Faithful denoising of sensitivity fields

Develop a denoising method for Monte Carlo sensitivity fields that supports faithful human inspection without changing the ranking of the parameters that dominate the sensitivity.

Background

The sensitivity buffers produced by the physically based path tracer inherit Monte Carlo path-tracing variance, particularly at low sample counts. Although denoising may improve visual quality, the paper notes that denoising a sensitivity field has a stricter requirement than denoising a beauty image: it must preserve the relative dominance of influential parameters.

The authors characterize this task as distinct from ordinary beauty-image denoising and explicitly identify it as unresolved.

References

Like all Monte Carlo derivatives they inherit path-tracing variance, so buffers computed at low sample counts may need denoising, and denoising a sensitivity field for faithful inspection, without reordering which parameters dominate, is itself an unstudied problem distinct from beauty-image denoising.

— Sensitivity as an Arbitrary Output Variable for Differentiable Rendering  (2610.10852 - Beresna et al., 7 Oct 2026) in Section 6, “Limitations and Future Work”