2D-SuGaR: Surface-Aware Gaussian Splatting

2D-SuGaR teaser results — 2D-SuGaR achieves state-of-the-art mesh reconstruction on the DTU benchmark while maintaining high-quality novel view synthesis.

Abstract

3D Gaussian Splatting (3DGS) has emerged as a powerful technique for generating photorealistic renderings of a scene in real-time. However, the volumetric nature of 3DGS limits its ability to accurately capture surface geometry. To address this, 2D Gaussian Splatting (2DGS) was proposed to enable view-consistent and geometrically accurate surface reconstruction from multi-view images. However, 2DGS can be sensitive to the initialization of the Gaussian primitives. Reliance on Structure-from-Motion (SfM) initializations, which can produce poor estimates on challenging image sets, may lead to subpar results. In this work, we enhance 2DGS by incorporating monocular depth and normal priors to improve both geometric accuracy and robustness. We propose a depth-guided initialization strategy for Gaussians and introduce a clustering-based technique for pruning degenerate Gaussians. We evaluate our method on the DTU dataset, where it achieves state-of-the-art results in mesh reconstruction while preserving high-quality novel view synthesis.

Quantitative Results

DTU Benchmark: Chamfer Distance

Chamfer distance (lower is better) across 15 DTU scenes. Bold = best per column.

Method	24	37	40	55	63	65	69	83	97	105	106	110	114	118	122	Mean ↓	Time ↓
2DGS	0.48	0.91	0.39	0.39	1.01	0.83	0.81	1.36	1.27	0.76	0.70	1.40	0.40	0.76	0.52	0.80	10.9m
MILo	0.43	0.74	0.34	0.37	0.80	0.74	0.70	1.21	1.22	0.66	0.62	0.80	0.37	0.76	0.48	0.68	35m
Ours w/o Refinement	0.50	0.67	0.33	0.44	1.24	0.70	0.63	1.24	1.13	0.63	0.49	1.29	0.43	0.56	0.44	0.72	10.5m
Ours (2D-SuGaR)	0.48	0.65	0.29	0.39	1.24	0.69	0.61	1.17	1.10	0.56	0.45	1.19	0.37	0.50	0.40	0.67	18.5m

Visual Comparison

Ours vs. 2DGS vs. MILo

DTU Scan 37

Mesh

Geometry

DTU Scan 69

Mesh

Geometry

Ours vs. 2DGS

DTU Scan 24

Mesh

Geometry

DTU Scan 106

Mesh

Geometry

DTU Scan 122

Mesh

Geometry

Method

How It Works

2D-SuGaR builds on 2D Gaussian Splatting and introduces three key contributions to overcome sensitivity to SfM initialization:

Depth-Guided Initialization

Monocular depth estimates from Metric3D replace sparse SfM point clouds, providing a dense and robust initialization for Gaussian primitives even on textureless or specular surfaces.

Normal Prior Loss

Per-image normal maps from Metric3D are used as a geometric supervision signal during training, constraining the surface orientation of 2D Gaussians to align with true geometry.

Cluster-Based Pruning

Degenerate Gaussians are identified via clustering and pruned during training, preventing floaters and improving mesh extraction quality via TSDF fusion.

SuGaR Mesh Refinement

Meshes extracted from 2DGS are further refined using SuGaR, yielding high-fidelity textured meshes with UV mapping accelerated by Nvdiffrast.

Results

Novel View Synthesis

Our method simultaneously renders high-quality color, depth, and surface normal maps.

DTU Scan 55

Color

Depth

Normal

DTU Scan 37

Color

Depth

Normal

Custom Datasets

Generalization to In-the-Wild Scenes

Frankfurt Fridge Magnet

A challenging small object on a plain surface, where poor SfM initialization leads to failure in 2DGS. Our depth-prior initialization recovers geometry cleanly.

2DGS (baseline)

Ours

Citation

BibTeX

If you find our work useful, please consider citing our Eurographics paper:

@inproceedings{10.2312:egs.20261022,
booktitle = {Eurographics 2026 - Short Papers},
editor = {},
title = {{2D-SuGaR: Surface-Aware Gaussian Splatting for Geometrically Accurate Mesh Reconstruction}},
author = {Gupta, C. R. Prajwal and Sheth, Divyam and Ha, Jinjoo and Ostrek, Mirela and Thies, Justus},
year = {2026},
publisher = {The Eurographics Association},
ISSN = {2309-5059},
ISBN = {978-3-03868-299-8},
DOI = {10.2312/egs.20261022}
}

Acknowledgements

This project is built upon 3DGS, 2DGS, and SuGaR. TSDF fusion for mesh extraction is based on Open3D. Depth and normal priors are estimated using Metric3D. The rendering script for MipNeRF360 is adopted from Multinerf. Evaluation scripts for DTU are taken from DTUeval-python. We thank all the authors for their valuable open-source contributions.