Physics-guided networks for underwater image restoration are designed to predict transmission and backscatter, then invert the image-formation model. The authors of this paper audit one representative design, PhysGuide-Net, to test whether that physics actually does any work. They run four controlled experiments, including five-seed ablations, a data-efficiency study with a capacity-matched control, and a validation of predicted maps against measured range.

With all 800 training images, the physics module gave no zero-shot gain: mean PSNR differences were at most 0.16 dB, and all 95% confidence intervals included zero, despite requiring 3.4x the FLOPs. With only 80 images, the module helped more (+0.65 dB on UIEB, +1.48 dB on AUIED3K), but the capacity-matched control did not reproduce this benefit, and the differences were not individually significant after correction.

The predicted transmission does not behave like a physical quantity. Its correlation with range is no stronger than for the range map of a different image, and a Beer-Lambert fit explains only 0.4% of its variance. The module's effect also changes sign with degradation severity between datasets. The authors conclude the benefit comes from a physics-shaped inductive bias, not from physics itself.

They also show that single-run ablations can reverse sign and document implementation defects that invalidate loss ablations. The findings suggest that gains attributed to physics in such models should be verified with controlled, multi-seed audits before being taken at face value.