JLD: Perceptual Distance Through A Jacobian Lens

Official fitted Jacobian lens for JLD and JLD-fast, by Shreshth Saini, Balu Adsumilli, and Alan C. Bovik.

Paper · Code · Project and visual examples · Hugging Face paper page · Interactive demo

JLD measures the perceptual distance between an aligned reference image and a distorted image. Lower scores mean more similar images. It projects early DINOv2 patch features onto directions selected by the encoder's output sensitivity. The lens is fitted from unlabeled images without human quality ratings.

Use with uv

uv add "jacobian-lens-distance @ https://github.com/shreshthsaini/jld/releases/download/v1.1.0/jacobian_lens_distance-1.1.0-py3-none-any.whl"

This installs the versioned release wheel and records it in uv.lock. The distribution is named jacobian-lens-distance; the Python import remains jld.

from jld import JLD

metric = JLD.pretrained("full")  # use "fast" for JLD-fast
distance = metric("reference.png", "distorted.png")
print(distance.item())

response = metric.map("reference.png", "distorted.png")

The package already includes this lens. The encoder weights are downloaded separately through timm on first use. Python 3.10 or newer is required; CPU and CUDA are supported.

To download the lens from this repository explicitly:

from huggingface_hub import hf_hub_download
from jld import JLD

lens_path = hf_hub_download(
    repo_id="shreshthsaini/JLD",
    filename="jld_dinov2_s14_block1_k64.npz",
)
metric = JLD.pretrained("fast", lens_path=lens_path)

Artifact and provenance

This repository contains the fitted projection, its estimated sensitivity matrix, eigenvalues, and fitting metadata. The encoder remains frozen; its weights are provided by timm/vit_small_patch14_dinov2.lvd142m, under that repository's Apache 2.0 license.

Setting	Released lens
Encoder	DINOv2-S/14, feature width 384
Captured features	Patch tokens after block 1
Projection rank	64
Fitting images	100 DIV2K validation images
Fitting crops	Four random 224 × 224 crops per image
Gaussian probes	Eight per crop
Seed	0

SHA-256 of jld_dinov2_s14_block1_k64.npz:

0ae11bc65fb590e6d7b23e3840258746910efd4da94e0e147014c28afbc9c26c

This is the same lens shipped in the official Python package. Fitting code and evaluation scripts are in the source repository.

Evaluation and scope

The paper reports image-quality correlations on TID2013, CSIQ, LIVE, a held-out KADID-10k reference split, and PIPAL validation. The repository contains commands to reproduce those evaluations. This Hub release does not add a new benchmark result.

Use native-resolution RGB inputs, with values in [0, 1]. Images are center-cropped to multiples of 14 pixels without resizing, and each pair must have matching spatial dimensions. Changes to resizing, cropping, or color preprocessing can change scores and comparisons.

JLD requires a reference image. It is intended for aligned image comparisons, such as restoration and compression evaluation. The patch response map shows the lens term only. Full JLD adds a final CLS cosine term, so the full score need not satisfy the triangle inequality. The lens term is a pseudometric.

The video's temporal encoder and fitted lens in the paper are separate from this image lens. The image lens alone does not reproduce the paper's temporal video results. See the project page for the evaluated video protocols.

Citation

If you use JLD, its fitted lens, or its integrations in research, please cite:

@misc{saini2026jld,
  title={JLD: Perceptual Distance Through A Jacobian Lens},
  author={Shreshth Saini and Balu Adsumilli and Alan C. Bovik},
  year={2026},
  eprint={2610.05967},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.05967}
}

JLD code and this fitted lens are released under the MIT license. See LICENSE.