White balance is the correction that makes a neutral object look neutral under a coloured light. The light's colour is described by its correlated colour temperature, the temperature in kelvin of an ideal hot "black body" that glows the same colour: a candle is about 1,850 K and orange, overcast daylight about 6,500 K and bluish. Your eyes adapt to the light automatically and imperfectly; a camera has to estimate the light from the image itself, which is why auto white balance fails in mixed or single-colour light.
- Low Kelvin = warm: a candle is about 1,850 K, overcast daylight about 6,500 K.
- A 3,200 K setting describes the light, so the camera adds blue to cancel it.
- Equal changes look equal in mireds (1,000,000 / K), not in kelvin.
- Human colour constancy has a blue bias: eyes forgive daylight, cameras forgive nothing.
- Learned auto white balance (FFCC) cut error to 1.80° from 6.36° for grey world.
- India has fitted over 1.34 crore LED street lights, so night scenes now mix cool and warm sources.
The confusing part is the direction of the numbers. Physically hotter means bluer, so "warm" orange light has the low number. And when you set a camera to 3,200 K you are not adding warmth; you are telling it the light is 3,200 K, so it adds blue to cancel it.
Once you see white balance as "describe the light, then subtract it", the controls stop being backwards.
Why does warm light have a low Kelvin number?
Because colour temperature comes from black-body physics, where cooler objects glow redder. Wien's displacement law puts the peak of a black body's emission at λ = b/T, with b ≈ 2,898 µm·K: a 2,000 K filament peaks around 1,450 nm in the infrared and spills mostly red and orange into the visible band, while the Sun's surface at about 5,778 K peaks near 500 nm and looks white.
That is the answer to the Photography Stack Exchange question about whether colour-temperature kelvin is "real" kelvin. It is the same unit, applied to an idealised radiator. Real lamps are matched to the nearest point on the curve of black-body colours (the Planckian locus), which is why the figure is called correlated colour temperature, or CCT.
Sources that sit off that curve, such as fluorescent tubes and many LEDs, also need a green-magenta "tint" correction that the Kelvin number cannot describe.
In the 1940s and 1950s, North American photo labs balanced the colour of their prints against a reference card. These Shirley cards showed a single light-skinned woman in brightly coloured clothes, and very few of them showed an adult man. Light skin became the standard a print was corrected to, and the bias went down into the film: colour stock for still cameras was designed with a positive bias toward light skin, because it reflects more light. By the mid-1990s Japanese companies had redesigned their cards from their own colour-preference tests, with Japanese women as the reference. In 1995 Kodak made a multiracial card showing three women, Caucasian, Asian and African, in contrasting clothes. The researcher Lorna Roth traced this history in a 2009 paper. A colour correction is only as neutral as the reference someone chose for it.
Wikipedia: Color chart, Shirley cards (after Lorna Roth, 2009)See the data as a table
| Colour temperature (K) | |
|---|---|
| Match flame | 1700 |
| Low-pressure sodium street lamp | 1700 |
| Candle flame | 1850 |
| Sunrise or sunset | 1850 |
| Standard incandescent bulb | 2400 |
| Warm white CFL or LED | 3000 |
| Studio tungsten lamp | 3200 |
| Cool white or daylight tube and LED | 5000 |
| Electronic flash | 5500 |
| Overcast daylight | 6500 |
| Clear blue sky (low) | 15000 |
Source: Wikipedia: Color temperature, table of common sources · Flash is 5,500 to 6,000 K and clear polar-facing sky 15,000 to 27,000 K; the lower value is shown. Low-pressure sodium is nearly monochromatic, so a CCT barely describes it.

How does white balance actually correct a photo?
It scales the red, green and blue channels so that the estimated colour of the light becomes neutral grey. Under a 3,000 K lamp the raw sensor data has strong red and weak blue for a white wall; the camera multiplies blue up and red down (usually keeping green fixed) until the wall's three channels match, and every other colour shifts with it.
That per-channel gain is the simplest model of chromatic adaptation, and it is what you change when you move the Kelvin and tint sliders in a raw editor:
- the Kelvin slider moves the target along the black-body curve, between orange and blue;
- the tint slider moves it across the curve, towards green or magenta.
A JPEG has already had one correction applied and its colours compressed, so large changes afterwards stretch noise and clip channels. Shooting raw keeps the gains reversible, one of several reasons covered in RAW, dynamic range and ETTR.
Why are equal Kelvin steps not equal changes?
Because perceived colour change tracks the reciprocal of temperature. The mired scale, one million divided by the Kelvin value, makes equal steps look roughly equal; Irwin Priest observed in 1932 that the just-noticeable difference between two lights depends on the difference of their reciprocal temperatures.
| Change | Kelvin step | Mired step | Visible? |
|---|---|---|---|
| 2,500 to 3,000 K | 500 K | 67 mireds | Very |
| 6,000 to 6,500 K | 500 K | 13 mireds | Hard to see |
Lighting gels are rated this way. Converting 5,700 K daylight to match 3,200 K tungsten takes about −137 mireds, a full colour temperature blue (CTB) filter. The calculator below does the conversion and also shows where each temperature's black-body emission peaks.
Kelvin, mireds and the shift between two lights
How do your eyes and the camera disagree?
Your visual system discounts the light colour automatically, a process called colour constancy, but it does so unevenly: it adapts almost perfectly to bluish daylight and less well to warm or greenish light. A camera records the cast faithfully, so a room that looked white to you shows up orange or green in the photo.
The authors measured colour constancy across 40 illumination hues using achromatic adjustments: observers set a patch to look grey under each light. Constancy was essentially perfect under bluish daylight-like illumination and fell off for lights in other directions, with participants undershooting and showing a strong bias towards accepting bluish light as neutral. Bluish illuminants were sometimes named as achromatic at all. Blue bias and naming consensus alone explained more than 68% of the variance in the adjustments, and all factors together more than 94%. The authors conclude that human colour constancy is tuned to daylight. The limit: lab displays with simulated illuminants, not real rooms.
Read the paper (free)This blue bias explains a familiar complaint. Photos under warm tungsten or sodium light look far more orange than the scene felt, because your eyes corrected most of the cast and the camera did not; shade under a blue sky looks slightly blue in photos but barely registers in person.
Many photographers deliberately leave some warmth in indoor shots because a fully neutral correction looks wrong against memory.
How does auto white balance work?
By guessing the light from the image. There are three families of method:
- Grey world, the oldest, assumes the average of a scene is neutral and scales the channels until it is.
- White patch assumes the brightest pixels are white.
- Learned models, used by modern cameras and phones, are trained on thousands of photos with a measured light colour.
The first two fail on scenes dominated by one colour, such as a blue wall or a field of grass, which the camera "corrects" towards grey.
The top answer on the Stack Exchange thread about auto white balance describes the same history: start from the grey-world assumption, then constrain guesses to plausible illuminants. The research benchmark below shows how far learned methods have moved. Error is the angle in degrees between the estimated and true colour of the light; lower is better.
See the data as a table
| Mean error (°) | |
|---|---|
| White-patch | 7.55 |
| Grey-world | 6.36 |
| Shades-of-grey | 4.93 |
| Bayesian | 4.82 |
| Cheng et al. 2014 | 3.52 |
| Convolutional Color Constancy (2015) | 1.95 |
| FFCC (2017) | 1.8 |
| FFCC with metadata and semantics | 1.61 |
Source: Barron and Tsai, Fast Fourier Color Constancy, CVPR 2017, Table 1 · Angular error between estimated and true illuminant colour. FFCC row is the full two-channel model (Model J).
FFCC treats white balance as a localisation problem: it builds a histogram of each image's colours in log-chrominance space, wraps it onto a small torus so the work can be done with fast Fourier transforms, and learns a filter that finds the most likely illuminant. On the standard Gehler-Shi benchmark its full model cut mean error to 1.80° (median 0.95°), against 6.36° for grey world, and a version given camera metadata and semantic features reached 1.61°. It lowered error by 13 to 20% versus the previous best while running 250 to 3,000 times faster, with its fastest thumbnail version taking about 1.44 ms per image on a 2016 Pixel XL (the abstract quotes around 700 frames per second on a phone), fast enough for a live viewfinder, and it outputs a full probability distribution that allows smooth white balance across video frames. The limit: it assumes a single light colour for the whole frame, which mixed lighting violates.
Read the paper (free)The single-illuminant assumption is exactly where phones still struggle, and why computational pipelines now segment scenes and treat faces, sky and background separately. I cover that pipeline in how phone cameras use computational photography.
Why do Indian streets at night photograph so strangely?
Because they mix light sources with very different spectra, and that mix has changed fast. Low-pressure sodium lamps emit almost all their light at 589.0 and 589.6 nm, a single yellow-orange colour that no white balance can correct, because there is no other wavelength to rescale.
High-pressure sodium, the familiar orange street light, has a broader but still warm spectrum; sodium lamps as a family span roughly 1,400 to 2,700 K. White LED street lights, typically 4,500 to 5,000 K, have replaced much of it.
India's LED switch
The scale of that change in India is large. Under the Street Lighting National Programme, launched in January 2015, Energy Efficiency Services Ltd (EESL) had installed more than 1.34 crore LED street lights by 6 January 2025, according to a Ministry of Power document:
- saving over 9,001 million units of electricity a year (Ministry of Power document);
- cutting peak demand by more than 1,500 MW;
- in Hyderabad alone, about 5.53 lakh lights retrofitted.
The result for photographers: a main road is now lit cool white, while shopfronts, dhabas and homes along it still use warm LEDs, old tube lights and the odd sodium lamp.
| Light you meet at night in India | Approx. colour temperature | What it does to a photo | Fix |
|---|---|---|---|
| Low-pressure sodium (older highways, industrial) | About 1,700 K, near-monochromatic | Everything turns one shade of orange; colours cannot be recovered | Convert to black and white, or add your own light |
| High-pressure sodium | Within the sodium range of about 1,400 to 2,700 K | Strong orange, weak blues | Set 2,500 to 3,000 K, then check skin |
| LED street light (SLNP era) | Typically 4,500 to 5,000 K | Near neutral, can lean green | Daylight setting plus a little magenta tint |
| Cool white tube light, CFL | About 5,000 K plus a green spike | Greenish skin | Add magenta in tint |
| Warm white LED bulbs in homes and shops | About 2,700 to 3,000 K | Orange interiors | Set 3,000 K, leave some warmth |
Three ways to handle mixed light
For mixed scenes, the highest-voted answer on the Stack Exchange mixed-lighting thread lays out the real choices:
- Control the light: switch off or gel one source.
- Pick the subject's light and balance for that.
- Fix it locally in editing with masks.
A practical habit from that thread is to note one neutral object in the subject's own light and click it with the eyedropper later. If you shoot on the street, my street photography guide for India covers working with this light rather than against it.

Should you use auto white balance or set it yourself?
Use auto for changing daylight and casual shooting, and lock a fixed Kelvin value for anything you will edit as a set. Auto re-guesses every frame, so a sequence of shots, or a video pan from a lamp-lit room to a window, drifts in colour. A fixed setting keeps frames consistent, and in raw it can be adjusted for all of them at once.
- Daylight sequences: set about 5,500 K and leave it; sunsets then stay golden instead of being "corrected" to grey.
- Indoors under one lamp type: use a custom white balance from a grey card or white paper in the subject's position.
- Video: never leave auto on for a take; set Kelvin manually so the colour cannot shift mid-shot.
- Night streets: start at a setting between the sodium and LED temperatures in the table above, shoot raw, and decide in editing which source should look neutral.
Colour is half of exposure; the other half is in the exposure triangle in stops. For how this fits with the rest of the design and photography writing, see Kala.
What's new
- Samsung researchers show a 5,000-parameter auto white balance model that uses capture time and location, trained on 3,224 phone images Samsung Research (ICCV 2025), Oct 2025
- Hyderabad's GHMC retrofits about 5.53 lakh street lights with LEDs, saving 238.83 million units a year NewsMeter, Dec 2025
- Ten years of UJALA and SLNP: over 1.34 crore LED street lights installed nationwide PIB / Ministry of Power, Jan 2025
Questions people ask
Is colour temperature kelvin the same as real temperature?
Yes, it is the same unit. A light's colour temperature is the actual temperature an ideal black body would need to glow that colour. Incandescent filaments really are near their rated temperature; LEDs and fluorescent tubes are cool to the touch and only match the colour.
Why does setting 3200 K make my photo bluer?
Because the setting describes the light, not the look. Telling the camera the light is 3,200 K makes it add blue to neutralise that orange light. If the scene is actually lit by daylight, the result turns blue.
Can white balance fix sodium street light?
Not for low-pressure sodium. It emits nearly all its light at about 589 nm, so every surface reflects the same single wavelength and there is no colour information left to rebalance. High-pressure sodium has a broader spectrum and can be partly corrected.
Why does auto white balance ruin sunsets?
Auto white balance assumes the scene should average to neutral, so it reads a golden sky as a colour cast and pulls it towards grey. Set daylight (about 5,500 K) to keep the warmth you saw.
Sources
- Wikipedia: Color temperature (table of source temperatures, Planckian locus)
- Wikipedia: Wien's displacement law (b = 2,898 µm·K, Sun peak near 500 nm)
- Wikipedia: Mired (definition, Priest 1932, −137 mired CTB example)
- Wikipedia: Sodium-vapor lamp (589.0 and 589.6 nm lines, LED replacement at 4,500 to 5,000 K)
- Wikipedia: Color chart, Shirley cards (story: 1940s–50s lab reference cards, film bias, mid-1990s Japanese cards, Kodak's 1995 multiracial card; after Roth, Canadian Journal of Communication, 2009)
- Barron and Tsai: Fast Fourier Color Constancy, CVPR (2017)
- Barron: Convolutional Color Constancy, arXiv (2015)
- Weiss, Witzel and Gegenfurtner: Determinants of Colour Constancy and the Blue Bias, i-Perception (2017)
- Ministry of Power / PIB: 10 years of UJALA and SLNP (2025)
- EESL: Street Lighting National Programme
- NewsMeter: GHMC replaces over 5.53 lakh street lights with LEDs (2025)
- Samsung Research: Time-Aware Auto White Balance in Mobile Photography, ICCV (2025)
- Photography Stack Exchange: How does Kelvin for color temperature relate to Kelvin for actual temperature?
- Photography Stack Exchange: Does auto white balance really work? How?
- Photography Stack Exchange: How to white-balance photos shot in mixed-lighting environments?
