ResGS: Residual Densification of 3D Gaussian for Efficient Detail Recovery

Accepted to ICCV 2025
Yanzhe Lyu      Kai Cheng      Xin Kang      Xuejin Chen
University of Science and Technology of China
Paper Code

Our ResGS proposes residual split to address the limitation of split and clone in 3D-GS. It can capture fine details and retrieve sufficient geometry effectively while reducing redundancy.

Your Page Title

Motivation

The threshold-based selection between split and clone during 3D-GS densification faces a trade-off between achieving sufficient structural coverage and preserving fine details. We address this challenge by introducing a new densification approach: residual split.

Your Page Title

Method

Overview of our ResGS. (a) The core of our pipeline, residual split, involves adding a downscaled replicate and then reducing the opacity of the original Gaussian. (b) We assign initial Gaussians a temporary attribute \( l_i=0 \) for densification selection, which is discarded after training. Next, the pipeline is split into \( L \) (\( L=3 \)) stages, with each stage trained on images downscaled using an image pyramid. Each single stage is further divided evenly into \( K\) (\( K=3 \)) substages, for selecting Gaussians to densify. (c) The points selected for densification are determined by the substage \( k \), \( l_i \) and viewspace gradients of Gaussians.

Results

The intermediate results of our ResGS during training. The video shows our method can construct missing structures and recover fine details at the same time.

Qualitive comparison of our ResGS with other 3D-GS varients. It is evident that our model captures finer details, as shown in the second, third, and fourth rows, constructs missing geometry more effectively (last row), and achieves more accurate geometry reconstruction (first row).

BibTeX

@inproceedings{lyu2025resgs,
  title={ResGS: Residual Densification of 3D Gaussian for Efficient Detail Recovery},
  author={Lyu, Yanzhe and Cheng, Kai and Kang, Xin and Chen, Xuejin},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={28093--28102},
  year={2025}
}