Laplacian Frequency Hierarchies for Efficient 3D Gaussian Splatting Training

A plug-and-play frequency-staged training scheme for accelerating 3DGS backbones such as FastGS and Taming-3DGS.

Yixiong Yang1,*, Sisheng Zhang1,*, Qingsong Yan2, Shaohuai Shi1, and Qiang Wang1,†

1Harbin Institute of Technology, Shenzhen, China    2XGRIDS, China
*Equal contribution    †Corresponding author

Code Watch Demo
Laplacian-GS teaser
Figure 1: Laplacian-GS trains frequency-staged Gaussian fields and reconstructs the final image with Laplacian-style image-domain composition.

Abstract

A key bottleneck in 3D Gaussian Splatting training is the continual growth of Gaussian primitives, which increases optimization cost and slows convergence, especially at high resolutions. We propose Laplacian Frequency Hierarchies, a simple yet efficient 3DGS scheme that combines Laplacian image decomposition with coarse-to-fine, frequency-staged training. After fitting lower-frequency structure, we archive the corresponding Gaussian field so that subsequent fields can optimize higher-frequency residuals without carrying the full primitive burden, and we compose the rendered components in the image domain via a Laplacian-style reconstruction at inference time. This design reduces the number of active Gaussians during training, lowering optimization overhead and accelerating training. The proposed scheme is plug-and-play and orthogonal to prior 3DGS accelerations: it can be directly combined with strong backbones such as Taming-3DGS and FastGS to improve training speed with competitive reconstruction quality.

Method

Laplacian-GS decomposes supervision into frequency bands and optimizes a separate Gaussian field for each stage.

Overview of the Laplacian-GS method
Figure 2: Overview of our method. We decompose each training image into a resolution pyramid and the corresponding Laplacian residual pyramid. During training, a sequence of Gaussian fields $\mathcal{G}^{L-1}, \mathcal{G}^{L-2}, \ldots, \mathcal{G}^{0}$ is optimized in a coarse-to-fine manner, with each field learning a specific frequency component. At inference time, the fields are rendered separately and their outputs are composed via Laplacian reconstruction to produce the full-resolution result.

Demo

Results

The released scripts reproduce the Mip-NeRF 360 1K setting with FastGS and Taming-3DGS backbones.

Laplacian-GS 1K qualitative results
Figure 3: Qualitative results of the 1K setting.
Laplacian-GS 2K qualitative results
Figure 4: Qualitative results of the 2K setting.
Laplacian-GS 4K qualitative results
Figure 5: Qualitative results of the 4K setting.
Backbone Script Setting
FastGS run_mipnerf360_1k_fastgs.sh Mip-NeRF 360, 1K/1.6K resize convention
Taming-3DGS run_mipnerf360_1k_taming.sh Mip-NeRF 360, 1K/1.6K resize convention

Acknowledgement

We would like to thank the authors of the open-source projects Taming-3DGS, DashGaussian, FastGS, and 3DGS for making their implementations publicly available.