Reading the Grade

Four colour studies sit beside this one. They are full of numbers — L* 9.41, σL* 38.2, C*ab 53.3, warm balance 0.19 — and every one of those numbers is doing real work. If you cannot read them, the studies are a slideshow of pretty frames with captions you have to take on trust.

This guide exists so you never have to take them on trust again. It assumes you know nothing about colour science and are not embarrassed about it. It explains one idea at a time, shows you the picture that idea was measured from, and tells you where the idea stops working.

Everything here is drawn from the four studies themselves. No figure has been invented for the sake of a tidy example: the ten frames below are the actual record-holders — the darkest, the brightest, the most saturated, the most contrasted — out of the ninety curated frames those studies published.

Open any frame of any film and there is no colour in it. There is a grid of positions, and at each position three whole numbers between 0 and 255 — how much red, how much green, how much blue. A 1920×1038 frame is about two million positions and six million numbers.

Colour is what happens when a screen turns those numbers into light and your eye turns that light into a sensation. Everything in these studies happens in the middle of that sentence: the numbers are measured, the sensation is inferred, and the gap between them is where all the interesting mistakes live.

A useful habit: whenever a study says a frame "is green", translate it in your head to "the green numbers in this frame are consistently larger than the red and blue ones." That is all any of this ever means.

Here is the first genuinely counter-intuitive thing, and most confusion downstream starts by skipping it.

The number stored in the file is not proportional to the light that comes out of the screen. A pixel of 128 — exactly half of 255 — does not produce half the light of a pixel of 255. It produces roughly a fifth of it.

This is deliberate. Your eye is far better at telling dark tones apart than bright ones, so storage that spent its numbers evenly across the light range would waste most of them on highlights you cannot distinguish and starve the shadows you can. Instead the numbers are bent — encoded — to match your eye, and the screen bends them back.

The curve that does the bending is the transfer function. These four studies all assume a plain 2.4 power curve, written gamma24 in the configuration. Get that assumption wrong and every absolute brightness figure in every study is wrong with it.

To get one brightness number out of three colour numbers you have to decide how much each channel counts. Rec.709 — the standard behind ordinary HD video — settles it like this:

luma = 0.2126 × red + 0.7152 × green + 0.0722 × blue

Green counts for about seven-tenths, red for a fifth, blue for barely a fourteenth. That is not a preference; it is roughly how sensitive your eye is to each. It has a consequence worth holding on to: a deep blue can be vivid and nearly black at the same time. It is loud in colour and quiet in light, and the two are measured separately for exactly that reason.

Luma is still a number about the file. L* is a number about your perception.

L* is the lightness axis of CIELAB, a colour space built so that equal steps in the number are roughly equal steps in what you see. It runs 0 (black) to 100 (white), and it is deliberately not proportional to light: L* 50 — the tone you would call halfway — reflects only about 18% of the light of white.

This is why the studies quote L* rather than raw pixel values. When Thirty Rooms of Colour says the median shot of The City of Lost Children is L* 9.41, that is a claim about how dark the film looks, not about how few photons it emits. And it is a startling claim: on a scale where 100 is white, the middle shot of that film sits below ten.

Contrast in these studies is σL* — the standard deviation of L* across the pixels of a frame. It answers one question: how far from the frame's own average does a typical pixel sit?

A frame with σL* near 1 is nearly uniform — a wall, a fade, an unbroken black. A frame near 38 has both deep shadow and bright highlight in quantity. It is a measure of spread, and it is completely independent of whether a frame is dark or bright: you can have enormous contrast entirely inside the shadows.

Two σL* traps worth knowing. A single blown highlight in an otherwise flat frame raises it sharply. And it says nothing about where the light and dark are — a checkerboard and a split screen can score identically.

A histogram tells you how many pixels sit at each brightness, but "how many pixels are between 0.22 and 0.60" is not a sentence anyone wants to read twice. So the scale is cut into five named bands, and the studies quote the share of the frame in each.

Toe is near-total black — the part of the picture that has gone. Shadow is dark but still readable. Mid is where faces and most of everything lives. Light is bright surfaces. Spec is sources and reflections — the top of the range.

"78% of shots are low-key" is this measurement, aggregated. Every one of these boundaries is a decision someone wrote down, not a natural feature of light. They are useful because they are held constant across all four studies, which makes films comparable.

Once brightness is separated out, what remains is colour, and CIELAB describes it with two numbers:

• a* runs green (negative) to red (positive).
• b* runs blue (negative) to yellow (positive).

Zero on both is a perfect neutral grey. Everything else is a position on a plane, and two useful quantities fall straight out of that position:

• Chroma (C*ab) is the distance from the centre — how much colour is there, regardless of which.
• Hue is the direction — which colour, regardless of how much.

This is worth pausing on, because it is the geometry the whole vocabulary rests on. When Gold Dug Out of Black says Sisu's median chroma is 6.71 while Thirty Rooms says Lost Children's is 11.38, it is saying the typical pixel of one film sits nearly twice as far from grey as the other's.

These studies use two different words for colourfulness, and treating them as synonyms will mislead you badly.

Chroma (C*ab) is an absolute distance from grey. A near-black pixel cannot have high chroma; there is not enough light in it to be far from anything.

Saturation (HSV) is a ratio: how far apart the largest and smallest of the three channels are, relative to the largest. It has no idea how bright the pixel is. Channels of 10, 3 and 2 give a saturation of 0.8 — and that pixel is, to your eye, black.

The frame below is the proof, and it is not a contrived one — it is a real curated frame from Zatoichi. Its mean saturation is 0.75. Its mean chroma is 0.14. Its mean L* is 0.21. One number says the frame is three-quarters saturated. The other says it has essentially no colour in it at all. Look at the frame and decide which one you believe.

The rule to carry away: saturation is only meaningful once there is light to be colourful with. Where a study wants a claim about colour it can defend, it reaches for chroma.

Warm balance is the number that appears most often in the four studies and is misread most easily. It is defined like this:

of the pixels chromatic enough to have an opinion, what fraction of them lean warm?

Three things are packed into that sentence, and all three matter.

"Chromatic enough to have an opinion" means pixels below saturation 0.08 are thrown out entirely. A near-grey pixel does not vote.

"Lean warm" means the hue falls in 300–360° or 0–90° — reds, oranges, yellows. Cool is 90–300°. That boundary is a configured choice recorded in the analysis run, not a fact about light. Move it and every warm balance in every study moves.

"What fraction" means it is a count, not an intensity. A frame at warm balance 1.000 is one where every voting pixel leaned warm. It says nothing about how warm, how bright, or how much of the frame was chromatic in the first place. This is why a frame can be overwhelmingly warm and almost entirely dark — as most of The City of Lost Children is, at a median warm balance of 0.982 and a median key of L* 9.41.

A scope answers a question your eye is bad at. Your eye is superb at comparison and terrible at absolutes: it will adapt to a cast within seconds and then swear it was never there. A scope does not adapt.

The three slide-lab studies show six of them beside every frame, always in the same order and always with the same convention: muted marks are the whole film, brighter marks are the one shot you are looking at. Once you have that, every panel is answering the question "how unusual is this frame, for this film?" rather than "what is in this frame?"

One thing they are not: these are analytical displays computed from an encoded file, not calibrated, colour-managed scopes in a grading suite. They describe the numbers that reached the analysis. They cannot see a node tree, a mask, a power window or an intention.

Take the picture. Take the leftmost column of pixels. Stack them by brightness — dark at the bottom, bright at the top — into a thin vertical smear. Do that for every column, left to right, and you have a waveform.

Horizontal position is still horizontal position in the frame. That is what makes it readable: a bright window on the left of the shot puts marks high up on the left of the scope. Vertical position is brightness, and nothing else.

What to look for: a floor that never touches the bottom means the blacks are lifted. A ceiling pressed flat against the top means highlights are clipped. A shape that hugs the bottom across the whole width is a genuinely dark frame rather than a badly exposed one.

The same idea, three times: one distribution per channel, side by side on a shared scale.

The parade is the fastest way to see a cast. If the red curve sits consistently further right than green and blue, the frame is warm — not as a judgement but as an arithmetic fact about which numbers are bigger.

It is also the fastest way to see a deliberate cast in the shadows, which is a signature of graded rather than merely captured images: the channels separate at the dark end and converge higher up. That pattern is consistent with a colourist's hand. It is not proof of one.

A histogram of brightness: how many pixels sit at each level. The five zones sit behind it as bands, and three percentile marks sit on it.

P10, P50 and P90 are the brightness levels below which 10%, 50% and 90% of the picture falls. P50 is the median pixel. The distance from P10 to P90 is a robust measure of how much of the tonal range the frame actually uses — robust because a handful of blown pixels cannot move it, the way they move a maximum.

In these films the shape is usually the same: a hard wall against the left edge, a long thin tail to the right. That is what a low-key film is, drawn.

Throw away brightness entirely and plot only colour: direction from the centre is hue, distance from the centre is how much of it. Neutral grey is the exact middle, so a frame with no colour is a dot in the centre.

The six labelled targets are the primaries and secondaries at fixed angles. The dashed line is the skin-tone line — the direction most human skin falls along, regardless of complexion, because what varies between people is mostly how far along the line they sit, not the angle. Colourists use it constantly to check that faces have not drifted.

Read it as an honest rule of thumb, not a measurement. In these studies the skin figures come from a colour heuristic — a test on pixel values — and not from face detection. A wooden door, a brick wall and a bowl of soup all pass it happily.

A hue compass answers "which colours does this frame actually spend its area on?" The ring is the colour wheel; the shape inside is the amount of each hue.

The weighting is the interesting part. Each pixel's vote is multiplied by its saturation and by the square root of its value, so near-grey and near-black pixels barely count. Without that, the noise in a dark frame — where hue is meaningless and unstable — would swamp the actual colour, and every low-key film would look like it contained every hue equally.

That weighting is a deliberate analytical choice. It is the right one for these films, and it is still a choice: it makes the compass a picture of where the colour is, not of where the pixels are.

The sixth panel is the only one that is not a picture of the frame. It is a picture of the frame's place in its own film, across six measurements at once.

Each row is one measurement. The thin bar spans the middle 80% of every shot in the film; the thicker box is the middle 50%; the white line is the film's median. The orange dot is the one shot on screen.

This is the panel that turns a number into a judgement. L* 28.7 means nothing on its own. L* 28.7, sitting above the 75th percentile of a film whose median is 14.8 means this is a bright frame, for this film — and that is a sentence worth writing down.

Almost every headline figure in these studies is a median, not an average, and the choice is doing real work.

The mean adds everything up and divides. One four-second explosion in a two-hour film pulls it upward, and the figure now describes a film that does not exist.

The median is the middle value: half the shots are darker, half are brighter. A single explosion moves it by one shot's worth — which is to say, not at all.

Film measurements are heavily skewed — a great mass of ordinary shots and a long thin tail of extremes — so the two genuinely disagree. The Warm Axis reports Delicatessen at a mean L* of 18.75 and a median of 16.47: the gap is the tail, doing exactly what tails do.

A median tells you the middle and nothing about the spread, and two films can share a median while being nothing alike. So the studies quote quantiles: the 10th, 25th, 50th, 75th and 90th percentiles.

Read them as a shape. The 10–90 band is where the bulk of the film lives. The 25–75 box is its habit. When Gold Dug Out of Black reports a 10–90 band of L* 5.0–46.8 against Lost Children's 4.1–20.6, it is telling you that Sisu ranges across four times as much of the brightness scale — before you know anything about either film's median.

A wide band is not better than a narrow one. A narrow band is a film with a firm habit; a wide one is a film that changes register. Both are decisions.

This is the distinction that separates a defensible study from a confident one.

The baselines in the slide labs are populations: every photographed shot of the film, 1,252 or 1,410 or 945 of them. When a scope says "the film", it means all of them, not a sample. Claims about the film as a whole are safe here.

The thirty curated frames in each lab are not a sample. They were chosen for what they demonstrate — they are the argument, not the evidence base. Nothing about the film may be estimated from them, and the studies say so explicitly. The ten frames in this guide are a selection from that selection, chosen precisely because they are extremes.

One more level worth keeping straight: thousands of shots inside four films are not thousands of independent observations of "a Jeunet film". They are four films. The unit of observation for a claim about filmmaking is the film, and four is a very small number.

A two-hour film is around 170,000 frames. Measuring all of them would be slow, and would mostly measure the same thing repeatedly — successive frames inside one shot are nearly identical.

So the film is cut into shots by detecting the edits, and exactly one frame is taken from each: the midpoint. It is reproducible, it is cheap, and it avoids the dissolve at a shot's edges.

It is also completely blind to anything that happens inside a shot. A camera that pans from a dark corridor into sunlight is measured once, in the middle, and the study will record a single moderate brightness that no viewer ever saw. Long takes are where this bites hardest, and every one of the four studies states the limit rather than hiding it.

Detected cuts are also proposals, not facts. Fades, whip pans, flashes and credits all fool edit detection, which is why the upstream method requires a human to review the cut QC before the shots are treated as observations.

A widescreen film inside a 16:9 file has black bars above and below. Those bars are not part of the picture — they are packaging — and they are pure black in enormous quantity.

Measure them and every film gets darker, less colourful and more contrasted, in direct proportion to how wide it was shot. Sisu, at 2.39:1, would be hit hardest; a 1.85:1 film barely at all; and the comparison between them would become nonsense.

So the active picture is detected and cropped before anything is measured, and the crop is recorded with the run. All three slide-lab studies resolved to a single crop across every shot of their film — which is itself worth checking, because a film with mixed aspect ratios would need per-shot handling and a much more careful claim.

The five-colour strip under every frame is a measurement, not a mood board. The frame's pixels are clustered into five groups by colour; each swatch is one group's average, and its width is the share of the frame that group covers.

This is why the biggest swatch is so often black or near-black, and why that is the honest result rather than a boring one. A palette that gave equal width to five colours would be telling you about a designer's taste; this one tells you what the frame is mostly made of.

The aggregate palettes in The Warm Axis are the same idea one level up, weighted by shot so that a long take does not count more than a short one.

The Warm Axis is the comparative one: four Jean-Pierre Jeunet films graded by Yvan Lucas, 1991 to 2004, across the change from photochemical timing to digital grading. Its question is what survives that change. The answer it reaches is a direction rather than a palette — mean b* (the yellow axis) rises by 7.59 between eras while mean a* (the red axis) barely moves at all, and key rises by nearly ten points of L*. Warmth persists; the way it is delivered changes.

Thirty Rooms of Colour takes the darkest film of that set — The City of Lost Children, median key L* 9.41 — and asks how colour survives inside darkness. Its answer is rationing: a median warm balance of 0.982 inside a median shadow field of 81%.

Gold Dug Out of Black is the counter-example, and the most useful one for calibrating your instincts. Sisu is far more open (median L* 23.20) and decisively cool (median warm balance 0.191) — a reminder that "dark and warm" is a house style of one director, not a property of cinema.

Indigo, Straw and Blood sits between them: Zatoichi at median L* 14.83 and warm balance 0.620, low-key without being black, warm-leaning without being committed.

Put the three side by side and the instrument proves itself. Same six measurements, same baselines, same displays — and three films that could not be confused for one another.

Median chroma tells the shortest version: 11.38 for Lost Children, 6.71 for Sisu, 6.15 for Zatoichi. The first film puts nearly twice as much colour into its typical pixel as the other two — while being by far the darkest of the three. That combination is the whole finding of Thirty Rooms of Colour in one comparison, and you can only see it because the three studies were measured the same way.

This is what a shared instrument buys: not better numbers, but comparable ones.

One. The transfer function is unratified. Absolute values carry an unquantified error; relative comparisons within a film do not.

Two. The edition is not the film. A WEB encode, a Blu-ray and a theatrical print are three different objects. Sisu was measured from a 1080p WEB encode; Zatoichi from a 720p source upscaled into a 1080p timeline. Cross-film absolute comparison inherits every one of those differences.

Three. The midpoint frame cannot see inside its own shot.

Four. Curated frames are an argument, not a sample. The population is the baseline.

Five. A skin heuristic is not face detection. It gates on pixel values, and wood, brick, sand and soup all pass.

Six. Four films by one director are four observations. A pattern across them is a pattern in a very small set, and the director, the cinematographer, the production designer, the stock, the lab, the restoration and the mastering are all still standing in the room with the colourist.

None of this is here to replace looking at films. It is here to make looking at them harder to fool.

The eye adapts. It adapts to a cast in seconds, to a low key in a minute, and then it reports confidently that nothing unusual was happening. A measurement does not adapt, and a baseline does not forget what the rest of the film looked like. Between them they let you say this frame is unusual, for this film and then go back and see why.

That is the whole trade. The numbers are not the point; they are the thing that keeps your attention honest long enough for the point to show up.