Holistic Large-Scale Scene Reconstruction
via Mixed Gaussian Splatting

Chuandong Liu    Huijiao Wang    Lei Yu*    Gui-Song Xia*   
Wuhan University   *Corresponding Author

NeurIPS 2025

  arXiv   Video   Code   BibTeX

Abstract

Recent advances in 3D Gaussian Splatting have shown remarkable potential for novel view synthesis. However, most existing large-scale scene reconstruction methods rely on the divide-and-conquer paradigm, which often leads to the loss of global scene information and requires complex parameter tuning due to scene par titioning and local optimization. To address these limitations, we propose MixGS, a novel holistic optimization framework for large-scale 3D scene reconstruction. MixGS models the entire scene holistically by integrating camera pose and Gaus sian attributes into a view-aware representation, which is decoded into fine-detailed Gaussians. Furthermore, a novel mixing operation combines decoded and original Gaussians to jointly preserve global coherence and local fidelity. Extensive experi ments on large-scale scenes demonstrate that MixGS achieves state-of-the-art ren dering quality and competitive speed, while significantly reducing computational requirements, enabling large-scale scene reconstruction training on a single 24GB VRAMGPU.

Watch MixGS reconstructed scenes

We visualize the reconstruction results of MixGS on the UrbanScene3D and Mill19 datasets. Use the controls to switch between scenes.

Training pipeline of MixGS

(1) Weintroduce MixGS, a novel holistic optimization method designed to overcome the limitations of existing approaches that rely on divide-and-conquer strategies for large-scale scene reconstruction.
(2) Wedevelop a mixed Gaussians rasterization pipeline that simultaneously captures global and local scene information through view-aware Representation modeling for implicit feature learning.
(3) Extensive experimentation demonstrate that MixGS achieves state-of-the-art results while maintain ing comparable rendering speed on established benchmarks for large-scale scene reconstruction.

MixGS novel view synthesis on large-scale datasets

CityGS MixGS GroundTruth

MixGS depth rendering on large-scale datasets

CityGS MixGS Original Image

Please consider citing our paper

@article{liu2025holistic,
          title={Holistic Large-Scale Scene Reconstruction via Mixed Gaussian Splatting},
          author={Liu, Chuandong and Wang, Huijiao and Yu, Lei and Xia, Gui-Song},
          journal={arXiv preprint arXiv:2505.23280},
          year={2025}
        }

Acknowledgement

This work is built upon CityGaussian, and 3D GS. The project page is based on ZeroGS. We sincerely thank all the authors for releasing their code.