T²-Splat: Adaptive Topology Mesh Splatting with Texture Residuals
Authors: Zhihao Tang, Youjia Zhang, Mingbo Zhao, Wei Yang
Subjects: Computer Vision and Pattern Recognition (cs.CV); Graphics (cs.GR)
Venue: Accepted at NeurIPS 2026
arXiv: 2611.06974 [cs.CV]
Submitted: 4 November 2026
Project Page: https://t2-splat.github.io/
Overview
Triangle-based splatting bridges novel view synthesis and conventional graphics pipelines, but typically demands excessive tessellation to preserve appearance details. T²-Splat addresses this limitation by decoupling topology adaptation from appearance refinement—an approach that keeps mesh complexity compact while maintaining high-fidelity rendering.
Method
T²-Splat introduces a three-stage pipeline:
- Safe Restricted Delaunay Triangulation (Safe RDT) — initializes mesh connectivity with a robust triangulation scheme.
- Dynamic Topology Evolution (DTE) — adapts triangle allocation through edge collapse and vertex split operations, allowing the mesh to allocate geometric detail where it matters most.
- Residual per-face Mip-textures — refine appearance on the fixed mesh using scale-adaptive filtering, capturing fine texture detail without additional geometry.
- +0.71 dB PSNR improvement over MeshSplatting on Mip-NeRF 360.
- Additional compression experiments show flexible control of mesh complexity, supporting face budgets as low as 10% of the original face count.
By separating topology from appearance, the method avoids the redundant tessellation that conventional triangle-based splatting requires to preserve visual quality.
Results
Experiments on Mip-NeRF 360, Tanks & Temples, and DTU demonstrate competitive results with compact meshes:
Significance
The decoupled design offers a practical path toward lightweight, graphics-pipeline-compatible scene representations—relevant as the field moves toward efficient, deployable 3D reconstruction and rendering in 2026. The ability to aggressively reduce mesh complexity while retaining appearance quality makes T²-Splat well suited to real-time and resource-constrained applications.
Citation
arXiv:2611.06974 [cs.CV]
https://doi.org/10.48550/arXiv.2611.06974
via ArXiv CV