Establish whether the input-envelope guard generalizes beyond the tested shift families

Establish whether the three-statistic input-envelope guard generalizes to distribution shifts beyond the tested noise, blur, and darkening families, including compression artifacts, color casts, and mixed corruptions.

Background

The guard detects out-of-envelope inputs using ranges of log-Laplacian energy, luminance, and contrast-normalized high-frequency energy estimated from clean LOL-v1 calibration images. In the registered stress grid, it reduces harm for the tested corruption families, but the guard was designed after observing those same failure modes.

No held-out shift family is evaluated. The paper therefore limits the claim to the tested families and explicitly rejects interpreting the results as evidence of a general out-of-distribution detector; the residual darkening case also demonstrates that an input-marginal detector can fail when the shifted distribution overlaps the calibration distribution.

References

It is not evidence of a general out-of-distribution detector, and the dark~s1 residual is the standing reason to expect it is not one.

— When Should a Low-Light Pipeline Generate? Auditable Risk Control for Selective Generative Enhancement  (2610.21335 - Xie et al., 15 Oct 2026) in Section “Distribution-Shift Stress Grid and the Input-Envelope Guard,” paragraph “Scope of the guard”