Adapting to a warm room deepens the shadows' tint

A display’s shadows sit on a floor of light: the panel’s own leaking black and the room’s light reflected off the screen. A colour engine that knows the floor can render shadows for it, and one that has the floor wrong renders them for a room that is not there. A warm room tints the black, not the halo found the commonest way to have it wrong: a room lit by a 2,700 K lamp reflects that lamp’s colour off the screen, and an engine that declares the floor’s level correctly but in grey leaves every zone of a dark picture tinted towards the lamp’s orange by more than a ΔEITP unit, up to 17.7 at the darkest grey in a room putting a tenth of a candela on the screen. Declared in the room’s colour, the tint nearly vanished.

ΔEITP is a colour difference for absolute light. It has no viewer in it: two stimuli are as different as their cone signals, run through the perceptual quantiser, say they are, whatever the eye has been looking at. A viewer who has watched a screen in a warm room for an evening has been looking at the room. Chromatic adaptation discounts the colour of the light an eye is adapted to, and a viewing condition is an argument is the reason to ask an appearance model rather than a difference formula what such a viewer sees. Adapted partly to a warm white, a viewer should find warm light less warm.

The prediction was that a viewer adapted a third of the way to the room sees under half the tint ΔEITP computes, and that at two thirds the darkest greys turn the other way, towards blue, because the declared grey floor then looks cool against an adaptation that expects warmth. If that held, the colour a declaration should give the room’s light would not be the lamp’s but the share of it the viewer has not adapted to.

The tint the viewer sees grows with adaptation #

For a viewer adapted to a mixture of the display’s D65 white and the room’s 2,700 K white, read through CIECAM16 at an adapting luminance of 10 cd/m² and a dim surround, the tint in the moon picture — the CAM16-UCS colour distance between a grey on the declared grey floor and the same grey on the true warm one — grows at every step of adaptation: at its worst, 2.75 units for a viewer adapted to the display, 3.35 a third of the way to the room, 3.92 two thirds and 4.42 fully adapted. More than nineteen zones in twenty pass one unit for every viewer. The darkest grey never turns blue: it looks orange to every viewer adapted up to two thirds of the way, with its colourfulness falling from 2.6 to 1.2, and is nearly neutral to the fully adapted one. The greys that turn blue are the mid-greys, from half a candela up, whose light is the display’s own. In the bright window picture the tint shrinks with adaptation, from 0.8 to 0.5; in the subtitles it rises and falls. Declaring the floor in the room’s colour keeps every viewer’s tint under half a unit.

• The prediction is wrong in direction for the dark picture. Adaptation enlarges the visible tint in the shadows rather than discounting it.
• The shadows do not turn blue. They fade towards neutral; it is the brighter greys that look blue to a warm-adapted eye.
• The repair is unchanged. A floor declared in the room’s colour removes the tint whatever the viewer is adapted to, so the declaration does not need to know the viewer.

The three pictures part. The moon picture, nearly all dark, rises steadily from 2.75 units to 4.42. The window picture, bright enough that the light it scatters inside the viewer’s eye puts a grey veil over every zone, starts under one unit and falls. The subtitles, a dark frame with a bright line, rise to 2.3 a third of the way to the room and fall back to 1.9. ΔEITP had ordered the three the same way — the moon tinted most, the window least — and adaptation pulls them further apart: the darker the picture, the more of its floor is the room’s light, and the more the viewer’s adaptation matters to it.

Why adaptation enlarges a tint it should discount #

The tint is a difference between two stimuli, and adaptation acts on both. An appearance model’s adaptation scales each cone signal by a gain set by the white the viewer is adapted to, and adapting towards a 2,700 K white turns the short-wave gain up, since a warm white has little short-wave light, and the long-wave gain down. The warm floor and the grey floor differ most in exactly the short-wave signal, where the lamp is weakest. Turning that signal’s gain up enlarges the difference between them in the very coordinate where it lives, at the same time as it makes the warm floor itself look less warm. The census does not take the model apart into its gains and compressions; what it shows is the net, and the net in the darkest greys is growth.

At the darkest grey the growth is 2.73 units to 4.36, and it fades with brightness: by a few candelas the grey’s own light dominates the floor and the tint falls under a third of a unit for every viewer. The shadows are where the floor is a large share of what the eye receives, and the shadows are where adaptation matters. A lit room brings the units’ medians together found the room’s light, given its share in an appearance model, doing much of what the eye’s own veil does to a display’s differences; here the same share, given a colour, does something a grey room’s light cannot.

The shadows stay warm and fade #

The prediction’s second half was a turn of hue in the shadows. The darkest grey looks orange — a CIECAM16 hue of 61 degrees for a viewer adapted to the display, 68 a third of the way to the room, 72 two thirds of the way — and only for a fully adapted viewer does it lose its hue, its colourfulness then a fifth of what it was. What does turn is everything above the shadows. From half a candela up, a viewer adapted a third of the way to the room or more sees the greys as blue, hues between 230 and 265 degrees, because those greys are lit by the display’s own D65 white, and to an eye expecting 2,700 K a D65 white is cool.

So a warm-adapted viewer of a dark picture sees a grey ramp that is not grey anywhere: orange at the bottom, where the room’s light is most of the floor, blue from the middle up, where the display’s white takes over, and nearly neutral only around the crossing. The darkest grey’s orange fades from a colourfulness of 2.58 to 1.99, 1.24 and 0.50 as adaptation grows, while a mid-grey’s blue grows from nothing to several units. The tint the census measures — the difference between what the declared floor and the true floor show — is largest where the two kinds of grey meet the floor’s error, at the bottom.

A share of adaptation is a number someone must choose #

The census reads four viewers because nobody knows which one is watching. CIECAM16 has a degree of adaptation of its own — the share of the way to complete adaptation its formula assigns from the surround and the adapting luminance — and a discount nobody measured found it at 0.94 in an ordinary room and the difference from complete adaptation worth a factor of 1.7 on what an observer would report. That number answers a different question, though: how completely the viewer adapts to whatever white they are adapted to. The share here answers which white that is, and a viewer in a warm room watching a screen has two candidates in view at once.

The shares the census reads span the plausible range rather than pick from it. A viewer who has just walked in from daylight is adapted to the display’s white, or near it; one who has sat for an hour with the screen filling a small part of the field is adapted mostly to the room. A gain has a time constant put a clock on the shift between them, and the clock runs over minutes. Whatever the share, the census’s answer runs the same way in the dark picture — the tint grows — so the conclusion does not hang on choosing it. What does hang on the share is how large the growth is, from a fifth more a third of the way to the room to three fifths more fully adapted.

Two units, two kinds of number #

The tint ΔEITP gave this same zone was 17.7 units; CIECAM16 gives a viewer adapted to the display 2.7 CAM16-UCS units. The two numbers are not a disagreement about size. They are on different scales — ΔEITP’s unit was scaled so that one is about a threshold at any luminance, CAM16-UCS’s for differences between surface colours under a stated viewing condition — and a model judged in another model’s unit is the warning against reading one against the other directly. What can be compared is how each moves. ΔEITP has no way to move with the viewer, and CAM16-UCS moves by three fifths across the range of viewers in the dark picture and by more than a third the other way in the bright one.

That is what an appearance model is for, and the reason the room is the illuminant once mattered for surfaces matters here for a screen: the light a viewer is adapted to is a fact about the room, not about the stimulus, and a difference formula that takes only the stimulus has assumed a room without saying which. ΔEITP’s 17.7 is the right number for a viewer adapted to nothing in particular; the census’s four numbers are each right for one particular viewer, and the dark picture’s grows with every step that viewer takes towards the room — the one direction a difference formula cannot follow.

Declaring the room’s colour works for every viewer #

The warm-room census’s repair was to declare the floor in the room’s colour as well as its level. Declared that way, the moon picture’s worst tint stays under half a CAM16-UCS unit for a viewer at every share of adaptation. That is the useful result for a display pipeline. A floor declaration has no way of knowing what the viewer is adapted to, and the prediction would have made the right declaration depend on it — declare the share of the room’s colour the viewer has not adapted to. The census says it need not: declaring the room’s own colour removes the floor’s error for every viewer, because the error is a difference between two floors and the right floor is right for everyone.

What adaptation does change is everything else about how the display looks. A viewer adapted to a warm room sees the display’s whites and greys as cool whatever the floor, and that is a mismatch between the display’s white and the room’s that a floor declaration does not address. The units disagree in the dim found CAM16-UCS and ΔEITP parting on exactly such questions of light level; here they part on viewer state, which only one of them has.

The adapting luminance does not decide it #

An appearance model’s answer depends on the adapting luminance, and a dim room’s is not well defined: the screen, the walls and the lamp all contribute. Between 5 and 20 cd/m² the tint at the darkest grey changes by less than a fifth at every share of adaptation, and grows with adaptation at every level. The finding is not a product of the level chosen.

How the viewers were modelled #

The panel, pictures, floors and eye veil are the warm-room census’s: a 1,000 cd/m² local-dimming LCD of 32 by 18 zones, the moon, window and subtitle pictures at the middle diffuser spread, each zone’s true floor the panel’s native black at its backlight plus a 2,700 K room’s reflected light of a tenth of a candela plus the eye’s veil from the picture’s bright content, and the declared floor the census’s two floors in grey. A grey of luminance Y on a floor is carried to XYZ relative to the display’s white and read through CIECAM16 with the white the viewer is adapted to set as the mixture (1 − s) D65 + s 2,700 K, for s of nought, a third, two thirds and one, an adapting luminance of 10 cd/m² (5 and 20 as a check), a background of a fifth and a dim surround. The tint is the distance between the two greys in CAM16-UCS’s a′ and b′, worst over grey levels from 0.005 cd/m² to the white, and the greys’ own appearance the CIECAM16 hue and colourfulness of the grey on the true floor.

What the census does not reach #

One appearance model. CIECAM16’s adaptation is a von Kries scaling in its own cone space with a degree of adaptation fixed by the surround; a model that adapted the short-wave signal differently, or less completely at low luminance, would give a different size of growth.

A uniform adaptation state. A viewer’s adaptation in a room is not one white: the screen fills part of the field and the room the rest, and the shadows on a screen are seen against both.

No observers. Whether viewers in warm rooms report shadows warmer or cooler than intended, and by how much, is the experiment the model stands in for.

Still open: whether viewers in a warm room see the shadows’ tint grow #

The census predicts something an experiment could see directly: the same dark grey on a screen, rendered for a grey floor, looks more wrong to a viewer adapted to a warm room than to one adapted to the screen.

The experiment is a matching task: viewers adapted for twenty minutes to a 6,500 K and to a 2,700 K surround each adjust a dark grey on a screen in a warm-lit room until it looks neutral, at three luminances from 0.005 to 0.5 cd/m², and the adjustment is read as a colour difference from the grey a grey-declaring engine would have shown. The prediction worth recording is that the warm-adapted viewers’ adjustments are larger at the darkest level, by about half again, and smaller than the cool-adapted viewers’ at half a candela, where the display’s own white starts to look blue to them. If the warm-adapted viewers’ adjustments are smaller at every level, real adaptation discounts the room’s colour in the shadows where CIECAM16’s does not, and the model’s short-wave gain is wrong for dark stimuli.

Discounting a light is not discounting a difference #

The habit here is about what adaptation acts on.

The prediction reasoned that adapting to a warm light discounts warm light, so a tint made of warm light should shrink. That is true of the warm light’s own appearance: the warm floor looks less warm to an adapted eye. But the tint was never the warm floor’s appearance. It was the difference between the warm floor and the grey one, and adaptation acts on both, scaling the very signal in which they differ. Discounting a colour and shrinking a difference are different operations, and an adapted eye does the first while, here, doing the opposite of the second.

The failure mode is to reason about a difference as if it were one of its terms. The test is to compute the appearance of both stimuli under the adaptation, rather than of the one that seems to carry the effect.