Upstream reshaping of inter-channel covariance geometry produces standard JPEGs that encode at 1.31 BPP mean through Facebook’s pipeline, down from 4.80 BPP without the use of new codec, neural network or quality reduction. Every existing decoder reads the output unchanged.
The repo contains all 24 images, 72 JSON measurements, and a Python verification script that reproduces every published number in under 2 minutes.
<a href="https://github.com/PearsonZero/kodak-pcd0992-spdr-verification-suite" target="_blank" rel="noopener nofollow">https://github.com/PearsonZero/kodak-pcd0992-spdr-verification-suite</a>
Sixth and final paper in a series measuring the decorrelation gap between BT.601’s fixed transform and per-image KLT optima across the Kodak PCD0992 suite.

Deleted user

A claimed 73% JPEG bits-per-pixel reduction across the 24-image Kodak PCD0992 benchmark suite is presented, achieved by reshaping inter-channel RGB covariance geometry upstream before standard JPEG encoding. The technique reportedly brings mean BPP from 4.80 down to 1.31 using Facebook's pipeline, without a new codec, neural network, or quality loss. All standard decoders read the output unchanged. A GitHub repo provides all 24 images, 72 JSON measurements, and a Python verification script. This is the sixth paper in a series measuring the decorrelation gap between BT.601's fixed transform and per-image KLT optima on the Kodak dataset.

73% JPEG BPP reduction across 24 Kodak images via upstream RGB covariance reshaping — for you to verify