Color temperature is the temperature you would have to heat a lump of idealised carbon to before it glowed the color of the light you are looking at. Heat it to 2700 kelvin and it glows the warm amber of a soft white bulb. Heat it to 6500 and it glows something close to overcast daylight. That is the entire idea, and it is why the numbers on the side of a light bulb box are in kelvin rather than in anything to do with color.
The chart below is the part everyone actually came for. What makes it different from the dozens of identical strips on lighting retailer blogs is that I did not eyedrop it from another chart. I generated it from Planck's law, integrated each blackbody spectrum against the CIE 1931 two degree color matching functions, and converted the result to sRGB. The hex values are what the physics says, not what a designer thought looked about right.
Doing it that way turned up three things the catalogue charts get wrong, and I will get to all three after the chart.
The color temperature chart
Each swatch is the color of an ideal blackbody radiator at that temperature, normalised to equal brightness, shown on a screen calibrated to D65. The mired figure is the same value in the reciprocal unit photographers use, and the last column is the CIEDE2000 color difference from the row above it.
Candle flame. Cannot be shown accurately on any sRGB screen.
Sunrise, low pressure sodium, a dimmed filament bulb.
Vintage Edison filament lamps.
Dimmed incandescent, decorative bulbs.
Soft white, warm white. The most common bulb sold.
Warm white halogen, bathroom and kitchen fittings.
Neutral white. Offices, retail.
Cool white. Garages, workshops, task lighting.
Early morning and late afternoon sun.
Horizon daylight. The print industry's viewing standard.
Electronic flash, mid morning sun.
Bright midday sun.
Labelled daylight. Not the same as D65, as covered below.
Light overcast.
Heavy overcast, open shade.
Deep shade under a clear sky.
Clear blue northern sky.
Very clear high altitude sky.
Bluest sky measured. Past this the color barely moves.
The one thing to notice before anything else: the scale runs backwards from how people talk. The low numbers are the ones we call warm, and the high numbers are the ones we call cool. That is not a mistake in the labelling, it is the physics being honest and the language being metaphorical. A hotter object genuinely does glow bluer. We just associate orange with fire and blue with ice, so the words point the opposite way from the numbers.
First problem: a kelvin is not a unit of color
Look at the step column in the chart. Moving from 2000 K to 2200 K, a gap of 200 kelvin, changes the color by 2.7 CIEDE2000. Moving from 10000 K to 12000 K, a gap of 2000 kelvin, changes it by 3.0. Ten times the number of kelvin buys you the same amount of visible color.
I measured this properly across the whole locus by taking a fixed 100 kelvin step at different starting points and recording how far the chromaticity moved in CIE 1960 uv, which is the space this kind of question is normally asked in:
- At 2000 K, a 100 kelvin step moves the color 0.0078 in uv.
- At 5000 K, the same 100 kelvin step moves it 0.0014.
- At 10000 K, 0.00036.
- At 15000 K, 0.00015.
That is a spread of nearly 52 to 1 between the warm end and the cool end. The kelvin is a perfectly good unit of temperature and a terrible unit of color appearance, because the color changes fast at the bottom of the scale and grinds almost to a halt at the top.
Put in terms of what you can actually see, using 2.3 CIEDE2000 as a rough threshold for a difference an ordinary person would notice with the two colors side by side:
- At 2000 K you need about 180 kelvin to see a change.
- At 3000 K, about 240 kelvin.
- At 6500 K, about 310 kelvin.
- At 10000 K, about 1450 kelvin.
- At 15000 K, about 6810 kelvin.
A kelvin at the warm end of the shelf is worth roughly 38 times as much visible color as a kelvin at the top of the sky. This is why arguing about 8000 K versus 9000 K in a monitor menu is close to pointless while the difference between 2700 K and 3000 K is something you will notice every evening.
The unit that does work
Photographers and film lighting crews solved this decades ago and mostly stopped mentioning it, which is why it never made it into the consumer charts. Instead of kelvin they use the mired, short for micro reciprocal degree, which is a million divided by the temperature in kelvin. 5000 K is 200 mired, 2500 K is 400 mired.
I ran the same uniformity test on the mired scale. A fixed step of 10 mired, taken at ten points from 2000 K all the way up to 16667 K, moves the color by:
- 3.29 in uv at the warm end (times a thousand)
- 3.70 at the peak
- 3.46 on average
A spread of 1.13 to 1, against 52 to 1 for the kelvin scale. The mired is close enough to perceptually even that you can treat it as even and be right almost all the time. If you ever wondered why colour correction gels are labelled in mired shifts rather than in kelvin, that ratio is the answer. A gel that takes 3200 K to 5600 K has to be labelled by how much it moves the light, not by where it starts, and only the mired scale lets you write that as a single number.
My honest caveat: the near-uniformity of the mired holds in the chromaticity space it was designed for. If you instead model the light as a color you are looking at on a fixed D65 screen and score it in CIEDE2000, the warm end needs bigger mired jumps to cross the visible threshold, roughly 28 mired at 2700 K against 7 mired at 6500 K, because CIEDE2000 deliberately compresses differences between highly saturated colors. Both results are correct, they answer different questions. The uv figure answers "how much did the light change". The CIEDE2000 figure answers "how obvious is that change on a screen".
Second problem: 6500 K is not daylight white
This one surprised me. Bulbs sold as 6500 K daylight are, in the trade, understood to be roughly D65, the standard daylight illuminant that your monitor and the entire sRGB standard are defined against. The number even matches: D65, 6500 K.
They are not the same color, and it is not close. I put the 6500 K blackbody and D65 side by side under the same measurement:
- D65 sits at x 0.3127, y 0.3290.
- A 6500 K blackbody sits at x 0.3135, y 0.3236.
- The color difference between them is 4.63 CIEDE2000, twice the threshold at which people call two colors visibly different.
More to the point, no blackbody temperature matches D65. I swept every value between 5000 K and 8000 K looking for the closest, and the best any of them managed was 4.45 at 6338 K. The gap never closes because D65 is not on the Planckian locus at all. It sits above it, on the green side, because real daylight is sunlight plus scattered blue sky plus whatever the atmosphere absorbed, and that mixture does not have a blackbody spectrum. The number 6500 describes its correlated color temperature, meaning the closest point on the locus, which is a deliberately loose relationship.
The practical version: a 6500 K bulb is measurably pinker than the white your screen is calibrated to. You can see this in the chart above, where the 6500 K swatch comes out as #fff8fe with a faint magenta lean rather than pure #ffffff. That is not a rounding artefact, it is the whole reason the lighting industry quotes a Duv figure alongside the kelvin rating. Kelvin tells you where along the locus a light sits. Duv tells you how far off the locus it strayed, and a light can be badly green or badly pink while still being honestly labelled 4000 K.
Third problem: the warm end of every chart is a lie
Candlelight is usually given as 1850 K to 1900 K, and it appears on every color temperature chart as a deep orange swatch. When I converted the 1900 K blackbody to sRGB, the blue channel came out negative.
A negative channel means the color is outside the sRGB gamut. It is a real color, your eye can see it, but no combination of your monitor's three primaries can produce it. I found the exact boundary by walking up the locus one kelvin at a time: the lowest color temperature an sRGB screen can show honestly is 1901 K. Everything below that, every candle swatch and oil lamp swatch and match flame swatch on every chart including the top of mine, has been clipped to the edge of the gamut. It is the right hue and it is not saturated enough.
This is the same gamut boundary problem that shows up when you try to take a saturated screen color to a printing press, which I measured separately in the piece on RGB versus CMYK. Different direction, same wall.
Why warm light distorts colors more than cool light does
Here is a result I did not expect and have not seen stated anywhere. The Planckian locus is not symmetric around white. Measuring how far each point sits from neutral in chroma:
- 2700 K sits 73.3 units from neutral.
- 3000 K, 61.6.
- 4000 K, 33.7.
- 6500 K, 3.7. Effectively neutral.
- 10000 K, 22.7.
- 20000 K, 40.9, and it never gets much past that.
A 2700 K bulb is 3.2 times further from neutral than a 10000 K sky is, and the warm end keeps going while the cool end runs out of road. The locus converges as temperature rises, so no amount of extra heat gets you a strongly blue light, while dropping the temperature gets you to a strongly orange one quickly.
That asymmetry has a consequence people feel without naming. Warm indoor lighting is a much larger color cast than any daylight condition, which is why paint you chose in the shop looks wrong at home in the evening and almost never looks wrong at home at noon. Your visual system corrects for it through color constancy, but the correction is imperfect and the errors it leaves behind are what makes metamerism such an expensive problem for paint and textiles.
The hue figures are worth noting too. The warm end sits at a hue angle around 61 to 72 degrees, which is orange running to yellow. The cool end sits around 280 degrees, which is violet, not blue. And the locus never passes anywhere near green or cyan. There is no green on the color temperature scale at any temperature. Every green cast you have ever seen in a photograph came from fluorescent tubes or foliage bounce, both of which sit off the locus, and neither of which a kelvin number can describe.
What this means when you are choosing a bulb
The most common practical question is whether the shelf ratings are genuinely different or just marketing. I ran every adjacent pair:
- 2700 K against 3000 K: 3.14. Visible, and the smallest real step.
- 3000 K against 3500 K: 4.77.
- 3500 K against 4000 K: 4.44.
- 4000 K against 5000 K: 8.47.
- 5000 K against 6500 K: 13.38.
- 2700 K against 6500 K: 30.07. Not remotely the same light.
Every step on the shelf clears the visible threshold, so the labels are doing real work. But the steps are wildly uneven. The jump from soft white to warm white is a tenth the size of the jump from soft white to daylight. If you are mixing bulbs in one room, mixing 2700 and 3000 is the only pairing that will look like a mistake rather than a decision.
Manufacturing tolerance is worth knowing about as well. The ANSI binning standard allows a nominally 2700 K lamp to land roughly 145 kelvin either side of its rating, which works out at about 1.55 CIEDE2000. That is below the threshold, so two bulbs from the same bin should look the same. Two bulbs from opposite edges of adjacent bins will not, which is the real reason a replacement bulb sometimes looks wrong next to its neighbour even though the boxes say the same number.
The bit that applies to screens
Night Shift, Night Light, f.lux and every other blue light filter work by sliding your display white point down the same locus, typically from 6500 K to somewhere around 3400 K. That is a shift of 25.7 CIEDE2000, which is enormous. It is most of the distance from daylight to a soft white bulb, applied to everything on your screen at once.
Because it moves everything together, it barely affects your ability to compare two colors that are both on screen at the same moment. Relative judgements survive a white point shift surprisingly well. What it destroys is any judgement that has to cross a gap in time. If you look at a color now and try to reproduce it in an hour, and the white point moved underneath you in between, your stored impression is being compared against a different reference than the one it was formed under.
That matters for the color memory game, where the whole task is holding a color across a gap with no reference to check against. If you want a clean read on your own color memory, turn the warm filter off first. The same advice goes for the Hex variant, where you are typing a number that means a specific thing under a specific white point, and for any attempt at the Farnsworth Munsell hue test on a screen.
How the chart was made
For anyone who wants to check the numbers or rebuild them, the method is four steps and no approximations.
- Compute the blackbody spectral radiance at each wavelength from Planck's law, using the CIE second radiation constant of 1.4388 by ten to the minus two metre kelvin.
- Integrate that spectrum against the CIE 1931 two degree color matching functions at 1 nanometre resolution from 360 to 830 nanometres, giving XYZ tristimulus values.
- Normalise to equal luminance, convert to linear sRGB with the standard matrix, scale to the brightest channel and apply the sRGB transfer function. Colors with a negative channel before scaling are flagged as out of gamut rather than silently clipped.
- For every color difference quoted here, convert to CIELAB against a D65 white point and run CIEDE2000 with all three weighting factors at one.
The sanity check that the pipeline is right: standard illuminant A is defined by the CIE as a Planckian radiator at 2856 K with chromaticity x 0.44758, y 0.40745. My code returns 0.4475, 0.4074 for 2856 K without being told anything about illuminant A. That agreement to four decimal places is the reason I trust the rest of the table.
One modelling choice worth stating plainly. Treating a light source as a color you look at, rather than a light you are adapted to, is a simplification. In a real room your eye adapts and the light looks roughly white whatever its temperature, which is exactly why you stop noticing your own kitchen lighting within a minute of walking in. The chart above is the correct model for a camera with its white balance locked, for a screen calibrated to D65, and for two lights compared side by side. It is the wrong model for asking what a room feels like after you have been sitting in it for ten minutes. That question needs a full appearance model, and no chart of any kind answers it.
Everything on one line
- Low kelvin is orange, high kelvin is blue, and the words warm and cool run the opposite way to the numbers.
- The kelvin scale is 52 times less uniform at the cool end than the warm end. Use mired if you need a scale where equal steps mean equal change.
- 6500 K is not D65. Nothing on the locus is. The closest any blackbody gets is 4.45 CIEDE2000 away, at 6338 K.
- Nothing below 1901 K can be shown accurately in sRGB, so every candle swatch you have ever seen is undersaturated.
- The warm end is 3.2 times further from neutral than the cool end, so warm light is the bigger distortion by a wide margin.
- There is no green anywhere on the scale. Green casts always come from somewhere off the locus.
If you want to find out how much any of this you can actually see, the color memory game is the short version, and training your eye for color is the long one. Both get harder under a warm bulb, and now you know by how much.