Short answer
A camera makes three decisions that fall hardest on faces: how bright to make the picture, what colour to call the light, and how to shape tone and saturation afterwards. Each is tuned against reference images, and for much of photography's history those references covered a narrow range of skin. A face that is much darker or lighter than the pipeline assumes, or lit against a bright background, is where exposure, flare and white balance go wrong. This page is about the engineering of cameras. It does not sort or describe people.
Automatic exposure chooses a single setting for a scene from the brightness of what it sees, usually weighted towards the centre and towards anything detected as a face. Google's published HDR+ exposure method does exactly this: it compares the frame's brightness distribution against thousands of hand-tuned examples and strongly boosts the weight of regions where faces are detected. Two things make that fragile. If the face is not detected, the weighting is lost and the background decides. And what counts as a well-exposed face comes from the examples the system was tuned on. A subject much darker than those examples in front of a bright wall is under-rendered; a much lighter one in a dark room is blown out. Neither is a property of the person. It is the meter's assumption meeting a scene outside it.
Some of the light entering a lens scatters inside it and lands as a thin, even veil over the whole image. A window or bright sky behind the subject makes far more of it. Because the veil is roughly the same everywhere, it is a small addition to a bright surface and a large one to a dark surface, which is lifted, greyed and robbed of contrast. The table puts numbers on this for plain surfaces of different reflectance: a veil equal to 3% of the scene's white raises a 3% surface by a full stop and a 60% surface by less than a tenth of one. Google lists a dedicated stray-light algorithm among its camera changes for this reason, describing backlit faces left washed out or in shadow.
The cure at capture is old and optical: shade the lens, keep the brightest light source out of frame, and clean the glass.
Skin gets its colour from two absorbers. Melanin in the outer layer absorbs more strongly the shorter the wavelength, and haemoglobin in the blood beneath is the main absorber in the deeper layer, so the light that comes back is always weighted towards the long-wavelength end. A close portrait is therefore a frame dominated by one warm colour, which is the case automatic white balance handles worst: it cannot tell warm skin in neutral light from neutral skin in warm light, and an algorithm that assumes scenes average to grey will cool the picture to compensate. Under mixed or strongly coloured light the error grows. A neutral reference in the frame settles it, which is why a grey card matters more for portraits than for landscapes.
Every automatic decision above is tuned by comparing output with references, and the references have a history. Lorna Roth's study of the 'Shirley' reference cards used to set colour balance describes film emulsions and later digital cameras as designed around light skin, with other skin rendered poorly until compensating practices and technical changes arrived. The ColorChecker chart of 1976 carries two skin patches among its 24. Google's account of its recent phone work lists the stages it retuned: face detection, white balance, exposure, stray light and low-light sharpness, using image sets it describes as 25 times more diverse than before. In 2022 it released the ten-step Monk Skin Tone Scale as an instrument for checking that technology performs evenly across a range.
The reliable methods take the guesses away from the camera. Expose for the face: meter it, or use exposure compensation until its brightest part sits below clipping in every channel, checking the red channel in particular. Set white balance from a grey card held where the face is. Keep bright windows and lamps out of the frame or shade the lens, since flare cannot be corrected well afterwards. Shoot raw when the light is difficult, because exposure and balance can then be revised without the damage an 8-bit JPEG suffers. Finally, judge the result on a calibrated screen next to the person or a physical reference, not from memory and not against another camera's version.
| Stage | What it decides | How it goes wrong | Documented engineering response |
|---|---|---|---|
| Face detection | Whether exposure and focus are weighted to the face at all | Undetected faces leave the background in charge | Detector retrained on more varied images and lighting (Google) |
| Automatic exposure | Overall brightness of the frame | A face much darker or lighter than the tuning examples is under- or over-rendered | Face-weighted exposure (HDR+ paper); exposure tuning revised (Google) |
| Automatic white balance | What colour the light is taken to be | A frame dominated by warm skin is cooled; mixed light leaves a cast | White balance models retuned against more varied portraits (Google) |
| Stray light | Contrast in the darker parts of the image | Backlight veils dark surfaces far more than light ones | Dedicated stray-light reduction (Google) |
| Tone and colour tuning | Warmth, saturation and contrast of the final picture | Tuned for a preferred look that suits some subjects better than others | Evaluation against a published ten-step scale (Google, 2022) |
| Surface reflectance | Flare = 1% of scene white | Flare = 3% of scene white |
|---|---|---|
| 3% | +0.42 | +1.00 |
| 9% | +0.15 | +0.42 |
| 18% | +0.08 | +0.22 |
| 36% | +0.04 | +0.12 |
| 60% | +0.02 | +0.07 |
| 90% | +0.02 | +0.05 |
Why: The bright background set the exposure, and flare from it then flattened what was left.
Fix: Turn the subject so the light falls on the face, add exposure compensation, and shade the lens.
Why: Automatic white balance read the frame's warmth as a colour cast and cooled it.
Fix: Set white balance from a grey card at the face, or use a preset for the actual light source.
Why: The red channel clipped on the brightest skin, so the channel ratios changed.
Fix: Reduce exposure until the red histogram clears the right edge; fill the shadows with a reflector instead.
Why: One exposure and one tone curve are serving surfaces of very different reflectance and lighting.
Fix: Even out the light on the two faces first; then shoot raw and adjust each area locally.
Each statement is labelled by kind — established fact, a standard’s requirement, observed market data, a convention, or Colourwise’s own interpretation or analysis — with the strength of the evidence behind it.
FactStrong evidence
The published HDR+ auto-exposure method matches a frame's brightness distribution against a database of about 5,000 hand-tuned scenes and strongly boosts the weight of regions where faces are detected.
Caveat: One manufacturer's method as published in 2016.
FactStrong evidence
Melanin in the epidermis absorbs visible light more strongly at shorter wavelengths, and haemoglobin is the main absorber in the dermis, so the spectrum returned from skin is weighted towards longer wavelengths.
Caveat: An optical mechanism, cited to explain camera behaviour and not to describe any person.
Source: The optics of human skin (Anderson & Parrish, Journal of Investigative Dermatology 77, 13–19, 1981); Skin Optics Summary (Steven L. Jacques, Oregon Medical Laser Center News, January 1998)
FactModerate evidence
Google states that for its Real Tone work it retrained face detection on more varied images, retuned automatic white balance and automatic exposure, and added an algorithm to reduce the effect of stray light, which it says washes out darker skin most.
Caveat: The manufacturer's own description of what it changed. It publishes no measurements, so this is evidence of which pipeline stages matter, not of how accurate the result is.
FactModerate evidence
A peer-reviewed history of colour-balance reference cards describes colour film emulsions and digital camera design as having been balanced around light skin, with compensatory practices and technical improvements developed later.
Caveat: Cited from the article's abstract; the full text could not be retrieved for this review, so no detail beyond the abstract is relied on.
FactModerate evidence
Google openly released the ten-shade Monk Skin Tone Scale on 11 May 2022 as a tool for developing and evaluating technology across a range of skin tones.
Caveat: An evaluation instrument for engineers. Colourwise does not use it, or any scale, to classify people.
Source: Improving skin tone representation across Google (Monk Skin Tone Scale announcement)
Colourwise analysisModerate evidence
A uniform veil of flare equal to 3% of scene white raises the recorded level of a 3% reflectance surface by 1.0 stop and of a 60% reflectance surface by 0.07 stop.
Based on: Colourwise arithmetic: log₂ of (reflectance + flare) ÷ reflectance, for plain surfaces under even light. The cited source supplies the observation that stray light affects darker subjects most; the numbers are ours. See the table.
Caveat: Real flare is uneven and depends on the lens and the position of the light source.
ConventionModerate evidence
Camera makers commonly tune for pleasing rather than accurate colour, and a camera-testing vendor notes that this includes warm, slightly saturated skin rendering.
Caveat: A testing vendor's general observation, not a survey of current cameras.
Reviewed 1 October 2026. Colourwise summarises its sources in its own words and does not reproduce standards text or proprietary colour data. Spotted an error? Tell us.