A lit room brings the units' medians together

The units part by hue, not by light level built pairs of colours that differ in lightness alone and in chroma alone, each exactly one ΔE₀₀, and read them in the two colour differences that take a light level — ΔEITP and CAM16-UCS — on displays from 1.5 to 10,000 cd/m². Both units say that as a display brightens, chroma differences gain on lightness differences: ΔEITP’s median balance falls 11 per cent and CAM16-UCS’s 23. They part by hue. The appearance model moves every base colour’s balance by the same factor; ΔEITP moves the violets, the reds and the cyan-blues the other way, and at nine of the twelve base colours the two go in opposite directions.

Every one of those readings had the viewer adapted to the display alone. CAM16-UCS takes its light level as an adapting luminance — what the eye is adjusted to — and it was set to a fifth of the display’s white throughout, as though the reader were sitting in a dark room with nothing else to look at. That is the convention and it is not a reader. The essay ended on the question: the viewer is adapted to something between the display and the room, so what does the room do?

What the room changes and what it does not #

Letting the room supply part of the viewer’s adaptation flattens the appearance model’s response to the display’s level at every step, and with the room supplying all of it the response vanishes. In a 20 cd/m² room supplying two thirds of the adaptation, CAM16-UCS’s median balance falls by the same 11 per cent as ΔEITP’s. Their disagreement about direction survives untouched: the model still moves every base colour towards chroma, ΔEITP still moves nine of the twelve towards lightness.

• With the display alone, CAM16-UCS’s median falls 23 per cent from a 1.5 to a 10,000 candela white. With a 20 cd/m² room supplying half the adaptation, 13; a quarter from the display, 10; none, 0.
• ΔEITP’s median falls 11 per cent in every case, because its light level is the stimulus’s own and no room enters it.
• The medians meet when the room supplies 81 per cent of the adaptation in a 5 cd/m² room, 68 in a 20 cd/m² office and 40 in a 60 cd/m² showroom.
• In a lit room the model’s per-colour moves shrink from about ×0.77 to about ×0.88 and none crosses to the other side. The units still move opposite ways at nine of twelve base colours.
• The best colour for an experiment costs more in a lit room: the dark violet goes from two observers to four, because the room moves the model towards ΔEITP’s median and away from ΔEITP where they disagreed about direction.

What the viewer is adapted to #

The viewer here takes all, three quarters, half, a quarter or none of the adaptation from the display, and the rest from a room whose adapting luminance is 20 cd/m². An office lit to about 300 lux, with ordinary surfaces, sits near that.

The mechanism is in this figure, not in the colour model. With the display alone, the adapting luminance tracks the display across four decades, from 0.3 to 2,000 cd/m². With half of it from the room, it starts at 10 — the dim display adds almost nothing to the room — and ends at 1,010, a range of two decades rather than four. With a quarter from the display, the dim end is held at 15 by the room and the bright end reaches 515. The room acts as a floor under the adaptation, and a floor under the adaptation is a floor under everything the model does with it.

That matters because of where the appearance model’s response to level lives. The units part by hue, not by light level found that CAM16-UCS’s reading of a pure lightness difference does not move with the level at all, and that its reading of a pure chroma difference grows by 30 per cent — the model’s colourfulness is scaled by a factor that rises with the adapting luminance, and its lightness is a ratio in which that factor cancels. Everything the model says about light level passes through the adapting luminance, so compressing the adapting luminance’s range compresses everything.

The model flattens #

The figure is the appearance model’s balance — its reading of a pure lightness pair over its reading of a pure chroma pair, median over twelve base colours — against the display’s white, for the same five shares, with ΔEITP’s beside it.

The appearance model’s curve flattens with every step of room. Adapted to the display alone, its balance slides from 1.13 to 0.87 across the range, a fall of 23 per cent. With the display supplying three quarters of the adaptation the fall is 15 per cent; half, 13; a quarter, 10. Adapted to the room alone, the model’s balance is 0.98 at every display level — the room’s value, whatever the display does.

ΔEITP’s curve does not move, because it takes the display’s light as the stimulus’s own absolute luminance and no room enters it. Its balance falls from 0.91 to 0.81, 11 per cent, in every one of these rooms.

So the question the earlier essays asked — which unit is right about how a display’s level moves the balance between lightness and chroma differences — has a second question inside it that neither unit asks: what is the reader adapted to? ΔEITP answers it by not needing to; its light level is the stimulus’s. CAM16-UCS needs an answer and has been given the darkest-room one.

How much room it takes to agree #

The appearance model’s fall shrinks from 23 per cent towards nought as the room takes over, and ΔEITP’s sits at 11. Somewhere they meet.

In a 20 cd/m² room the medians agree when the room supplies 68 per cent of the adaptation. In a dim 5 cd/m² room it has to supply 81 per cent, because a dim room is a low floor and the display pulls the adaptation up past it. In a 60 cd/m² showroom, 40 per cent is enough.

Those numbers turn an abstract parameter into something a session can control. A viewer looking at a display that fills a quarter of the visual field, in an office, is adapted largely to the room; a viewer at a reference monitor in a dim suite, with the display filling the field, is adapted largely to the display. The earlier comparisons described the second viewer. The first is the usual one, and for that viewer the two units’ medians agree about how much a brighter display favours chroma differences — which is a finding about the median, and nothing more.

A tolerance has no light level found the same distinction on surfaces: a colour difference that takes the stimulus’s luminance and one that takes the room’s are answering different questions, and on a display the two arguments are not even the same physical quantity. What this adds is how far apart the answers are in an ordinary room — for the median, not at all.

The disagreement that survives #

The median agreement leaves the base colours exactly where they were in one unit and moves them in the other.

The room pulls every one of the appearance model’s moves towards no change, from about ×0.77 to about ×0.88, and it does so uniformly — the model still moves every colour by nearly the same factor, because the room compresses the one quantity through which the model moves anything. None of the model’s moves crosses the line. In a lit room, as in the dark, CAM16-UCS says every pure lightness pair loses a little on every pure chroma pair as the display brightens.

ΔEITP’s moves are untouched: ×0.68 to ×0.85 at the yellow-greens, ×1.04 to ×1.16 at the reds and cyan-blues, ×1.17 to ×1.30 at the violets. So the units still move in opposite directions at nine of the twelve base colours. In a lit room the median agreement is exact and the per-colour disagreement is as large as it was — which makes the disagreement by hue the whole of what separates the two units, rather than the part of it that a median concealed.

What it does to the experiment #

A forced choice between the two units’ predictions is settled fastest where their predictions are furthest apart, and the room moves the predictions.

At the violets, the room makes the experiment harder. The dark violet needs two observers in the dark and four in a lit room; the mid and pale violets three and five, three and six. The reds and cyan-blues go from four to six observers in the dark to eight to twenty in a lit room. In every one of these colours ΔEITP moves towards lightness and the model towards chroma, and the room, by pulling the model’s move towards no change, pulls it towards ΔEITP’s.

At the yellow-greens it goes both ways. There the two units already agreed in direction, so the room can move the model past ΔEITP’s move or towards it: the dark yellow-green goes from 31 observers to 8, the mid one from 323 to 50, and the pale one from 35 to over a thousand.

The design consequence is plain. An experiment meant to separate the two units should be run with the viewer adapted to the display — a dim room, the display filling the field — at the violets. That is not a nuisance to be designed around; it is where the appearance model makes its strongest claim about light level, and a lit room is where it makes its weakest. An experiment run in an office would find the two units hard to tell apart at every colour except the violets, and at the violets it would need twice the observers.

Why the dark room was the default #

The convention of adapting a model to the display alone is not an oversight. It comes from the reference viewing environments that display standards specify — a dim surround, a display filling much of the field, a grading suite rather than an office — and from the history of the models themselves, which were fitted to patches viewed in controlled surrounds. The surround is three rows of a table is where the model’s own treatment of the surround was laid out, and it offers three surrounds, not a continuum between a display and a room.

The dark room is the setting in which a model’s light-level behaviour is largest, which is exactly why a standard picks it: it is the condition under which the display’s own encoding is most responsible for what the reader sees. How bright is white followed the move from relative to absolute display encoding, and that move was made for reference viewing. An absolute encoding says what a display emits; it cannot say what the reader in front of it is adapted to, and a colour difference that needs the second quantity inherits whatever the standard assumed.

The practical difference is visible in the numbers. In a grading suite, the appearance model says a brighter display favours chroma differences twice as strongly as ΔEITP does; in an office, the two agree. A tolerance written for HDR content in CAM16-UCS terms is therefore a tolerance for one viewing condition, and applied to another it over-weights the effect of display brightness on chroma by a factor that depends on how bright the room is. ΔEITP’s tolerance does not carry that dependence, and whether that is a virtue depends on whether readers in lit rooms really do see what ΔEITP says — which is the question the violet experiment would answer for the median and the hue disagreement alike.

Brighter looks more colourful is the effect underneath all of this: the appearance model’s colourfulness rises with the adapting luminance, and that is the whole of its response to light level. A lit room raises the adapting luminance at the dim end of the range, so the dim display is already seen with the colourfulness boost the dark room would only have given the bright one. A patch is not a scene is the standing caution about taking any of these numbers to a real image, where the adaptation is to the image as well as to the room.

How the readings were made #

The twenty-four pure pairs are those of the earlier essay: twelve base colours in CIELAB, three lightnesses by four hues at a chroma of 40, each with a pair stepped in L* alone and a pair stepped in C* alone to exactly one ΔE₀₀. ΔEITP reads them at the display’s absolute luminance through the perceptual quantiser, exactly as before. CAM16-UCS reads them relative to a D65 white with an average surround and a background of 20 per cent, and with its adapting luminance set to a share of a fifth of the display’s white plus the remaining share of the room’s adapting luminance. The balance is the median over base colours of each lightness pair’s reading over its chroma pair’s; a base colour’s move is its own balance at the brightest display over its balance at the dimmest.

Observer counts use the same stated scatter as before, a quarter in the logarithm of the ratio, over the full range of display whites.

What this leaves out #

The mixture is linear, and the share is a free parameter. What fraction of a viewer’s adaptation comes from a display depends on how much of the visual field it fills, how bright it is against the surround, and how long the viewer has been looking at it, and no single number describes a session. The shares here span the range from a reference suite to an office, and the finding that matters — that the model’s median response shrinks continuously and its per-colour sign does not change — holds at every share tried.

A room’s light does more than set an adapting luminance. It reflects off the screen and lifts the display’s blacks, which the shadows the unit counts are the ones a room removes priced for ΔEITP. That veil acts on the stimulus, which both units would see; the adaptation acts on the viewer, which only one unit models. The two corrections are separate and both belong in a full account of a display in a room.

And CAM16’s surround parameter is held at average. A viewer in a dim suite is in a dim surround, which the model treats as a further parameter; setting it would move the dark-room readings slightly and not the lit-room ones, and would not change which colours separate the units.

Still open: whether the hue disagreement is ΔEITP’s quantiser #

Everything the room did, it did to the appearance model, because that is where the room enters. What it could not touch is the disagreement by hue, which is ΔEITP’s: it moves the violets, reds and cyan-blues towards lightness and the yellow-greens towards chroma, and the appearance model moves them all alike.

The candidate mechanism is the perceptual quantiser acting on three cone-like signals that sit at very different levels for a saturated colour. A violet’s three signals are far apart, a yellow-green’s close together, and the quantiser’s slope changes with level, so a brighter display stretches a violet’s signals unequally. If that is right, the hue dependence should grow with a base colour’s chroma and vanish at a neutral, and it should follow how unequal the colour’s three signals are rather than its hue angle as such. The computation is the pure pairs at several chromas, with each base colour’s move plotted against the spread of its three signals in the quantiser’s own coordinates. If the dependence follows the spread, the disagreement between the units is a claim of the quantiser’s, and the violet experiment tests that claim directly.

The argument a unit leaves out #

The habit is about parameters that one of two models takes and the other does not.

ΔEITP and CAM16-UCS were compared as though they answered one question with two answers. They answer it with different inputs: one needs to know what the reader is adapted to and the other does not. Holding that input at its most extreme value — the darkest room — made the appearance model’s response to light level as large as it could be, and the comparison then found the models disagreeing about how much a brighter display favours chroma. Letting the input take a realistic value made that disagreement disappear, and left standing the one that no choice of adaptation could touch.

The failure mode is to compare two models at a setting of a parameter only one of them has, without saying the setting was a choice. The move is to sweep the parameter and see which of the disagreements it moves. The ones it moves are about the parameter. The one it cannot move is about the models.