Research map / 3D vision / Gaussian splatting
Gradient-Direction-Aware Density Control for 3D Gaussian Splatting
Zheng Zhou, Yu-Jie Xiong, Jia-Chen Zhang, Chun-Ming Xia, Xihe Qiu, Hongjian Zhan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| nonlinear (power) weighting instead of linear weighting (ablation) | GDAGS · Mip-NeRF360 (9 scenes) | PSNR / SSIM / LPIPS: SSIM 0.814 -> 0.839; PSNR 27.55 -> 28.02; LPIPS 0.248 -> 0.145 | Improved |
| apply GCR weighting during splitting only (ablation) | GDAGS · Mip-NeRF360 (9 scenes) | memory / SSIM / LPIPS: memory 734MB -> 441MB; SSIM 0.815 -> 0.819 | Mixed |
| apply GCR weighting during cloning only (ablation) | GDAGS · Mip-NeRF360 (9 scenes) | PSNR / memory: PSNR 27.21 -> 27.46; memory 734MB -> 615MB | Mixed |
| apply GCR weighting in both splitting and cloning (ablation) | GDAGS · Mip-NeRF360 (9 scenes) | SSIM / PSNR / LPIPS: SSIM 0.839, PSNR 28.02, LPIPS 0.145 vs best single-sided 0.819, 27.52, 0.217 | Improved |
| increase p from 10 to 30 (ablation) | GDAGS · Mip-NeRF360, Tanks&Temples, Deep Blending | memory and rendering quality | Mixed |
| increase beta from 15 to 35 (ablation) | GDAGS · three datasets | performance and storage | No clear effect |
| use exponential instead of power function in weighting (ablation) | GDAGS · three datasets | training time: 1140s -> 1200s (+5.2%) on Mip-NeRF360 | Hurt |
| exponential instead of power function (ablation) | GDAGS · Mip-NeRF360 (9 scenes) | PSNR: 28.02 -> 27.96 | Hurt |
| add GDAGS to MCMC-3DGS | MCMC-3DGS · Mip-NeRF360 (9 scenes) | LPIPS: 0.190 -> 0.150 | Improved |
| add GDAGS to MCMC-3DGS | MCMC-3DGS · Mip-NeRF360 (9 scenes) | PSNR: 29.89 -> 29.80 | Hurt |
| add GDAGS to Compact-3DGS | Compact-3DGS · Mip-NeRF360 (9 scenes) | LPIPS: 0.247 -> 0.227 | Improved |
| add opacity-weighted Gaussian dropout 5% (ablation) | GDAGS · Deep Blending | PSNR: 29.70 -> 30.03 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- mini-splatting — Mini-Splatting: Representing Scenes with a Constrained Number of Gaussians (ECCV 2024) · better on 2 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Steepest Descent Density Control for Compact 3D Gaussian Splatting (CVPR 2025 · similarity 0.94)
- Efficient Scene Modeling via Structure-Aware and Region-Prioritized 3D Gaussians (TPAMI 2025 · similarity 0.94)
- Pixel-GS Density Control with Pixel-aware Gradient for 3D Gaussian Splatting (ECCV 2024 · similarity 0.93)
- ReorgGS: Equivalent Distribution Reorganization for 3D Gaussian Splatting (NeurIPS 2026 · similarity 0.92)
- Frequency-Aware Density Control via Reparameterization for High-Quality Rendering of 3D Gaussian Splatting (AAAI 2025 · similarity 0.92)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (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.
