SIGGRAPH 2026

Faster 3D Gaussian Splatting Convergence via Structure-Aware Densification

Max Planck Institute for Informatics, Cambridge University, and Saarbrücken Research Center for Visual Computing, Interaction, and Artificial Intelligence (VIA).

Paper PDF Supplement arXiv GitHub

Overview of multiscale structure analysis and structure-aware Gaussian densification. — We analyze multiscale image structure in the input views and use it to guide anisotropic densification during 3D Gaussian training.

Abstract

3D Gaussian Splatting has emerged as a powerful scene representation for real-time novel-view synthesis. However, its standard adaptive density control relies on screen-space positional gradients, which do not distinguish between geometric misplacement and frequency aliasing, often leading to either over-blurred high-frequency textures or inefficient over-densification. We present a structure-aware densification framework. The key insight is that the decision to subdivide a Gaussian should be driven by an explicit comparison between its projected screen-space extent and the local structure of the texture it seeks to represent. We introduce a multi-scale frequency analysis combining structure tensors with Laplacian scale space analysis to estimate dominant local frequency, define a per-Gaussian, per-axis frequency violation metric, and perform anisotropic splitting with a multiview consistency criterion. By densifying early and more directly, the method skips lengthy iterative densification phases and achieves faster convergence with superior reconstruction quality, particularly in high-frequency regions.

Real-Time Optimization

Training visualizations show fast convergence on real-world scenes while preserving perceptual detail.

0:00 / 0:00

Quantitative Results

Compared with the SOTA baseline FastGS-Big, our method is 2.8x faster on average and improves LPIPS by 0.017 absolute, an 8.2% relative reduction.

Quantitative comparison table on Mip-NeRF360, Deep Blending, and Tanks and Temples. — Table 1. Quantitative comparisons on Mip-NeRF360, Deep Blending, and Tanks & Temples. We report optimization time, PSNR, SSIM, LPIPS, the number of Gaussians, and training iterations.

Convergence Curves

The method converges faster than strong acceleration baselines while maintaining image quality.

0:00 / 0:00

Qualitative Comparison

Compare our reconstruction against baseline methods across representative scenes.

Select Scene

Select Baseline

Ours

Time: 51.9s | LPIPS: 0.300

FastGS-big

Time: 193s | LPIPS: 0.341

BibTeX

@inproceedings{lyu2026faster,
  author = {Lyu, Linjie and Tewari, Ayush and Chen, Jianchun and Leimk{\"u}hler, Thomas and Theobalt, Christian},
  title = {Faster 3D Gaussian Splatting Convergence via Structure-Aware Densification},
  booktitle = {Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers},
  series = {SIGGRAPH Conference Papers '26},
  year = {2026},
  month = jul,
  location = {Los Angeles, CA, USA},
  publisher = {Association for Computing Machinery},
  address = {New York, NY, USA},
  isbn = {979-8-4007-2554-8},
  doi = {10.1145/3799902.3811212},
  url = {https://doi.org/10.1145/3799902.3811212},
  numpages = {10}
}