AI

JLD: Perceptual Distance Through A Jacobian Lens

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 07/10/2026, 06:18:24 UTC.

Briefings

A new metric called Jacobian Lens Distance (JLD) measures perceptual differences using a frozen vision encoder's Jacobian tensor, achieving state-of-the-art performance on four standard databases while remaining stable under resolution changes. The method combines early patch features with later encoder representations to create a fixed geometric interpretation for image comparison.

Key points

• JLD achieves state-of-the-art performance across four standard perceptual databases, consistently outperforming LPIPS, DISTS, PieAPP, and DreamSim.
• JLD-fast achieves 4 times faster performance than LPIPS-VGG while maintaining a mean correlation of 0.911.

Context

Image compression, restoration, and generation require accurate perceptual distance metrics. Existing methods like DISTS and LPIPS rely on human judgments, which ties them to fixed data and resolution. JLD derives its geometry from a frozen vision encoder, offering a fixed metric tensor that is more robust to resolution changes.

Limitations

JLD was trained on 100 unlabeled images and requires a frozen vision encoder. Performance on TID2013 drops from 0.815 to 0.717 when resolution doubles, though the lens-term correlation remains nearly unchanged at 0.845. Video performance on Waterloo IVC 4K shows a correlation of 0.786 compared to 0.611 for VMAF.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 25/10/2026, 13:16:39 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeiafzh255z4tnqed6dafkboeufr36gz7y3iinicd6kn47vyfsawic4