The screen is not the room

A display’s contrast ratio is measured in a dark room, and a display is not used in one. The number on the specification is a property of the panel; the number a reader gets is a property of the panel and the room together, and the room usually wins.

The claim #

Light falling on a screen is reflected back at the observer and adds to every pixel equally, so it is negligible for the bright ones and ruinous for the dark ones. A calibration cannot remove it, because it is not in the signal.

The arithmetic is one line. An illuminance $E$ falling on a surface returning a fraction $R$ of it diffusely leaves as a luminance

$$L_{\text{added}} = \frac{E R}{\pi}$$

and that luminance is added to whatever the panel is emitting. At 300 lux — an ordinary office — and a two per cent screen reflectance, it is 1.91 candelas per square metre.

That is nothing beside a 300-candela white and it is six times a 0.3-candela black.

What it costs #

Two of those rows deserve to be read twice.

At 300 lux the thousand-to-one panel is a 137-to-one panel, which is less than a printed sheet of paper at about 320 to 1. The medium usually described as having the worse dynamic range has the better one, in the room both are used in.

And at 2,000 lux — the illuminance a print is judged at in a standard booth — the display is at 24 to 1. A screen placed beside a viewing booth so a proof can be compared with a print, which is what the standard recommends, is a screen whose contrast has been reduced by a factor of forty.

The surprise: the code values survive #

The natural expectation is that the flare eats the bottom of the code range — that the darkest few dozen of 256 levels all come out the same. Measured, it barely does.

Counting the 8-bit codes whose lightness falls within one L* of the black:

Three codes lost in a dark room and three at 300 lux. The room does not eat the code space; it lifts the floor the whole scale sits on.

And the flatness has a sharp end, which is worth locating because the threshold is not where anybody would look for it.

Reconstructing the count directly — a code step of one 255th through the sRGB toe, against a one-L* window at the lifted black — reproduces the published column to within a single code on one row, and shows why it is flat and then is not.

Below the threshold, the count cannot change at all. CIELAB’s lightness has a linear segment for luminance ratios under 0.008856, where $L^*$ is simply proportional to $Y$. In that region a one-lightness-unit window is a fixed increment of luminance, and the sRGB toe’s code steps are also fixed increments of luminance — so the number of codes inside the window is a ratio of two constants and does not depend on the flare at all. Adding light moves the black up the scale and takes the window and the code steps with it, unchanged.

Above it, the window widens. Past 0.008856 the lightness is a cube root, so the same one-unit window covers progressively more luminance as the black rises, and progressively more code steps fall inside it.

For this panel the crossing is at a black of 2.67 candelas a square metre, which is 374 lux:

So the surprise in this section is a fact about the metric rather than about the display. It is not that the room happens to be gentle on the code space up to some level; it is that below a stated ratio the measuring stick is linear, and a linear stick cannot report a compression that is not happening. The panel’s dark-room contrast decides where the crossing falls — a display with a deeper black crosses later, and one whose black is already above a hundredth of its white has no flat region at all.

That also says what the count is worth as a diagnostic. Three codes lost is the answer in a dark laboratory and in an ordinary office, and it is the same answer for the same reason in both — which makes it a poor way to detect that anything has happened. The quantity that moved between those two rows is the black’s lightness, from 0.9 to 6.6, and it moved by a factor of seven while the code count sat still.

The reason is in the encoding. The sRGB transfer function has a linear toe near black, so its lowest code steps are already large in luminance — the step from code 0 to code 1 is a bigger fraction of the black level than the flare is. Adding a constant to everything shifts the curve up without compressing the spacing at the bottom, and the codes stay distinguishable from each other while the whole group moves away from black.

So the damage is not banding. It is that the darkest thing the display can show is now a mid-dark grey, and every tonal relationship in the picture is compressed against that floor.

The screen’s own surface #

The other variable is the screen, and it is the one a purchaser can choose.

A factor of nine in reflectance is a factor of six in contrast, and it is entirely a property of the surface rather than of the display electronics. A cheap panel with an excellent coating beats an excellent panel with a poor one, in any room with light in it.

There is a second, harder distinction the model here does not make. The number above is diffuse reflectance, which veils the whole screen evenly. A glossy screen also has a specular component, which does not veil — it produces an image of the lamp, in one place, which the observer moves to avoid. The two are traded against each other by the coating: a matte finish scatters the specular reflection into diffuse veiling, which removes the distracting image and lowers the contrast everywhere.

What calibration can and cannot do #

A calibrated display is one whose signal-to-luminance behaviour has been measured and corrected. Every measurement in that process is made with an instrument held against the screen, which excludes ambient light, and the correction is applied to the signal.

The flare is not in the signal. No lookup table, no matrix and no profile can subtract a luminance the panel is not emitting.

What can be done is to compensate: raise the display’s black level and adjust the tone curve so that the sum of emitted and reflected light follows the intended curve. That is what ambient-aware display modes do, and it works — at the cost of the contrast that was lost anyway, since the compensation cannot put back light the flare added.

The other thing that can be done is to state the condition. The ICC’s profile format has a viewing conditions tag, largely unused; the video standards specify a reference environment in detail — five candelas of surround for critical work, and a stated screen reflectance — and the whole grading industry observes it, which is why a film graded in a reference room and watched in a kitchen looks different in a way nobody considers a fault.

Where the light comes from #

The single number in the arithmetic is the illuminance falling on the screen, and it is not the illuminance anybody would quote for a room.

A room described as 500 lux is usually being described by the light on a horizontal work surface, which is where a lighting engineer measures it. A screen is vertical, and a vertical surface in a room lit from above receives a fraction of that — perhaps a third to a half, depending entirely on the fittings. The same room with a window behind the observer can put far more on the screen than any lighting design intended, and the same room with the window behind the screen puts almost none on it and dazzles the observer instead.

So the ambient figure in the table is a property of an arrangement of furniture, and it is the reason two people in the same office can have quite different displays. It is also why the advice given to anybody doing critical colour work — a dim room, no window in the field of view, a hood over the screen — is not fussiness: every item on that list is a term in one equation.

The apparatus this page is on #

There is a reflexive version of all this that this site cannot avoid, and it is worth stating rather than working around.

Every figure here is being rendered on a display in a room, and neither is known. The site’s hatched out-of-gamut regions mark colours no display can show; the arithmetic above says that in a bright room a great many colours the display can show are arriving veiled, with their darkest values lifted to a mid-dark grey and their contrast reduced by a factor of seven.

That does not invalidate any figure — the site’s swatches are computed from spectra and are correct instructions to a display in any case — but it does mean the reader’s room is a variable in every dark figure on this site, in the same way the reader’s display is a variable in every one. A dark swatch that reads as black on a calibrated screen in a dim room reads as grey beside a window, and no page can compensate for the difference because no page can measure it.

The site’s one available response is the same one it applies to the display: say so, and put the number in reach.

Where this model stops #

The screen is treated as a Lambertian diffuser. Real screens are not: their reflectance is a mixture of diffuse and specular with an angular distribution, and the observer’s position relative to the lamp decides which dominates. The number here is the veiling term, which is the one that applies everywhere on the screen at once.

The eye’s adaptation is not modelled here. An observer in a bright room is adapted to the room, not to the screen, which changes the appearance of everything on it — the appearance model’s account of that is a separate and roughly comparable effect, and the two compose.

Illuminance is quoted as though it were uniform, and a real room is not: light comes from windows and fittings at particular angles, so a screen may be at 200 lux and a sheet of paper on the desk beside it at 500. That is a large part of why a print and a screen compared in an ordinary room are being compared under two different conditions.

The 374-lux threshold is this panel’s. It is the ambient at which the lifted black reaches 0.008856 of the lifted white, so it moves with the panel’s own black, its peak and its coating — a display with a tenth the black level crosses at about ten times the illuminance, and one with a glossy surface crosses at a quarter of it. What does not move is that a threshold exists and that it sits in CIELAB rather than in the hardware.

And nothing here covers a reflective display. An electronic paper display has no emission at all, so the room’s light is not veiling — it is the illumination, and the arithmetic is a print’s rather than a screen’s.

The number that should be on the box #

It is worth writing down what an honest display specification would say, since every ingredient is in the arithmetic above and none of it is difficult.

A useful specification has three numbers rather than one: the peak luminance, the dark-room black, and the screen’s diffuse reflectance. From those, any contrast ratio in any room follows in one line, and a purchaser can compute the number that applies to their own desk rather than to a laboratory.

The third is the one that is almost never published, and it is the one that a purchaser cannot estimate: a matte and a glossy panel look alike and differ by a factor of nine in the quantity that decides the result. Professional reference displays sometimes publish it; consumer ones almost never do, and the contrast figure they publish instead is the one measured where the term is zero.

The generalisation #

An additive contamination is invisible where the signal is large and dominant where it is small, so it destroys ratios at one end of a range and nothing at the other.

Stated that way it is a familiar shape, and this site has met it twice already in other fields. A camera’s read noise is a fixed quantity added to every pixel, so it is irrelevant in the highlights and decides the shadows — which is why the noise argument in photography is always about the dark end. Stray light in a spectrophotometer is the same thing in an instrument: a small constant that matters only where the sample is dark, which is exactly where a dense ink is measured.

The general moral for a specification: a ratio quoted for a device is a ratio measured in an environment, and if the environment is not stated the ratio is the best one available rather than the one anybody will get. A contrast ratio measured in a dark room, a signal-to-noise ratio measured at full scale, a gamut measured with no flare — all three are honest numbers about conditions nobody uses.

The useful habit is to ask what the additive term is in the situation of use, and then to compute the ratio again.

There is a second habit in the code-count result, and it is about instruments rather than about devices. A quantity that refuses to move is not always reporting that nothing is happening; it may be reporting that the scale it is measured on is linear over the range in question, and a linear scale cannot see a shift. The right response to a flat column is to ask what the column is a ratio of — here, two luminance increments that move together — rather than to conclude that the effect is absent. The lightness of the black moved by a factor of seven across four rows in which the code count did not move at all, and only one of those two numbers was worth watching.

Who found it, and when #

Veiling glare was measured long before displays existed. The photographic literature of the early twentieth century treats it as a property of lenses and of viewing conditions, and the standard model — a constant added to every luminance, with contrast falling accordingly — is the same one used here.

For displays it entered the standards in the 1990s, when portable computers made bright-environment use ordinary. The relevant measurements now specify an ambient illuminance and a screen reflectance and report a contrast ratio for each; the useful ones report several.

The number on a consumer specification is still the dark-room one, and it has grown enormously — a modern panel may claim a million to one, which is a true statement about a laboratory and reduces to a few hundred in a room. That is not deception exactly: the dark-room number is the only one that measures the panel alone, and any other number is a measurement of a room the manufacturer chose.

The honest version, and it is occasionally published, is a curve: contrast against ambient illuminance, for a stated reflectance. Everything in this essay is that curve, computed rather than measured.

Where the ladder goes next #

This rung sits on the display is an unknown, which is where this site established how little a page knows about its apparatus, and on how bright is white, where a code value stops being a ratio and starts being a luminance.

Beside it, the proof is a different object handles what the same room does to the observer rather than to the light, and the two effects add.

Below it, a black that is not black is the print’s version of the same limit, and the comparison between the two is the essay’s most surprising number: in an ordinary room, the sheet of paper has the deeper black.