Original title: MeshSplatBench: A Unified Benchmark for Triangle-Based Neural Rendering
Kaixuan Zhang, Minxian Li, Mingwu Ren, Xiatian Zhu · September 2026
Triangle-based neural renderingAI methods that create 3D scenes using triangles compatible with standard graphics hardware. creates 3D models using triangles that work with standard graphics software. This paper introduces a benchmark called MeshSplatBench to test these models from creation to use in game engines. It finds that even though the triangles are standard, the models often have structural errors like broken or incomplete meshes. These errors make them unusable in real games, despite working in research setups.
Imagine you're building a custom bike using parts designed to fit any standard bike frame. But when you try to ride it on rough terrain, some parts break because they have hidden cracks that weren't obvious during assembly. In this paper, triangle-based neural rendering is like using standard bike parts (triangles) that should work with any graphics engine. The benchmark is like testing the bike on rough terrain (game engines), revealing that hidden mesh flaws (cracks) cause failures, even though the parts are standard.
Game developers might adopt neural rendering thinking it's ready for production, but this benchmark reveals that standard triangles aren't enough—mesh problems can crash engines. It saves teams from building on flawed methods by highlighting real-world compatibility issues.
The benchmark only evaluates triangle-based neural rendering methods, not other types of neural rendering. It focuses on testing within the Unity engine, which may not represent all game engines, and it identifies topology problems without proposing solutions to fix them.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
MeshSplatBench: A Unified Benchmark for Triangle-Based Neural Rendering — Kaixuan Zhang, Minxian Li, Mingwu Ren et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.