S2Gaussian: Sparse-View Super-Resolution 3D Gaussian Splatting

CVPR 2025

Yecong Wan1, Mingwen Shao1, Yuanshuo Cheng1, Wangmeng Zuo2,
1China University of Petroleum (East China) 2Harbin Institute of Technology
Paper Code (Soon)

Abstract

In this paper, we aim ambitiously for a realistic yet challenging problem, namely, how to reconstruct high-quality 3D scenes from sparse low-resolution views that simultaneously suffer from deficient perspectives and clarity. Whereas existing methods only deal with either sparse views or low-resolution observations, they fail to handle such hybrid and complicated scenarios. To this end, we propose a novel Sparse-view Super-resolution 3D Gaussian Splatting framework, dubbed S2Gaussian, that can reconstruct structure-accurate and detail-faithful 3D scenes with only sparse and low-resolution views. The S2Gaussian operates in a two-stage fashion. In the first stage, we initially optimize a low-resolution Gaussian representation with depth regularization and densify it to initialize the high-resolution Gaussians through a tailored Gaussian Shuffle Split operation. In the second stage, we refine the high-resolution Gaussians with the super-resolved images generated from both original sparse views and pseudo-views rendered by the low-resolution Gaussians. In which a customized blur-free inconsistency modeling scheme and a 3D robust optimization strategy are elaborately designed to mitigate multi-view inconsistency and eliminate erroneous updates caused by imperfect supervision. Extensive experiments demonstrate superior results and in particular establishing new state-of-the-art performances with more consistent geometry and finer details.

Method Overview

The S2Gaussian initially optimizes an LR GS and densifies it to initialize the HR GS through a tailored Gaussian Shuffle Split operation. Then, the original sparse views along with the pseudo views rendered by the LR GS are super-resolved together to refine the high-resolution texture with 3D robust optimization. In which an inconsistency modeling module (IM) and a blur proposal module are incorporated to mitigate inconsistency and blurriness, aiming to create 3D scenes with high-fidelity texture details.

Main Results

Visual Results

Visual Results

Visual Results

BibTeX

@article{wan2025s2gaussian,
  author    = {Yecong Wan and Yuanshuo Cheng and Mingwen Shao and Wangmeng Zuo},
  title     = {S2Gaussian: Sparse-View Super-Resolution 3D Gaussian Splatting},
  journal   = {CVPR},
  year      = {2025},
}