MeshSplatBench: A Unified Benchmark for Triangle-Based Neural Rendering

MeshSplatBench is a new benchmark for triangle-based neural rendering that tries to answer a question game teams actually care about: how well do these methods survive outside a custom research renderer? The benchmark evaluates methods across the full path from native optimization to game-engine deployment, instead of stopping at isolated lab metrics.

A key part of the setup is a standardized evaluation protocol that preserves each method’s native optimization semantics. The authors say they can reproduce published results within 0.8% PSNR deviation, which makes the comparisons much more credible for anyone trying to judge whether a technique is genuinely improving or just benefiting from a bespoke pipeline.

The Unity deployment side is split into three tiers: native CUDA renderers, method-specific dedicated engine shaders, and standard opaque mesh pipelines. That separation is useful because it isolates where fidelity is lost — whether from adapting the method to an engine, or from reducing the representation to something more conventional. For engine teams, that distinction matters when deciding whether a neural rendering approach is worth integrating at all.

The benchmark also includes a topology audit of reconstructed surfaces, and the takeaway is blunt: explicit connectivity and shared indexing do not guarantee production-ready assets. Non-manifold geometry, fragmented components, and boundary artifacts remain common, which means “rasterizable” is only the first hurdle. The source code is slated for release, so this could become a practical reference point for...