A Stable Transport-Mechanism Descriptor for Per-Pixel Rendering Difficulty

Per-pixel rendering difficulty is usually estimated with sample variance, but that signal gets shaky exactly where the hard pixels live: heavy-tailed transport. A new descriptor tackles the problem by labeling each contribution event using its end-vertex BSDF lobe, whether a delta-specular event is present, and whether the path is single- or multi-bounce. That produces seven mutually exclusive labels, with six named mechanisms covering all observed energy in the tested scenes.

The big takeaway for rendering engineers is stability. Across seven scenes, the dominant label agreed 87% to 99.6% between 64 and 4096 samples per pixel, while quantile-binned variance agreement could fall to 21%. The descriptor also held up when the estimator’s MIS half was restored, which matters if you’re trying to use pilot renders to drive adaptive sampling without chasing noise.

The method is not just a better scalar. It exposes structure that variance cannot represent, including a geometry-driven sign flip in the correlation between delta-mediated and glossy transport. In practice, using the label to correct a noisy pilot variance beat pilot-variance allocation at equal budget on every heavy-tailed test-matrix scene, while falling back to the incumbent behavior when those buckets were absent.

The robustness claims are unusually broad for this kind of work: the gains survived a random-partition placebo, persisted over a robust median-of-means pilot baseline, and were checked with pre-registered third-party sentinel tests. For teams building adaptive sampling, path guiding, or denoising...