ΔE Metrics from CIE76 to CIEDE2000 and ΔEITP

Delta E is only useful when it is connected to a real decision. The same two samples can pass on a web UI, fail in a textile lab, look unstable under retail lighting, or need an HDR-aware metric because luminance dominates the match. This lab treats color difference as a quality-control system: formula choice, tolerance, observer uncertainty, viewing condition, metamerism, and instrument repeatability all move the final call.

Visual comparison works until it doesn't. Two humans looking at two paint chips might disagree about whether they match. The same human might disagree with themselves across morning and evening light. Different cultures use different color names for what is measurably the same hue. The industry needs a number - one anyone, anywhere, can compute from instrument readings and use to settle a question that subjective judgement leaves open.

ΔE (read "delta E," from the German Empfindung, "sensation") is that number. It is the distance between two colors in some perceptually meaningful space. ΔE = 0 means identical; ΔE = 1 is the threshold of just-noticeable difference; ΔE = 5 is a clear shift; ΔE = 20 is "different colors entirely." Every paint contract, print spec, display certification, and color-difference benchmark hangs on a chosen ΔE formula and a chosen threshold.

Different formulas exist because perception is not Euclidean. CIELAB tries to build a uniform space; ΔE in CIELAB is approximately right but breaks down in saturated colors, dark tones, and the blue corner. Each newer formula corrects specific weaknesses. Choosing the wrong one for the wrong job produces decisions that look numerically rigorous but visually wrong.

Published with CIELAB itself in 1976, ΔE76 (also called ΔE_ab) is simply the Euclidean distance between two Lab points:

The simplicity is the strength. ΔE76 is fast, symmetric (color A vs color B = B vs A), well-behaved as a distance, and easy to implement. For decades it was the only formula most workflows had.

Its weakness is the same as CIELAB's: equal numerical distances do not always correspond to equal visual differences. The space is more compressed in blues than in greens; saturated colors have less perceptual room than neutrals. A ΔE76 of 3 in pale gray looks much more obvious than a ΔE76 of 3 in deep saturated red. For sorting nearly-matching swatches this matters; for ballpark comparison it's often fine.

In 1984 the UK's Colour Measurement Committee published a refined formula designed for fabric and textile matching. CMC(l:c) introduced two weighting factors - l for lightness and c for chroma - that let users tune the formula for the perceptual realities of their materials. CMC(2:1) is the textile default, meaning lightness differences count half as much as chroma differences (because textile observers naturally tolerate more lightness variation than chroma).

CMC was the first formula to use lightness, chroma, and hue weighting based on the color's position in Lab space. A red and a pink can share the same ΔL*, Δa*, Δb* values but mean very different things visually; CMC scales the terms based on where in space the comparison sits.

CMC remains the de-facto standard in the textile industry. Most fabric acceptance criteria still use CMC(2:1) with a pass threshold around 1.0. Outside textiles, it has been mostly superseded by CIE94 and CIEDE2000 - which were designed by similar reasoning but with broader applicability.

In 1994 the CIE published its first official refinement of ΔE76. The structure of ΔE94 mirrors CMC: split the Cartesian Δa*, Δb* into chroma difference (ΔC) and hue difference (ΔH), then weight each term by lightness, chroma, and hue-dependent factors. Three application-specific parameter sets were defined: graphic arts (kL=1, K1=0.045, K2=0.015) and textiles (kL=2, K1=0.048, K2=0.014).

ΔE94 dramatically improved the saturated-color accuracy of the formula. Its limitation: not symmetric. ΔE94(A, B) ≠ ΔE94(B, A) in general, because the weighting depends on the "first" color's chroma. The textbook convention is to use the reference (target) sample as color 1; but for arbitrary pairwise comparison, this becomes awkward.

CIEDE2000 (also written ΔE₀₀) is the CIE's current recommendation and the formula most modern color-difference work uses. Published in 2001 (the year after its name), it incorporates four refinements over ΔE94:

1. Symmetrized weighting via averaged chroma C̄ so ΔE₀₀(A,B) = ΔE₀₀(B,A)
2. Hue rotation term R_T that handles the blue-purple region where Lab is most distorted
3. Tunable kL, kC, kH that default to 1 but allow adjustment for special viewing conditions
4. Recompensed a* axis via a factor G that corrects the well-known stretching near low-chroma saturated reds

The full formula is intricate but mechanical. Implementations exist in every serious color-management toolkit. Most production code uses a tested library rather than rewriting the math.

For most color-critical work in 2026 - print contracts, paint matching, display certification - ΔE₀₀ is the default. When in doubt, compute and report it. CIE76 lives on for quick estimates; CMC for textile contracts that predate ΔE2000.

CIEDE2000 works inside CIELAB, which assumes SDR viewing at moderate luminance. HDR breaks that assumption. A color difference that looks small at 100 cd/m² may look enormous at 4000 cd/m². The same ΔE₀₀ can mean very different things depending on absolute brightness.

ΔEITP (ITU-R BT.2124) was published in 2019 specifically for HDR and wide-gamut color difference. It operates in ICTCP space - a perceptually quantized opponent space that scales with absolute luminance via PQ. The result is a ΔE that remains meaningful across the entire HDR dynamic range.

Björn Ottosson's 2020 Oklab color space (and its polar form OKLCh) is designed to be perceptually uniform from the start. Where CIELAB needed CIEDE2000's elaborate corrections to repair its non-uniformities, Oklab aims for a simpler Euclidean ΔE in its native space to suffice.

The trade-off: Oklab's scale is different from CIELAB's. A ΔE_OK of 0.01 corresponds roughly to a ΔE₀₀ of 1 - the numerical thresholds are completely different. Oklab has gained traction in CSS Color 4 (color-mix(in oklab, ...)), modern design systems, and palette generation tools. For interactive UI work where speed matters and CIEDE2000's cost is noticeable, Oklab is increasingly the choice.

Just-noticeable difference (JND) is the smallest color difference a typical observer can reliably detect under controlled viewing conditions. For CIEDE2000, JND is generally taken as ΔE₀₀ ≈ 1.0. Differences below that are statistically invisible to most people most of the time.

Industries have converged on ΔE thresholds appropriate for their materials and visual demands. These numbers appear in contracts, QC procedures, and ISO specifications.

Different formulas produce different numbers for the same pair. Knowing which disagrees and by how much is the difference between a defensible measurement and an argument.

Every formula assumes things that are not always true. Knowing the assumptions keeps you honest about when the number means what it claims.

Six questions on ΔE formulas and thresholds. Wrong answers come with brief explanations.