JLD: Perceptual Distance Through A Jacobian Lens
JLD measures perceptual image distance from a frozen vision encoder’s Jacobian, fitted once on 100 unlabeled images.
The paper says this “Jacobian lens” keeps early patch locality while weighting directions that most affect later encoder representations. It reports state-of-the-art results across four perceptual databases, ahead of LPIPS, DISTS, PieAPP, and DreamSim. On TID2013, the lens term barely changes when resolution is doubled, dropping from 0.850 to 0.845 correlation. The authors also describe JLD-fast as 4x faster than LPIPS-VGG, and extend the method to video with a 0.786 correlation on Waterloo IVC 4K. HF Daily Papers' note
The paper says this “Jacobian lens” keeps early patch locality while weighting directions that most affect later encoder representations. It reports state-of-the-art results across four perceptual databases, ahead of LPIPS, DISTS, PieAPP, and DreamSim. On TID2013, the lens term barely changes when resolution is doubled, dropping from 0.850 to 0.845 correlation. The authors also describe JLD-fast as 4x faster than LPIPS-VGG, and extend the method to video with a 0.786 correlation on Waterloo IVC 4K. HF Daily Papers' note
score 4