Short answer
White balance multiplies the camera's red, green and blue channels by different amounts so that something neutral in the scene comes out neutral in the picture. It stands in for the adaptation your visual system does automatically. The hard part is not the maths but the guess: from the image alone, the camera cannot always tell a white object in blue light from a blue object in white light, so automatic white balance fails in dim, strongly coloured or mixed light.
Under a warm tungsten lamp a white wall reflects more red than blue, and the raw sensor data records exactly that. Your eyes adapt within seconds and the wall still looks white; the camera does not adapt, so it must correct. White balance scales each channel — in a published smartphone pipeline, a linear scaling of the four RGGB channels so that greys in the scene map to greys in the image — and every other colour shifts along with the neutral. Kelvin and tint sliders in editing software are a friendlier interface to the same gains, expressed as the colour temperature and green–magenta offset of the light being corrected for.
Google's Night Sight engineers put the problem plainly: is the snow blue, or is it white snow lit by blue sky? The pixel values can be identical. Auto white balance is an ill-posed problem, solved with assumptions — that the average scene is grey, that the brightest patch is white, that certain colours such as skin and foliage are likely — and increasingly with learned models trained on correctly balanced photographs. Those assumptions hold for ordinary scenes and fail for scenes dominated by one colour: a close-up of an orange wall, a green-lit stage, a sunset, sodium street light.
Fully neutralising the light is not what a person remembers seeing. Candlelight looks warm even after adaptation; dusk looks blue. Many cameras deliberately leave some of the warmth in low-colour-temperature scenes, and photographers often choose a balance for mood rather than neutrality. For documentation — product colour, artwork, evidence of a defect — neutral is the goal; for a portrait at a fireside it may not be. The choice is part of the photograph, which is why no single balance is correct for every purpose.
White balance corrects for one light. When two lights of different colour fall on a scene, no global setting neutralises both; see mixed lighting.
The most reliable white balance is measured, not guessed. Photograph a grey card or a neutral target in the same light as the subject and set a custom white balance from it, or click it with the eyedropper when editing a raw file. Choosing a preset for the light source is the next best option. If the camera is left on auto, shooting raw keeps the choice open, because white balance in a raw file is metadata applied at conversion rather than baked into the pixels.
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
White balancing in a published smartphone pipeline linearly scales the RGGB channels so that greys in the scene map to greys in the image, using scale factors supplied by the image signal processor.
FactModerate evidence
Auto white balance is an ill-posed problem: the same image data can arise from a coloured object under neutral light or a neutral object under coloured light, and conventional algorithms struggle in very dim or strongly coloured light.
FactModerate evidence
Google developed a learning-based auto white balance for Night Sight, trained to distinguish well-balanced from poorly balanced images captured on Pixel phones.
Caveat: One manufacturer's published approach; others are not publicly documented in the same detail.
Colourwise interpretationModerate evidence
Camera white balance is an engineered substitute for the visual system's chromatic adaptation, and like adaptation it can only discount a single dominant illuminant at a time.
Based on: Follows from white balance being a global per-channel scaling, as described in the HDR+ paper, compared with the colour-constancy literature summarised on the lighting pillar.
Reviewed 29 September 2026. Colourwise summarises its sources in its own words and does not reproduce standards text or proprietary colour data. Spotted an error? Tell us.