Some additional notes on the topic:

The author describes color rendering index as a percentage; it is not. It's 100 - [a calculation of average color error on 8 color samples]. Reporting the color error directly would probably be better; 90% and 95% would both be pretty good grades on most tests, but a rendering error of 10 sounds a lot worse than a rendering error of 5.

Being based on only 8 samples, CRI Ra doesn't tell the whole story. There are other standards like TLCI that use more samples, and six supplemental samples some sources report. R9 (strong red) and R12 (strong blue) are particularly relevant to modern lighting, as LED sources often struggle with them despite high CRI (Ra) ratings.

CRI is a comparison to blackbody radiation (like an incandescent bulb) or simulated daylight at the same correlated color temperature as the sample light source. An incandescent bulb and midday sunlight at the equator both have CRI of 100, but a CCT of 2700K and a CCT of 5700K will render colors very differently. White balance in a camera's settings or raw development software can make one look like the other.

CCT is roughly blue-yellow balance, but light sources can deviate from blackbody radiation on a red-green axis. This is most often expressed as "delta u v" or "Duv", as it refers to distance in the CIELUV color space[0]. Duv numbers are small, with differences of less than 0.001 being noticeable in a side by side comparison and more than +/- 0.006 falling outside the ANSI standard for white light. Positive numbers are green, and negative numbers are red/pink.

[0] https://en.wikipedia.org/wiki/CIELUV

▲

wlesieutre 2 hours ago | parent | next [-]

Continuing on CRI’s limitations, since 2015 the new and improved metric is TM-30’s “fidelity index” (Rf) which is calculated based on 99 samples, among many other improvements.

The larger number of samples makes it harder for an LED manufacturer to “cheat” by optimizing for the specific reflectance spectrums used to score CRI, getting a higher score while not actually rendering colors of most real world materials that well.

https://www.energystar.gov/sites/default/files/asset/documen...

Unfortnately it’s still much less popular than CRI.

▲

Zak an hour ago | parent [-]

I like the goal of making it harder to overfit, but there are a couple significant weaknesses.

A big one is TM-30 Rf is still an average, and with a larger number of samples, the average minimizes the effect of strong deviations that only affect one or two samples. Actually looking at the color vector graphic reveals it, but some formats, like putting a bunch of different light sources in a table make that information harder to convey.

Probably worse when it comes to LED sources is that it doesn't include a primary red or primary blue. R9 (strong red) is a classic weak point for LEDs, with many "high-CRI" LEDs measuring in the 50s, and some other white LEDs having negative scores. R12 (strong blue) is a different kind of weak point; negative scores aren't common, but neither are values over 80.

So I'd like to a standard that uses more values, but it needs to include the red and blue primaries, and it needs to clearly report significant deviations - "Rworst" or some such.

	▲	wlesieutre 33 minutes ago \| parent [-]
		Yes, ultimately a limitation of trying to boil a huge space down to a single value, it can’t represent everything. But I think they had to do Rf anyway, it’s not perfect but it’s better than CRI by miles. The color vector graphics are a great feature for understanding the details better, but it would be hard to write one of those into a specification.

▲

cure_42 2 hours ago | parent | prev [-]

Fascinating, thank you!