Illumination, surface reflectance, sensor response, and display processing all affect recorded color.
By the end: Explain why a global white-balance change cannot repair every mixed-light photograph.
A surface reflects some wavelengths more than others. The illuminant determines which wavelengths arrive in the first place. Sensor channels have broad, overlapping spectral responses, rather than measuring pure red, green, and blue wavelengths separately. Processing then maps those responses into an output color space.
Two surfaces can match under one light and differ under another because their spectra differ. A light source with missing spectral energy cannot always reproduce colors that a broader source reveals. White balance is not a way to recover all missing spectral information.
Imagine two matte cards under the same lamp. To make the arithmetic visible, divide the arriving light into three broad wavelength bands. Suppose the lamp supplies 100 units in each band. A neutral card reflects one half of each, so 50, 50, and 50 units leave it toward the camera in this simplified arrangement. A second card reflects eight tenths, four tenths, and one tenth, so its reflected amounts are 80, 40, and 10. These are invented light quantities, not screen RGB values.
Now replace the lamp with one supplying 100, 60, and 20 units in those same bands. Keep the cards, geometry, and camera position fixed. The neutral card now returns 50, 30, and 10 units. The second returns 80, 24, and 2. Multiply the arriving amount by the reflected fraction separately in each band. Neither surface changed, but both reflected spectra changed. A photograph can therefore show different colors without any change to the objects themselves. The illumination is part of the measurement.
The three bands only explain multiplication. Real light contains a continuous spread of wavelengths, and real sensors do not place every wavelength into one exclusive bucket. Each sensor channel responds to a broad, overlapping range. Light at one wavelength can contribute to more than one channel. The channel value sums contributions across wavelengths, weighted by that channel's response. Three recorded numbers therefore compress a much richer physical description. They cannot specify every wavelength that reached the camera or reconstruct the surface's complete reflectance.
This compression explains a useful failure case. Imagine two different reflected spectra that both produce channel readings of 60, 60, and 60 under one illumination. The camera cannot distinguish them from those readings alone. A different lamp can change their readings unequally because their hidden spectral differences interact with the new light. A match under one lamp is not proof of identical material color under every lamp. This conditional matching is called metamerism. Human observers can also encounter it, with responses different from a camera's.
Consider an even stronger limit. In the three-band example, let a replacement lamp supply 100, 60, and zero units. A card reflecting everything in the last band and a card reflecting nothing there both return zero from that band. Multiplying the recorded zero by a white-balance gain still gives zero. The missing illumination prevented that difference from being measured. With real sources and overlapping channels the failure need not be so absolute, but a channel correction cannot create an independent measurement of absent spectral information.
Use these cases to interpret colored fabric under shop lighting and then daylight. First ask whether the illumination changed, rather than assuming the camera selected an incorrect color setting. Then distinguish two goals. One is to record appearance under the actual light. The other is to compare materials under a controlled reference light. Those goals can justify different renderings. A schematic can explain why the readings change, but neither its colored shapes nor an uncalibrated display certifies how the fabric would look to a particular observer.
| Last-band quantity | Card A | Card B |
|---|---|---|
| Arriving light | 0 units | 0 units |
| Reflected fraction | 1 | 0 |
| Returning light | 0 × 1 = 0 | 0 × 0 = 0 |
| After a gain of 5 | 0 | 0 |
White balance attempts to make an assumed neutral surface appear neutral under the chosen rendering. A simple model multiplies channels by different gains. If a gray patch records 120, 100, and 80, gains of 1, 1.2, and 1.5 equalize them at 120. Real processing includes color matrices and nonlinear operations as well.
A raw white-balance setting usually guides later rendering, though camera behavior varies. In a JPEG, much of the processing is already baked in and strong corrections can expose clipping or quantization. Kelvin settings roughly describe a blue-yellow axis related to illuminants; a separate tint adjustment often handles a green-magenta axis. Setting a higher Kelvin value commonly warms the correction, rather than making the photograph bluer.
The original reference has channel readings of 120, 100, and 80. Assume these numbers are proportional to measured light, have any electronic offset removed, and remain below a recording limit. To equalize them at 120, leave red unchanged, multiply green by 120 divided by 100, and multiply blue by 120 divided by 80. The gains are 1, 1.2, and 1.5. This calculation describes a small part of rendering. It does not change the light already collected by any sensor area.
Equalizing at 120 is a choice, not a unique physical solution. Equalizing at 80 would use gains of two thirds, 0.8, and 1. Equalizing at 60 would use 0.5, 0.6, and 0.75. Each set has the same proportions between gains. The 60-target set is half the 120-target set, so every corrected channel is half as large. The patch remains neutral within the model, but its linear level changes. A separate tone mapping determines how bright either numerical result appears on the final display.
| Target level | R gain | G gain | B gain | Result |
|---|---|---|---|---|
| 120 | 1 | 1.2 | 1.5 | 120, 120, 120 |
| 80 | 2/3 | 0.8 | 1 | 80, 80, 80 |
| 60 | 0.5 | 0.6 | 0.75 | 60, 60, 60 |
This distinction matters when a supposedly neutral reference is too bright. Take a constructed rendering stage whose largest allowed value is 255. A pale patch measured as 240, 200, and 160 becomes 240, 240, and 240 with the original gains. A brighter colored patch measured as 240, 180, and 190 becomes 240, 216, and 285. If this stage clips at 255, the last value becomes 255. The correction has changed a channel relationship because 285 cannot survive that limited output stage.
Reducing all corrected values together before that output limit can preserve their ratios. Half of 240, 216, and 285 is 120, 108, and 142.5, all below 255. That operation does not recover anything if the capture itself already clipped a channel. It only prevents an avoidable later clipping step in this invented pipeline. Raw measurements, working editing values, and exported image values can have different limits. Do not read this arithmetic as a claim that raw files contain ordinary display values capped at 255.
A channel gain also scales variation in that channel. If blue has a mean reading of 80 and a typical fluctuation of 8, multiplying by 1.5 gives a mean of 120 and a fluctuation of 12. The signal divided by fluctuation remains 10 in this simple scaling example. White balance can make weak blue-channel noise more visible without having caused the original shortage of blue light. Increasing ISO or an editing brightness setting also cannot replace photons that were never collected in that channel.
A useful reference must be neutral for the intended task, under the light reaching the subject. A white page may have optical brighteners, a glossy object may reflect a lamp directly, and an apparently gray wall may contain a color cast. A dedicated reference can help, but it does not remove every spectral or processing limit. If no reference is available, compare plausible corrections and describe the uncertainty. Choosing a pleasing face color is a rendering judgment, not proof that you recovered a uniquely correct illumination spectrum.
| Processing stage | R | G | B |
|---|---|---|---|
| Unclipped input | 240 | 180 | 190 |
| Gains 1, 1.2, 1.5 | 240 | 216 | 285 |
| Clip directly at 255 | 240 | 216 | 255, ratio lost |
| Instead halve before limit | 120 | 108 | 142.5 |
| Channel | Measured neutral patch | Gain | Corrected |
|---|---|---|---|
| R | 120 | 1 | 120 |
| G | 100 | 1.2 | 120 |
| B | 80 | 1.5 | 120 |
Extend the neutral-reference example to two identical gray cards. Place one where lamp A dominates and the other where lamp B dominates. Suppose their measured linear channels are 120, 100, 80 and 80, 100, 120 respectively. These values are assumptions, not a prediction for a window or a household bulb. The first card needs blue multiplied more than red. The second needs red multiplied more than blue. Because both cards are known neutral, their differing channel ratios are evidence of different illumination in this constructed case.
Apply the first card's gains of 1, 1.2, and 1.5 to the entire image. The first card becomes 120, 120, 120, while the second becomes 80, 120, 180. Alternatively, gains of 1.5, 1.2, and 1 neutralize the second card at 120 but change the first to 180, 120, 80. Both calculations use the same multiplication rule. The failure is not that the gains were applied incorrectly. Each correction answers a different local illumination problem, while a global operation treats both regions alike.
You can prove the conflict without trying many settings. To neutralize the first card, the blue gain must be 1.5 times the red gain, because 120 divided by 80 is 1.5. To neutralize the second, the blue gain must be two thirds of the red gain, because 80 divided by 120 is two thirds. A single positive gain ratio cannot equal both numbers. A compromise can reduce the casts, but it cannot make both of these references exactly neutral within this multiplication-only model.
Real mixed lighting can overlap on one surface instead of occupying separate regions. Assume the contributions add linearly. At a location receiving half of each example illumination, the neutral response is the average of the two cards' readings, or 100, 100, 100. Applying the first correction gives 100, 120, 150 and introduces a cast there. Not every mixture is neutral. It says the suitable correction can change continuously as the relative light contributions change.
This creates a practical difficulty for local editing. A hard boundary around half a face rarely matches the gradual illumination mixture across its curved surface. Hair, translucent skin, and shiny highlights can have different contributions again. A local correction with a soft transition may look better, but it is an estimate of appearance, not a full reconstruction of the original light field. If accurate comparison matters, changing or matching the lights before capture may be more reliable than trying to assign a separate correction to every image region.
Try the model as a prediction task before using any camera. Compute the corrected values for a third neutral location with readings 110, 100, and 90. The first gains give 110, 120, and 135. It is less unequal than the second card's result, yet still not neutral. That location is closer to the first illumination in channel ratio, but not identical to it. The lesson is about evidence for a local correction. A genuinely colored surface could produce the same numbers and would not justify neutralization.
| Neutral location | Before gains | After A gains |
|---|---|---|
| All A | 120, 100, 80 | 120, 120, 120 |
| Three quarters A, one quarter B | 110, 100, 90 | 110, 120, 135 |
| Half A, half B | 100, 100, 100 | 100, 120, 150 |
| All B | 80, 100, 120 | 80, 120, 180 |
A window and a warm lamp can illuminate different regions of one face. One global set of channel gains cannot necessarily neutralize both. You can change or match the lights, choose one as the intended dominant source, apply local adjustments, or preserve the color difference deliberately.
For a controlled test, photograph a known neutral matte object under two lights, keeping other settings stable if possible. Avoid using an arbitrary colored wall as a neutral reference. Display calibration and ambient viewing light also affect judgment; no exact color match is guaranteed by an uncalibrated screen or by these schematic patches.
A controlled comparison starts with a question narrow enough for the available evidence. Ask whether one neutral reference changes relative to another across a scene, rather than whether every color is correct. Put the same matte object in the two locations in turn, or use two matching references if available. Keep its orientation similar so that reflections do not become a new variable. A face is not an established neutral reference.
If your device allows white-balance lock, you can keep one rendering choice across the comparison. If it does not, include both references in one frame so a global automatic change affects them together. Even then, a phone may process regions differently, merge frames, or apply local color adjustments. A visible difference is evidence about that finished image, not a complete measurement of the light sources. Without a camera, use the two-card calculations to predict which reference each correction would neutralize and which it would leave colored.
Suppose two references look similar in one image and different in the next. Automatic white balance alone could move both toward another global balance, but our multiplication-only model cannot arbitrarily change their relative channel ratios. Local processing, changed lighting, exposure clipping, or different reflections are candidate explanations. Do not identify a single cause from appearance alone. Compare the scene arrangement and any available metadata, then repeat a simpler version under one light. Hidden processing limits the conclusion.
| Comparison | Evidence available | Still unresolved |
|---|---|---|
| Two known grays in one frame | Relative casts in that output | Local processing and spectra |
| Same gray in separate automatic frames | Combined capture and rendering changes | Whether the correction stayed fixed |
| Arbitrary wall made gray | A chosen edit | Whether the wall was neutral |
| Repeat under one light | A simpler illumination comparison | Display accuracy and material spectra |
For an expressive sunset image, the objective is different. An exactly neutral card can remove some of the warmth that communicates the illumination. Keep a neutral rendering as a comparison if useful, then judge the warmer alternative against the intended appearance. For product documentation, consistent lighting and a reference are more important because viewers may compare samples. Neither goal guarantees accurate reproduction on another person's screen. A color space describes how numbers relate to colors, while a particular display still has its own response and viewing conditions.
Output also changes what counts as an acceptable result. A small image can hide fine colored noise or narrow colored edges that a large crop exposes. Conversely, judging only a magnified crop can exaggerate defects that are invisible in the delivered photograph. Compare the intended output with a closer inspection, keeping the purposes separate. A warm cast is not the same failure as a soft edge. White balance cannot restore optical focus, and sharpening cannot identify missing color differences from an already clipped channel or an unilluminated spectral region.
Finish a comparison with a specific statement supported by what you observed or calculated. For example, one gain set neutralized the reference beside the lamp but left the reference near the window blue in the schematic model. A useful next capture would reduce the competing illumination or include both references again after a lighting change. That conclusion is narrower than claiming the camera records wrong colors. It names the correction's scope, preserves intentional lighting color when appropriate, and leaves device-specific processing and display accuracy as separate questions.
Apply gains 1, 1.2, 1.5 to another patch measured as 80, 100, 120. Does it become neutral?
Multiply each channel separately.
| R | G | B |
|---|---|---|
| 80 × 1 = 80 | 100 × 1.2 = 120 | 120 × 1.5 = 180 |
It becomes 80, 120, 180, which is not neutral. A correction chosen for the first illumination does not necessarily neutralize another. The difference could also be a genuinely colored surface, so establish the reference before interpreting it as mixed light.
A sunset scene looks warm. Must a technically correct edit make a white object perfectly neutral?
Separate a measurement goal from an expressive goal.
| Purpose | Possible choice |
|---|---|
| Product comparison | Controlled light and neutral reference |
| Sunset atmosphere | Retain intentional warmth |
No. Neutrality may be appropriate for documenting a product, while warmth may be essential to representing the sunset’s appearance. Keep a reference when accurate reproduction matters, but state the intended result instead of treating neutral white balance as a universal artistic rule.
The lessons and answers stand alone. These optional secondary references explain the optics and imaging concepts. Numerical examples are constructed, not measured camera specifications. Visual judgments depend on the intended image; the course does not claim a universal aesthetic score.
Channel scaling, neutral references, and color correction. The three-number example is not a spectral camera profile.