1: \begin{abstract}

2: 

3: 3D representation is essential to the significant advance of 3D generation with 2D diffusion priors. As a flexible representation, NeRF has been first adopted for 3D representation. 

4: With density-based volumetric rendering, it however suffers both

5: intensive computational overhead and inaccurate mesh extraction.

6: Using a signed distance field and Marching Tetrahedra, DMTet allows for precise mesh extraction and real-time rendering but is limited in handling large topological changes in meshes, leading to optimization challenges.

7: Alternatively, 3D Gaussian Splatting (3DGS) is favored in both training and rendering efficiency while falling short in mesh extraction.

8: In this work, we introduce a novel 3D representation, Tetrahedron Splatting (\textbf{\model{}}), that supports easy convergence during optimization, precise mesh extraction, and real-time rendering {\em simultaneously}.

9: This is achieved by integrating surface-based volumetric rendering within a structured tetrahedral grid while preserving the desired ability of precise mesh extraction,

10: and a tile-based differentiable tetrahedron rasterizer.

11: Furthermore, we incorporate eikonal and normal consistency regularization terms for the signed distance field to improve generation quality and stability.

12: Critically, our representation can be trained without mesh extraction,

13: making the optimization process easier to converge.

14: Our \model{} can be readily integrated in existing 3D generation pipelines, along with polygonal mesh for texture optimization.

15: Extensive experiments show that our \model{} strikes a superior tradeoff among convergence speed, render efficiency, and mesh quality as compared to previous alternatives under varying 3D generation settings.

16: \end{abstract}

17: