SuGaR is a rendering technique gaining popularity for its realistic output and faster training compared to Neural Radiance Fields (NeRFs). A key contribution involves a regularization term ensuring alignment of 3D Gaussians with the scene surface, facilitating precise mesh extraction. Leveraging this alignment, SuGaR introduces a fast and scalable Poisson reconstruction method, outperforming traditional Marching Cubes algorithms. An optional refinement strategy further binds Gaussians to the mesh, enabling seamless editing and manipulation using standard 3D software.

This rendering method significantly reduces mesh extraction time, minutes versus hours with state-of-the-art methods on Neural SDFs, while improving rendering quality based on PSNR, SSIM, and LPIPS metrics. The regularization term optimizes Gaussians' distribution, capturing scene geometry effectively. Efficient point sampling on visible parts of density function level sets enhances scalability, leading to highly detailed surface mesh reconstruction.

The optional refinement strategy binds new Gaussians to mesh triangles, facilitating joint optimization for high-quality rendering through Gaussian Splatting. This novel approach outperforms radiance field models relying on underlying meshes, offering enhanced performance in rendering quality. Additionally, it allows the utilization of traditional mesh-editing tools for flexible and versatile Computer Graphics applications.