Research map / 3D vision / Gaussian splatting
Grid4D: 4D Decomposed Hash Encoding for High-Fidelity Dynamic 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.
Grid4D decomposes 4D space-time input into one spatial and three temporal 3D hash encodings with a directional attention decoder and a feature-space smoothness regularizer, for Gaussian-splatting dynamic scene rendering.
Problem
Plane-based explicit deformation fields for dynamic Gaussian splatting rely on a low-rank assumption and over-decompose 4D space-time encodings, causing feature overlap and weaker rendering quality; explicit grids also lack smoothness.
Approach
Decompose the 4D (x,y,z,t) encoding into one spatial 3D hash grid and three temporal 3D hash grids (no low-rank assumption), aggregate them with a directional attention score in (-1,1) generated from spatial features, and regularize hash features with a neighborhood-consistency loss.
Key results
- On D-NeRF, Grid4D reaches mean PSNR 42.00 / SSIM 0.994 / LPIPS 0.008, versus 41.65 PSNR for SC-GS and 40.16 for DeformGS.
- On HyperNeRF interpolation split, Grid4D achieves 28.56 PSNR versus 27.54 for 4D-GS; on the rig split it gets 25.50 versus 24.99.
- Grid4D renders faster than 4D-GS and DeformGS in the reported FPS table on most D-NeRF scenes, but training time (55 min) is longer than 4D-GS (20 min) and shorter than DeformGS (33 min) and SC-GS (75 min).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Grid4D instead of SC-GS | D-NeRF · D-NeRF mean | PSNR: 41.65 -> 42.00 | Helped |
| use Grid4D instead of SC-GS | D-NeRF · D-NeRF mean | LPIPS: 0.009 -> 0.008 | Helped |
| replace four 3D hash encodings with a 4D hyper-grid hash encoding (ablation) | Grid4D · D-NeRF | PSNR: 42.00 -> 28.45 | Hurt |
| add smooth regularization (ablation) | Grid4D · D-NeRF | PSNR: 39.47 -> 42.00 | Helped |
| use attention rather than direct concatenation decoding (ablation) | Grid4D · D-NeRF | PSNR: 41.37 -> 42.00 | Helped |
| use directional attention scores in (-1, 1) rather than scores in (0, 1) (ablation) | Grid4D · D-NeRF | PSNR: 41.32 -> 42.00 | Helped |
| set temporal hash encoder max level L to 8 instead of 32 (ablation) | Grid4D · D-NeRF | PSNR: 41.34 -> 42.00 | Hurt |
| set temporal hash encoder max level L to 16 instead of 32 (ablation) | Grid4D · D-NeRF | PSNR: 41.69 -> 42.00 | Hurt |
| set multi-head decoder depth to one (ablation) | Grid4D · D-NeRF | PSNR: 41.96 -> 42.00 | Hurt |
| set multi-head decoder depth to two (ablation) | Grid4D · D-NeRF | PSNR: 41.75 -> 42.00 | Hurt |
| replace rotation-and-translation position deformation with direct position offsets (ablation) | Grid4D · D-NeRF | PSNR: 41.80 -> 42.00 | Hurt |
| add smooth regularization to the MLP decoder (ablation) | Grid4D · D-NeRF | PSNR: 42.00 -> 41.92 | Hurt |
Surprising findings
- Adding smooth regularization directly to the MLP decoder did not help (41.92 vs 42.00 PSNR), which the authors attribute to MLPs' inherent smoothness.
- Deeper multi-head decoders (d=2) performed worse than shallow ones, attributed to training difficulty of deep MLPs.
Limitations
- No training speed improvement; smooth regularization requires encoding input twice.
- Artifacts remain on scenes with large and complex motions.
- Mean rendering quality on HyperNeRF rig split is not best in all per-scene results; the authors note Grid4D does not surpass others in several scenes.
Experimental setup
Datasets: D-NeRF HyperNeRF Neu3D
Benchmarks: D-NeRF HyperNeRF Neu3D
Metrics: PSNR SSIM LPIPS MS-SSIM FPS
Baselines: HexPlane TiNeuVox 4D-GS DeformGS SC-GS
Builds on: 3D Gaussian splatting multiresolution hash encoding 4D-GS DeformGS SC-GS
Closest papers by meaning
- Swift4D: Adaptive divide-and-conquer Gaussian Splatting for compact and efficient reconstruction of dynamic scene (ICLR 2025 · similarity 0.93)
- 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering (CVPR 2024 · similarity 0.92)
- Temporal Smoothness-Aware Rate-Distortion Optimized 4D Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- Fully Explicit Dynamic Gaussian Splatting (NeurIPS 2024 · similarity 0.90)
- Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting (ICLR 2024 · similarity 0.90)
- Motion Decoupled 3D Gaussian Splatting for Dynamic Object Representation (AAAI 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.