LoopExpose learns arbitrary-length exposure correction without paired labels through a nested correction-fusion loop. Multi-exposure fusion iteratively refines pseudo-labels, while luminance-ranking supervision preserves the relative brightness order of the input sequence.

ZeroEC removes dependence on curated multi-exposure data by synthesizing continuous exposure degradations online and enforcing consistent correction across them. An inverse-exposure constraint exploits complementary observable regions to improve severely degraded areas while remaining applicable to different correction backbones.

This work reformulates exposure correction around spatially non-uniform degradation. It predicts region-adaptive modulation signals for explicit 3D LUT transformation, supplements color fidelity in HSL space, and uses uncertainty-aware optimization to balance heterogeneous local restoration demands.

This work challenges ground-truth fidelity as the objective of exposure correction through an input-conditioned human preference study. It introduces ECIQA-Pref, a benchmark of 5,200 images and over 60,000 pairwise judgments, and ECIQA, a degraded-reference quality metric that combines global appearance with native-resolution local structure to evaluate perceptual improvement.