Measuring Light Quality
Single-phosphor budget LED: blue spike, cyan gap, red deficit. Ra 82 sounds respectable, but the vector graphic shows what it hides: reds (h1, h15–h16) collapsing inward, and R9 barely above zero.
Dashed circle = reference source at matched CCT. Each arrow is one hue bin's average shift: outward = oversaturated, inward = desaturated, sideways = hue rotation. Exaggeration scales the drawing only; Rf and Rg are always computed at ×1.
Both spectra are normalized to equal power inside the 375–675 nm SSI window, binned to 10 nm, and their weighted, smoothed relative difference is folded into one number: SSI = 100 − 32·‖v‖. The shaded area between the curves is what SSI punishes. Beyond 675 nm the index is deliberately blind; cameras and film barely respond there.
One light, six verdicts
CRI asks eight pastel chips. TM-30 asks 99 real objects. TLCI asks a television camera. SSI ignores objects entirely and interrogates the spectrum itself. Pick a source on the left and watch the metrics disagree, then read down for why each one was invented.
CRI, R9 and TLCI are typical published values for this source class. Rf and Rg are computed live from the hue-bin vectors at left; SSI is computed live from the spectra below using the SMPTE ST 2122 procedure.
Why CRI stopped being enough+
CRI Ra (CIE 13.3:1995) is 1974 mathematics: eight pastel Munsell samples, color differences taken in the long-obsolete 1964 U*V*W* space with a von Kries adaptation transform, averaged into one number. Saturated red is sample 9, and R9 is optional, not part of Ra.
The averaging is the trap. A narrow-band LED spectrum can be tuned to render the eight pastels acceptably while failing everything between them, so two Ra 90 sources can render skin, wood, and produce visibly differently. CRI was designed to compare fluorescent tubes against incandescent, broad spectra against broad spectra. It was never built for the spiky emission of solid-state lighting, and by the 2010s it was actively misleading buyers.
Inside TM-30: 99 samples, modern math+
ANSI/IES TM-30 (2015, revised 2018 and 2020; its fidelity index harmonized internationally as CIE 224:2017) replaces the 8 pastels with 99 Color Evaluation Samples, reflectance curves measured from real objects (skin, textiles, paints, plants) selected so their spectral features are evenly distributed across wavelength. That uniformity is the anti-gaming clause: there is no gap between samples for a peaky spectrum to hide in.
Color shifts are computed in CAM02-UCS, a modern uniform color space, against a reference at the test source's own CCT: Planckian below 4000 K, CIE daylight above 5000 K, blended between. Rf is a transform of the mean color difference across all 99 samples; 100 means indistinguishable from the reference. Typical Ra 80 lamps land in the mid-70s Rf, but rankings reshuffle for narrow-band spectra, which is the point.
Rg and the no-free-lunch trade+
The 99 samples are averaged into 16 hue bins. Connect the 16 bin-average points and you get the polygon in the vector graphic; Rg is that polygon's area divided by the reference polygon's area, ×100. Above 100 the source oversaturates on average; below 100 it dulls. Rg says nothing about which hues moved; that is the vector graphic's job.
Fidelity and gamut are chained: Rf 100 forces Rg 100, and every point of added gamut is bought with lost fidelity: a source at Rg 120 cannot exceed roughly Rf 80, because the same chroma shifts that inflate the polygon are counted as color error. "Vivid" is not free; it is a controlled distortion.
Reading the vector graphic like a spec sheet+
Check bin h1 first; it holds reds and most skin reproduction. Real specifications call it out by name: lighting specs for critical color environments demand skin-bin fidelity Rf,h1 ≥ 90 and a local chroma shift Rcs,h1 held within about −9% to +9%. An inward-pointing h1 arrow is dull, gray skin; a hard outward one is the flushed, sunburned look.
Outward bulges in h2–h3 (orange) are the grocery-store trick: oversaturated produce sells. Inward h15–h16 is dead wood tones and lifeless reds, the classic budget-LED signature. The polygon is the whole story in one shape: Rf is essentially the average arrow length, Rg the enclosed area, both single-number projections of what you are looking at.
Design intents: P, V, F+
TM-30-20's Annex E turns the metrics into specification language: three design intents (Preference, Vividness, Fidelity), each with graded priority levels (P1–P3, V1–V3, F1–F3) built from thresholds on Rf, Rg, and the skin bin. A P1 space (hospitality, residential) wants high Rf with a slight, controlled chroma lift; a V-intent retail install may spec Rg 102–108 while protecting h1; a museum or grading suite specs F: maximum Rf, Rg pinned near 100.
This is the practical upgrade over "CRI ≥ 80": instead of one blunt average, a spec now states what colors are supposed to do under the light, and the vector graphic shows whether a luminaire complies at a glance.
TLCI: judging light through a camera+
Your eye adapts; a camera does not. High-CRI LED fixtures that look fine on set were producing footage that needed expensive grading, so in 2012 the EBU published TLCI (Tech 3355 / Recommendation R 137). Its move: replace the human standard observer with a standard camera, a modelled three-chip HDTV chain plus Rec. 709 display, deliberately echoing how the CIE once standardized a human.
Pipeline: measure the SPD with a spectroradiometer → derive CCT → pick the reference illuminant → simulate a 24-patch ColorChecker through camera and display under both sources → average the color errors into a 0–100 score. Read it in bands: ≥85 needs no correction (>90 is "television-ready"), 75–85 minimal grading, 50–75 correctable at real cost, below 50 sub-broadcast. Its known limit: it models a three-chip broadcast camera, and single-chip cinema sensors respond differently, which is exactly why the Academy went one step further.
SSI: skip the samples, match the spectrum+
The Academy's Spectral Similarity Index (2016, revised 2020; standardized as SMPTE ST 2122) discards observers, cameras, and samples alike. It compares the SPD itself against a reference: studio tungsten (SSI[P3200]), a daylight illuminant, or any Planckian radiator over 375–675 nm, binned to 10 nm, both normalized to equal power; then a weighted relative-difference vector (the spectrum's ends are de-weighted, where cameras and film respond least), smoothed and folded into SSI = 100 − 32·‖v‖.
The logic is airtight: if two spectra match, they render identically on any camera, film stock, or observer, no model needed. The trade: SSI below 100 warns that rendering errors are possible, not which ones. And it exposes what sample-based metrics hide: a white LED scoring CRI 92 can post SSI[P3200] ≈ 76, because its blue spike and phosphor troughs sit exactly where the eight pastels never look. The notation always names the reference: SSI[P3200] = 100 means "is tungsten," not "is good."
Which metric, when: the calibrator's read+
Broadcast and studio luminaires: TLCI first, TM-30 for how the room reads to the people in it. Cinema sets mixing fixture vendors: SSI against the gaffer's chosen reference; it predicts whether two "5600 K" panels can intercut without a grade. Architectural and retail work: TM-30 design intents in the spec, vector graphic in the submittal review.
Displays are the deliberate exception: a display is judged by ΔE against signal targets, not by rendering indices (see the ΔE-ITP module). But these spectra are exactly why measurement hardware matters: a colorimeter's three filters can be fooled by the same spiky spectra that fool CRI, which is why per-display spectroradiometer profiling exists. Every number on this page began life as a spectroradiometer measurement. If it isn't measured, it isn't calibrated.
IF IT ISN'T MEASURED, IT ISN'T CALIBRATED. · Color Volume Explorer · ΔE2000 vs ΔE-ITP Explorer · A History of Color