DashGaussian: Optimizing 3D Gaussian Splatting in 200 Seconds

TL;DR

The paper addresses the slow optimization of 3D Gaussian Splatting (3DGS) by introducing DashGaussian, a scheduling framework that reduces optimization complexity through frequency-guided rendering resolution and adaptive primitive growth. It reformulates 3DGS optimization as progressive fitting to higher-frequency components and couples this with a momentum-based primitive budgeting mechanism, enabling substantial speedups without compromising rendering quality. Across multiple backbones and datasets, it reports an average 45.7% acceleration and, in many cases, improved or preserved PSNR, SSIM, and LPIPS metrics while reducing the final primitive count. The approach demonstrates strong transferability as a plug-in missing only minor backbone modifications, and scales to large-scale reconstruction tasks, offering practical benefits for real-time or resource-constrained 3D scene synthesis and analysis.

Abstract

3D Gaussian Splatting (3DGS) renders pixels by rasterizing Gaussian primitives, where the rendering resolution and the primitive number, concluded as the optimization complexity, dominate the time cost in primitive optimization. In this paper, we propose DashGaussian, a scheduling scheme over the optimization complexity of 3DGS that strips redundant complexity to accelerate 3DGS optimization. Specifically, we formulate 3DGS optimization as progressively fitting 3DGS to higher levels of frequency components in the training views, and propose a dynamic rendering resolution scheme that largely reduces the optimization complexity based on this formulation. Besides, we argue that a specific rendering resolution should cooperate with a proper primitive number for a better balance between computing redundancy and fitting quality, where we schedule the growth of the primitives to synchronize with the rendering resolution. Extensive experiments show that our method accelerates the optimization of various 3DGS backbones by 45.7% on average while preserving the rendering quality.

Figures (5)

• Figure 1: We propose DashGaussian, a fast 3D Gaussian Splatting (3DGS) optimization method that can be easily plugged into existing 3DGS backbones. DashGaussian significantly boosts the training speed of various 3DGS backbones by 45.7% on average without trading off rendering quality. Equipping DashGaussian to prior-art 3DGS methods, we reduce the optimization time of a 3DGS model with millions of primitives to 200 seconds on a consumer-grade GPU. The figures above show the scene "stump" in the Mip-NeRF 360 dataset.
• Figure 2: The pipeline of DashGaussian. 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.
• Figure 3: Qualitative results of \ref{['tab:main-result']}. We show the rendering results of the "treehill" scene in the Mip-NeRF 360 dataset Barron_2022_CVPR, the "playroom" scene in the Deep Blending dataset DeepBlending2018, and the "truck"scene in the Tanks&Temples dataset Knapitsch2017, respectively.
• Figure 4: Variation in the optimization complexity. The left shows that DashGaussian scene-adaptively schedules the rendering resolution for the scenes "bicycle" and "stump" from Mip-NeRF 360 dataset. Please refer to implementation details (beginning of \ref{['sec:exp']}) for the definition of 'contiguous' and 'modulated'. The right shows the comparison on primitive growth between 3DGS kerbl3Dgaussians and 3DGS+DashGaussian on these two scenes.
• Figure 5: Profiling of the average optimization time on the Mip-NeRF 360 dataset. We can see DashGaussian effectively reduces the per-iteration elapsed time of all three major operations. It verifies that DashGaussian can effectively save the overall computational complexity of 3DGS optimization.