DashGaussian: Optimizing 3D Gaussian Splatting in 200 Seconds

Youyu Chen1#  Junjun Jiang1†  Kui Jiang1  Xiao Tang2  Zhihao Li2  Xianming Liu1  Yinyu Nie2 
1Harbin Institute of Technology     2Huawei Noah's Ark Lab    
chenyouyu@stu.hit.edu.cn   jiangjunjun@hit.edu.cn   tangxiao12@huawei.com  
† Corresponding author   # This work was done during he was an intern at Huawei.
CVPR 2025
arXiv Code (Coming Soon)

What is DashGaussian?

DashGaussian is a plug-and-play 3DGS training acceleration method which smartly allocates the computational complexity over the optimization process. DashGaussian reduces the training time-cost by 45.7% on average over various datasets and different backbones, while preserving and even improving the rendering quality.

Training Process Visualization

We demonstrate how DashGaussian accelerates the training process of different backbones on various scenes.


'Ours' denotes enhancing the backbone (Taming-3DGS) with DashGaussian.
Fullscreen the videos to view better.
'Ours' denotes enhancing the backbone (3DGS) with DashGaussian.
Fullscreen the videos to view better.
'Ours' denotes enhancing the backbone (Mip-Splatting) with DashGaussian.
Fullscreen the videos to view better.

Tour over the Optimized Scenes

Please click the videos to fullscreen for better view.

Method Overview

Architecture

DashGaussian determines the rendering resolution for each 3DGS optimization step with our resolution scheduling method. The insight of the resolution scheduling is to gradually fit 3DGS to higher level of frequency components in the training views throughout the entire optimization process. By directing the downsampling of training views with our scheduler, we significantly reduce the time cost for 3DGS optimization while preserving the rendering quality. We further manage the growth of Gaussian primitives, which cooperates with the scheduled rendering resolution. It prevents possible over-densification issues during the low-resolution optimization phase and further accelerates the optimization with suppressed primitive growth.

DashGaussian Improving SOTA

Quantitative comparisons with existing 3DGS fast optimization methods. With DashGaussian, we finish 3DGS optimization within 200 seconds while achieving significantly higher rendering quality compared with the other methods. ``Ours'' in the table reports the performance of equipping DashGaussian to Taming-3DGS. Time is reported in minutes.
Architecture

DashGaussian Enhancing Backbones

Quantitative results of accelerating various 3DGS backbones using DashGaussian. Plugging DashGaussian into different backbones, their optimization speed is largely improved by 45.7% on average. Most prominently, DashGaussian achieves acceleration with negligible compromise and even improvement in the rendering quality. The results advocate that a reasonable computational resource allocation across the optimization process can largely boost the optimization while preserving the rendering quality. ''Revising-3DGS*'' denotes our reproduction of the paper based on the Taming-3DGS codebase. Time is reported in minutes.
Architecture

BibTeX

@inproceedings{chen2025dashgaussian,
  title     = {DashGaussian: Optimizing 3D Gaussian Splatting in 200 Seconds},
  author    = {Chen, Youyu and Jiang, Junjun and Jiang, Kui and Tang, Xiao and Li, Zhihao and Liu, Xianming and Nie, Yinyu},
  booktitle = {CVPR},
  year      = {2025}
}