MeshSplatBench: A Unified Benchmark for Triangle- and Mesh-Based Neural Rendering

arXiv cs.GR details MeshSplatBench, a benchmark built to evaluate triangle- and mesh-based neural rendering under both native rendering and graphics-engine deployment. The goal is practical: these methods are often pitched as a bridge between neural scene representations and standard engines like Unity and Blender, but they have been compared under inconsistent conditions.

The benchmark introduces two deployment paths. Standard deployment uses a conventional opaque mesh pipeline with vertex colors and hardware Z-buffering, while dedicated deployment adds method-specific engine support to preserve effects such as alpha blending and compositing. The authors also add a structural audit for mesh splatting to check whether exported surfaces are actually usable downstream, not just renderable in isolation.

The headline finding is that engine deployment can noticeably degrade quality across methods, even when the output is technically rasterizable. Mesh splatting appears more robust under standard deployment, but preserving fidelity with dedicated integration comes at a steep cost: roughly 6-30x slowdown. The structural audit also suggests that current explicit connectivity and shared vertex indexing are still not enough to guarantee manifold or globally connected meshes.

For developers, the takeaway is straightforward: “graphics-ready” needs a stricter definition than “can be rasterized.” If these methods are going to matter in production pipelines, they need to survive the same constraints artists and engine programmers deal with every day, not just benchmark scenes. The...