher-fidelity trade-off ablation | PSNR | No clear effect |
| Increase the ECVQ codebook size (ablation) | RD4DGS · N3V, higher-fidelity trade-off ablation | PSNR: minor additional PSNR gain | Helped |
| Increase the ECVQ codebook size (ablation) | RD4DGS · N3V, higher-fidelity trade-off ablation | storage size: nearly doubles the model size | Hurt |
| Compress the scene with RD4DGS rather than 4DGaussians at Level 6 | N3V · compression Level 6 | PSNR: 28.63 dB -> 29.66 dB | Helped |
Surprising findings
- Applying the wavelet transform to trajectories improved PSNR as well as reducing storage at every compression level tested.
- Multi-level (2- and 3-level) wavelet decompositions reduce size further but do not surpass the RD curve of the single-level method.
Limitations
- Wavelet high-frequency filtering struggles with excessively fast motion, causing motion blur on fast-moving objects at high compression levels.
- Dynamic points still account for a substantial share of storage, leaving room for improvement.
- No error bars or statistical significance reported due to compute cost.
Experimental setup
Datasets: Neural 3D Video (N3V) Technicolor
Benchmarks: Neural 3D Video (N3V) Technicolor
Metrics: PSNR SSIM LPIPS model size (MB) FPS
Baselines: Ex4DGS 4DGaussians Light4GS RD3DGS QUEEN
Builds on: Ex4DGS 3DGS compression with RD optimization (Wang et al.) RD3DGS Haar wavelet transform
Closest papers by meaning
- 1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering (NeurIPS 2025 · similarity 0.93)
- 4DGC: Rate-Aware 4D Gaussian Compression for Efficient Streamable Free-Viewpoint Video (CVPR 2025 · similarity 0.93)
- Fully Explicit Dynamic Gaussian Splatting (NeurIPS 2024 · similarity 0.93)
- Fast Feedforward 3D Gaussian Splatting Compression (ICLR 2025 · similarity 0.93)
- 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering (CVPR 2024 · similarity 0.92)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 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.