ΔE2000 vs. ΔE-ITP: Scoring Color Error in SDR & HDR

A color-difference metric is a promise: compute one number, and it tells you how visible an error is, regardless of where in color space it sits. The MacAdam ellipses showed that raw distance in xy breaks that promise. Every ΔE formula since is an attempt to build a space where straight-line distance IS visibility, so that "ΔE < 2" can be written into a delivery spec and mean the same thing for a skin tone, a sky, and a shadow.

ΔE2000 measures distance in CIELAB, then repairs Lab's known warps with weighting functions (SL, SC, SH) and a rotation term that fixes the notorious blue region. It is excellent at what it was built and validated on: reflective samples and SDR displays, roughly 0.2–100 cd/m², viewed against a defined white. Its Achilles' heel is baked into Lab's first step: every value is computed relative to a reference white (Y/Yₙ). Ambiguity about that white is ambiguity about every number it produces.

HDR has no single reference white; diffuse white sits near 100–203 cd/m² while highlights run to thousands. Anchor Lab at 100 nits and everything brighter compresses into L* > 100 where the formula was never validated: the sweep plot shows it going numb exactly where HDR lives. In the shadows the cube-root lightness overstates differences PQ-graded material renders invisibly. And wide-gamut primaries push chroma into regions where even the 2000 corrections drift. The metric isn't wrong; it's out of its jurisdiction.

ICtCp rebuilds the space for HDR from the ground up: absolute light → cone responses (LMS) → the PQ curve applied to each cone signal → an opponent transform. I is intensity; Ct is the tritan (blue–yellow) axis; Cp the protan (red–green) axis. Because PQ is inside the space, it is perceptually even from 0.0001 to 10,000 cd/m² with no reference white needed. ΔE-ITP (Rec. ITU-R BT.2124) is then nearly Euclidean: scale T = 0.5·Ct, and ΔE = 720·√(ΔI² + ΔT² + ΔP²), normalized so 1.0 ≈ one JND. That flat line in the sweep plot is the whole argument.

A probe ultimately reports CIE XYZ, and Z is effectively the blue-weighted channel. It is the smallest signal of the three on most content, and the biggest lever: errors in Z swing blue chromaticity hard, precisely where MacAdam's ellipses are tiniest and vision least forgiving. Wide-gamut displays make it worse: narrow-band OLED/QD/laser primaries land on the steep slopes of a colorimeter's filter curves, so small filter-match errors become large XYZ errors. That is why serious WCG work uses a spectroradiometer as reference and per-display calibration matrices (four-color/FCMM) for the fast colorimeter, and why a probe that was fine on sRGB can lie on a QD-OLED. (JzAzBz, the Jz lightness correlate, attacks the same HDR problem from the XYZ side and appears in research metrics.)

Saturated primaries carry startlingly little luminance: in BT.2020, blue is 5.9% of white; at a 1,000-nit peak a full blue patch is ~59 cd/m², and near black it's milli-nits. Verifying the color volume therefore means measuring accurately at both extremes: long integration times and low-light-capable probes in the shadows, no flare or ND ambiguity at HDR peaks, and primaries measured at multiple luminance levels, not just 100%. Skip that and the tops and ridges of the volume model are decoration, not data.

SDR / Rec.709 work: ΔE2000 remains the lingua franca; grade-reference monitors typically hold < 1.5–2 across the gray ramp and ColorChecker set. HDR / PQ / WCG verification: use ΔE-ITP, where 1 ≈ 1 JND; ≤ 3 is a common studio tolerance for reference displays, and dynamic content is judged against it shot-wide. Quote the metric with the number: "ΔE 2" means nothing until you say which ΔE, against which white, at which luminance. A proper report states all three. Book a calibration →