# 3D Gaussian Flats: Hybrid 2D/3D Photometric Scene Reconstruction

**Authors:** Maria Taktasheva, Lily Goli, Alessandro Fiorini,  Zhen,  Li, Daniel Rebain, Andrea Tagliasacchi  
**arXiv:** [2509.16423](https://arxiv.org/abs/2509.16423)  
**Date:** 2025-09-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2509.16423v1/

## Key points

- Recent advances in radiance fields and novel view synthesis have enabled the creation of realistic digital twins from photographs.
- Challenges with flat, texture-less surfaces lead to uneven and semi-transparent reconstructions due to ill-conditioned photometric reconstruction objectives.
- A novel hybrid 2D/3D representation has been proposed to address these limitations.
- The approach optimizes constrained planar (2D) Gaussians for modeling flat surfaces and freeform (3D) Gaussians for the rest of the scene simultaneously.
- Dynamically detecting and refining planar regions enhances visual fidelity and geometric accuracy.
- Significant improvements in reconstructed surface geometry were demonstrated on common indoor scene datasets like ScanNet++ and ScanNetv2.
- Superior performance was shown in rendered image quality metrics such as PSNR, SSIM, and LPIPS compared to state-of-the-art fully 3D representations and 2D surface reconstruction approaches.
- Strong performance in depth estimation metrics like RMSE, MAE, AbsRel, and depth accuracy percentage was achieved.
- The method excelled at mesh extraction for planar surfaces within indoor scenes without overfitting to specific camera models.
- An ablation study validated the robustness and efficacy of the proposed method's design choices.
- Implementation details are provided for transparency and reproducibility.

## Summary

Recent advances in radiance fields and novel view synthesis have enabled the creation of realistic digital twins from photographs. However, current methods face challenges with flat, texture-less surfaces. This often results in uneven and semi-transparent reconstructions due to an ill-conditioned photometric reconstruction objective. While surface reconstruction methods can address this issue, they often compromise visual quality. To tackle these limitations, a novel hybrid 2D/3D representation has been proposed. This approach optimizes constrained planar (2D) Gaussians for modeling flat surfaces and freeform (3D) Gaussians for the rest of the scene simultaneously. By dynamically detecting and refining planar regions, this end-to-end method enhances both visual fidelity and geometric accuracy. The effectiveness of this proposed method has been validated through the task of novel view synthesis on common indoor scene datasets. Evaluation on benchmarks such as ScanNet++ and ScanNetv2 demonstrates significant improvements in reconstructed surface geometry while maintaining high visual quality. In the evaluation process, comparisons were made with state-of-the-art fully 3D representations and 2D surface reconstruction approaches. The method showcased superior performance in terms of rendered image quality metrics such as PSNR, SSIM, and LPIPS. Depth estimation was also a key focus, with metrics including RMSE, MAE, AbsRel, and depth accuracy percentage indicating strong performance in reconstructing surface geometry accurately. Furthermore, the proposed approach excelled at mesh extraction for planar surfaces within indoor scenes. By leveraging a combination of 2D and 3D Gaussian representations,<DateTime>, it achieved state-of-the-art depth estimation on ScanNet++ and ScanNetv2 datasets without overfitting to specific camera models. Through an ablation study that delved into different aspects of the method's design choices, its robustness and efficacy were further validated. Implementation details have been provided in supplementary material for transparency and reproducibility. Overall, this hybrid 2D/3D representation offers a promising solution to the challenges posed by flat surfaces in scene reconstruction. Its ability to optimize planar and freeform Gaussians jointly results in high-quality reconstructions of indoor scenes with improved visual fidelity and geometric accuracy compared to existing methods.