Why a CRI 95 Light Still Renders Red Badly

This page contains affiliate links. If you buy through one we may earn a commission, at no cost to you.

No, and not because the correlation happens to be weak. It is because R9 is not one of the numbers that Ra is the average of. CRI, written Ra, is the arithmetic mean of eight fidelity scores, R1 through R8, taken on eight test colour samples of low to moderate saturation. The saturated samples are numbered R9 through R14 and they are excluded from the average by definition. So a light can publish Ra 96 with an R9 of 30, and a light can publish Ra 96 with an R9 of 95, and the published figure is identical in both cases. There is nothing in the number to decode, because the information was never put in.

What the eight samples are, and what they leave out

The method is laid down in CIE 13.3. You compute, for each of eight specified test colour samples, the colour difference between how that sample looks under the light being measured and how it looks under a reference illuminant of the same correlated colour temperature. For a source at or above 5000 K the reference is a CIE daylight illuminant, and below that it is a blackbody radiator. Each difference is turned into a score with the formula Ri equals 100 minus 4.6 times the colour difference, and Ra is the mean of the eight.

Two things fall out of that formula straight away. The first is that a score can be negative, because there is no floor at zero in 100 minus 4.6 times anything, which is why you occasionally see an R9 published as a negative number and it is not a misprint. The second is the sensitivity: with eight samples averaged, a single sample has one eighth of the weight, so moving one sample by eight points moves Ra by one. Divide by the 4.6 and one point of Ra is roughly 1.7 colour difference units on a single sample. That is a difference you can see.

The eight samples themselves are a pale set. Light greyish red, dark greyish yellow, strong yellow green, moderate yellowish green, light bluish green, light blue, light violet, light reddish purple. They are pastels, chosen in the 1960s to sample the hue circle evenly at modest saturation, and they are the whole of what the published number covers. The saturated set that follows them is where the interesting colours live: R9 strong red, R10 strong yellow, R11 strong green, R12 strong blue, R13 light skin, R14 leaf green. None of those six is in Ra.

Work the arithmetic the drawing shows. Seven samples at 97 and one at 79 give an average of 94.75, which a listing rounds and prints as CRI 95. Eight samples all sitting at exactly 95 give an average of 95 too, and get printed the same way. The two lights are not remotely alike and the number cannot tell them apart, because an average cannot be decomposed. Meanwhile the R9 bar in the drawing is outside the box: slide it from 100 to below zero and the printed figure does not change by a single point.

Why deep red is the sample an LED is worst at

This would be a pedantic complaint if the excluded sample were a colour that sources happen to render well. It is the opposite. Almost every white LED in consumer lighting is a blue pump design: an indium gallium nitride die emitting at roughly 450 to 460 nanometres, with a phosphor layer on top that absorbs some of that blue and re emits a broad yellow band peaking somewhere around 550 to 570 nanometres. Your eye adds the leftover blue to the broad yellow and reads white.

That construction has two signature weaknesses. There is a dip around 480 to 500 nanometres, between the blue peak and the start of the phosphor band, and there is a steep roll off past about 650 nanometres, where the phosphor has run out of emission. Deep saturated red is exactly the region that reflects strongly beyond 650 nanometres. So the one sample that the published average leaves out is the one sitting on the part of the spectrum where this entire family of light sources is weakest. It is not a coincidence that R9 was left out and reds look flat; the exclusion is what lets the flat reds go unreported.

That is also the reason the metric has been superseded twice over without the old number going away. The EBU defined TLCI, which runs the same comparison through a model television camera and display rather than the human eye, because a camera's three channels fail differently from your retina. The IES defined TM-30, which replaces eight samples with 99 colour evaluation samples and reports two numbers instead of one: Rf, a fidelity average over all 99, and Rg, a gamut index where 100 means the source reproduces the same saturation as the reference, above 100 means it oversaturates and below 100 means it washes colour out. And the Academy defined SSI, which compares the whole spectral power distribution against a reference rather than scoring samples at all. Film crews ask for R9 as a separate figure for the same reason.

What our own lights publish, and the one that publishes enough

Start with the cleanest case on the site, because the claim is in the product name. A Godox SL60IID LED video light, 70 W at a fixed 5600 K is sold with CRI 95+ in its title. Read further into the same listing and the colour fidelity bullet states a CRI and TLCI rating of 96 and 97. So the listing carries two different CRI numbers for the same light, one in the name and one in the specification, and they are not even the same kind of claim: 95+ is a floor that any unit must clear, and 96 is a point value. If you were comparing this against something advertised as CRI 96 you would be comparing a floor with a figure, and neither tells you where the deep red sample landed. What is genuinely useful here is the 97 next to the TLCI, because that is the camera referenced number and most listings never give it. The light also publishes 18,600 lux at one metre with the standard reflector, and our own record flags its fixed 5600 K as a drawback, which is the subject of the companion post below. Full record in our Godox SL60IID review.

Now the one listing on the site that publishes enough to answer the question. An Aputure amaran 100x S bi colour LED light states CRI 95+, TLCI 97+, TM-30 Rf 96 and TM-30 Rg 102, plus SSI figures of 89+ against tungsten and 87+ against D56 daylight, across a 2700 K to 6500 K range. That is four independent metrics where almost everything else on the market gives you one. The Rf 96 is the meaningful one: it is a fidelity average over 99 samples rather than eight, so the saturated colours are inside it and cannot hide outside a box. And Rg 102 says the gamut is very slightly wider than the reference, so saturated colour is being pushed out a touch rather than being flattened, which is the failure mode that a poor R9 produces. None of that is a measurement we made. It is what the listing publishes, and the point is that it publishes it at all. Our record is at the Aputure amaran 100x S review. One caution on reading our own record for this light: two of the drawbacks we list are annotated in the record itself as not specified but common for the class, which means they are inference rather than published fact, and they should be read that way.

Then the silence. An 18 inch NEEWER ring light kit publishes 55 W, 5600 K, 240 LEDs, dimming from 0 to 100 percent and a pair of filters, and does not publish a CRI at all. Not a low one. None. Its listing also cannot agree with itself on its own diameter, giving 18 inches as 48 cm in the kit contents line and as 45 cm in the feature line, when 18 inches is 45.7 cm. A Weilisi 10.2 inch desk ring light publishes ten brightness levels, three colour temperatures and an edge lit panel, and likewise no colour fidelity figure of any kind. And a NEEWER two pack bi colour 660 LED kit publishes CRI 96+ without saying at which point in its 3200 K to 5600 K range that figure was taken, which matters more than it sounds and is the subject of the companion post. Records for those are in the NEEWER 18 inch ring light review, the Weilisi desk ring light review and the NEEWER 660 kit review.

Two things we have said on this site that are wrong

We should correct ourselves before correcting anyone else, because both of our lighting guides assert exactly the premise this page takes apart.

Our ring lights guide says, in as many words, that "a light with a high CRI rating ensures that colors appear natural and accurate in your photos or videos." It does not ensure that. It reports an average over eight pastel samples and is silent about every saturated colour, and not one of the five products listed underneath that sentence publishes a CRI figure at all, so the advice cannot be applied even to the picks it sits above.

Our LED video lights guide is worse, because it attaches the claim to the one subject where the exclusion bites hardest: "look for lights with a high Color Rendering Index (CRI) of 95 or above to ensure skin tones look natural." The red in skin is haemoglobin, which is why R13 light skin exists as a separate special index and why R9 deep red is the sample cinematographers ask about before any other. A high Ra with a poor R9 renders skin as grey and lifeless, and faces are the first place anyone notices it. Asking for Ra 95 to protect skin tone is asking the one number that excludes skin for a guarantee about skin.

The honest limit

Not one listing on this site publishes R9. Our Aputure publishes TM-30 Rf and Rg, which is the next best thing and a real reason to prefer it if colour fidelity is the deciding factor, and everything else gives you an Ra, a plus sign and nothing more. So the practical position is not that you should hunt for better numbers. It is that the numbers available to you cannot answer the question, and you should stop treating Ra as though it had.

Which leaves two moves, and the free one is better than the purchase. First, the test: put something deep red next to something pale, under the light, and look at it. A red mug, a red shirt, a tomato, your own lips. A source with a collapsed R9 renders them brownish, muddy and oddly dark while the pastels in the same frame look fine, and that split is the signature. Shoot it, white balance on a grey or white card so the overall cast is removed, and the red either survives or it does not. Five seconds of looking beats any figure a box will print.

Second, and this is the part that makes most of the spec hunting pointless: daylight scores 100 on every one of the fourteen samples by construction, because for any source at or above 5000 K daylight IS the reference the test compares against. A north facing window at midday is not a good light with a high CRI. It is the thing the CRI of a 5600 K light is a measure of its distance from. If you can shoot a face next to a window, you have already beaten every panel on this page for colour fidelity, for free, and the only reasons to buy a light are the ones daylight genuinely cannot give you: an output you can set and repeat, a colour that does not change when a cloud passes, and a time of day of your choosing. Buy for control, not for colour. And if you do buy for colour, ask the maker for R9 or Rf and treat a refusal as the answer. More on why the dial on a bi colour panel costs you something in what a bi colour LED costs you for the dial.