Standard low-light enhancement benchmarks contain hidden label shifts

The critical finding isn't just that labels differ between train and test, but that this divergence is systematic and directional (saturation reduction, brightness shift), suggesting the test sets may have been processed through a different pipeline entirely rather than random variation.

This connects directly to a pattern we've covered repeatedly: benchmarks and evaluation infrastructure are becoming a primary source of unreliability in published claims. The ANT-A paper from late October exposed how annotation benchmarks treat labels as singular ground truth when real-world labeling involves legitimate disagreement. The latent world models evaluation framework from October 27 showed how leading implementations achieve task performance through shortcuts rather than genuine capability. And the label-free monitoring work from October 26 revealed that standard metrics can fail to detect actual degradation. What ties these together is a shared insight: the infrastructure we use to measure progress is often misaligned with what we're actually trying to measure, and this misalignment compounds across the field.

If the eight major low-light enhancement methods are re-evaluated on a corrected version of LOL-v1 and LOLv2-Real with consistent pipelines across train and test, watch whether the performance rankings change materially or whether relative ordering holds. If rankings invert significantly, it confirms the benchmark flaw was masking real capability differences; if they hold, it suggests the methods learned robustness to the pipeline shift, which is a different (and more interesting) claim than current papers make.