The Architecture of White

Colourists talk about highlight roll-off as though it were a property a film has, the way a film has an aspect ratio. This study set out to find that property in 52,527 shots from 43 films, using every measurement the corpus already carried and re-measuring 9,856 frames pixel by pixel to get at the things it did not.

The property is not there. Almost everything the phrase is normally used to mean — how much clips, how far the top of the picture reaches, how much of the frame is bright — turns out to belong to the shot rather than to the film. Ninety-five per cent of the variation in how much a frame clips is variation between shots of the same film.

What does belong to the film is colour: which white the picture is heading for, and how many distinguishable steps survive inside it. Those are the measurements that hold still across a running order, and they are also the ones that separate one bright object from another when the display has almost no room left to separate them by level.

That is the report's central claim. Highlight rendering in the finished image is not a curve. It is an arrangement — of colour, of size, of local structure and of what the picture puts next to what — inside the last fifteen L* the format has to give.

Corpus. 43 films, 52,527 analysis-eligible shots

Re-measured. 9,856 frames, pixel by pixel

Complete pixel coverage. 7 films, 8,537 frames

Measured resolution floor. ΔE00 0.61 between neighbouring near-white codes

1. corpus — A film's clipping is not a property of the film: 95% of the variation in it is between shots of the same film.
2. corpus — What is a property of the film is the colour of its whites — the strongest film-level signature measured, by a distance.
3. corpus — White is a minority event. Only 15.9% of shots reach L* 90 at their ninety-ninth percentile, and where clipping happens at all it takes 0.08% of the frame.
4. census — Once you match frames on how much of the picture is bright, clipping predicts more highlight texture, not less. The folklore has the sign backwards. (overturns the folklore)
5. census — Flat white almost never means empty: 0.9% of frames have whites without structure at either scale measured.
6. census — 84% of frames that carry any highlight carry two or more perceptually distinct whites, a median ΔE00 11.9 apart, of which a median 38% is not lightness.
7. census — A compact bright source does not outrank an extended bright field by level. The median difference between them is under one L*.
8. corpus — The largest axis of variation between films is how much bright material they contain — exposure, not roll-off. Colour is the second.
9. corpus — Sicario never reaches L* 90 in any of its 1,205 shots and never clips. O Brother, Where Art Thou? reaches it in 43% of its shots. Same cinematographer.
10. corpus + census — High-frequency energy in the delivered master does not order by decade or by capture medium, so no film-versus-digital claim about highlights survives measurement on masters. (overturns the folklore)
11. census — Thirty curated frames do not stand in for a film: the median curated statistic sits 19 percentile points away from that film's own distribution. (about the method)

Seven separable properties are hiding inside one phrase, and in this corpus they do not travel together.

The phrase is doing at least seven jobs. It is used for the rate at which luminance compresses near the top of the range; for whether anything reaches the ceiling; for how well two bright objects stay apart; for whether a white surface keeps its texture; for what happens to colour as pixels get brighter; for the order in which bright materials sit; and for whether compressed highlights still look like they are emitting light. Those are seven different things, they are measured differently, and in this corpus they do not travel together.

The looser claim underneath the phrase is that a film has a roll-off the way it has a look — a stable property you could hear described and then recognise. That is testable. If it is true, shots inside one film should agree with each other about their highlight behaviour more than shots drawn from different films do. If it is false, the phrase is describing something that changes from shot to shot, and attributing it to the film is a category error.

This report tests it, and then asks what is left when the answer comes back mostly negative. Something is clearly there — the whites of Delicatessen do not look like the whites of Sicario, and no one has trouble telling them apart. The question is what that something is measured in.

That hypothesis, which this study set out to test, survives the corpus better than the curve account does — but not in the form it is usually stated. The differences that get preserved are mostly not differences in level, because by the time two objects are both in the highlight band the display has almost no level left to give them. They are differences in hue, in chroma, in size, in local structure, and in what the picture puts around them.

The rest of this report is the evidence for that, the several places where the evidence contradicts the folklore, and a proposal for saying it more precisely.

Three tiers of evidence that can answer different questions, and a measured cost for the one that looks like a sample and is not.

Three tiers of evidence, kept apart throughout, because they can answer different questions and the report never lets one stand in for another.

The corpus tier is every analysis-eligible shot of every film — 52,527 rows — carrying the measurements the extraction pipeline already made. It is complete and unselected, and its resolution is limited to the columns that exist. The census tier is 8,537 frames re-measured pixel by pixel: every photographed shot of the seven films whose Resolve exports were on local disk. It is complete for those films and only those films. The curated tier is 1,319 frames re-measured the same way — the thirty-odd frames each published slide lab selected. It is the only pixel access forty of the films have, and it is not a sample of anything.

Four films are in both the census and the curated tiers, which makes the cost of the curated tier measurable rather than assumed. For each measurement, the curated median was located inside that film's own full distribution: fifty would mean the selection is typical.

It is not typical. The median displacement is nineteen percentile points, and only 29% of the film-and-measurement pairs land within ten points of typical. The City of Lost Children's selected frames sit at the 89th percentile of the film's own ninety-ninth percentile and the 87th of its own clipping. Kwaidan's sit at the 86th percentile of near-white tonal diversity while Sisu's sit at the 14th. The displacement is large and its direction is not even consistent between films, which rules out correcting for it.

So the curated tier is used here for what a selected frame can honestly support — what is happening inside that frame — and never for what a film typically does. Every film-level distribution in this report comes from the corpus tier or the census tier.

Two limits are structural and are measured rather than glossed.

The first is eight bits. Every frame in the corpus is delivered as an eight-bit encode, and between L* 92 and the ceiling a gamma-2.4 encode has about twenty code values. Any statement about tonal diversity inside the whites is bounded by that. Every measured frame therefore carries the number of distinct luma codes its near-white pixels actually use, and a frame whose whites are built from too few codes is reported as unanswerable rather than answered.

The second is what the picture is. A rendered frame is the output of a whole imaging chain — exposure, capture, grade, display transform, master, encode — and no single frame separates them. The report therefore describes the rendered upper-luminance distribution and refuses to call it a shoulder or a response curve. Where the word compression appears, it means values crowded together in the delivered file, not an operation anybody performed.

The resolution floor comes free from the data. For every frame, the distinct quantised colours present in its near-white set were collected and the median nearest-neighbour CIEDE2000 between them recorded. That median is 0.61 across the corpus, and the 95th percentile is 0.79. Two whites in this report have to clear ΔE00 3 — about five times the floor — before they are counted as two whites.

• Films: 43 — 1964–2025
• Shots aggregated: 52,527 — analysis-eligible rows
• Frames re-measured: 9,856 — at native resolution
• Complete pixel coverage: 7 films — 8,537 frames
• Quantisation floor: ΔE00 0.61 — measured, median over frames
• Separation threshold: ΔE00 3.0 — ≈5× the floor

Defined here and nowhere else. Every later use links back to this list rather than re-explaining itself in a caption.

Highlights are a minority event. Only one shot in six reaches L* 90, and where clipping happens at all it takes a thousandth of the frame.

Before asking how cinema handles the approach to display white, it is worth establishing how often cinema goes anywhere near it. The answer is: rarely, and when it does, in very small quantities.

Across 51,453 SDR shots, 15.9% have a ninety-ninth percentile at or above L* 90, and 7.4% reach L* 95. Clipping — any channel at the top code — touches more than a tenth of a per cent of the frame in 15.7% of shots. Where clipping is present at all, the median clipped area is 0.08% of the picture: on a 1920-wide frame, roughly sixteen hundred pixels out of two million.

• Shots reaching L* 90: 15.9% — at the 99th percentile
• Shots reaching L* 95: 7.4%
• Shots clipping >0.1%: 15.7% — of the frame area
• Shots clipping >1%: 6.0%
• Median clip where present: 0.08% — of the frame

This matters for how the rest of the report should be read. A measurement of highlight behaviour is, in most shots, a measurement of a very small and very unrepresentative part of the picture — often a few hundred pixels of a lamp, a window edge, a rim on a shoulder. Film-level averages of such a quantity are averages over a highly intermittent event, which is one reason they are so unstable.

It also puts the folklore in proportion. Arguments about clipping concern a phenomenon that, when it occurs, occupies about a thousandth of the frame. Whatever people are responding to when they call a film's highlights harsh or beautiful, it is unlikely to be that thousandth on its own.

Films order themselves neatly by where they put the top of the picture — and on most of them, the spread within the film is wider than the gap to the next film.

Three percentiles are available for every shot of every film, which makes one comb the whole corpus can draw: where a film puts the tenth, fiftieth, ninetieth and ninety-ninth percentile of a typical shot, and — more informative — how much those placements move.

The films order themselves cleanly enough. First Cow, Suspiria, Enter the Void and Sicario sit at the bottom; O Brother, Where Art Thou?, No Country for Old Men and Another Round at the top. But the pale bar behind each row is the middle half of that film's own shots, and on most films it is wider than the gap between neighbouring films.

The Banshees of Inisherin is the extreme case and a useful one, because it is not an incoherent film. Its interquartile range for the ninety-ninth percentile spans 55 L*, from a dim interior at 35 to a blown coastal sky at 90. Both belong to the film; neither is the film's highlight behaviour. Point Blank runs from 37 to 80 across the same measure. Films with narrow ranges exist — Hero's is 20 L* wide, Sicario's 25 — but they are the minority, and narrowness is itself a strategy rather than the default.

On the seven films with complete pixel coverage the comb can be drawn finely, from the ninetieth percentile of the median frame up to its maximum. What that picture shows is worth stating carefully, because it is easy to misread as a tone curve.

Every film's rungs crowd at the top, and the crowding says almost nothing about grading. A rendered frame's quantile function near q = 1 is dominated by how few pixels are up there, and the ceiling is fixed by the format. Reading a shoulder out of one frame's upper percentiles is reading the format and the subject, not the transform.

Aggregating over a whole film helps, because content averages out and the ceiling does not. The survival curve — the share of every pixel of a film at or above each L* — is the honest version of the picture people have in mind when they say roll-off. It still is not a tone curve. It is a joint statement about what the film photographed and how the result was finished, and the two cannot be separated by looking at masters.

White is not one destination. Eighty-four per cent of frames that carry any highlight carry two or more distinct whites, and a median 38% of what separates them is not lightness.

White is not one destination. The near-white pixels of a frame were clustered in CIELAB and the clusters kept only where every retained pair was at least ΔE00 3 apart — five times the measured quantisation floor — and each held at least 3% of the near-white pixels.

Of the 4,199 re-measured frames that carry any highlight population at all, 84.4% carry two or more distinct whites. The median separation between a frame's two most distant whites is ΔE00 11.9. The median difference in lightness between them is 8.6 L*. And a median 38% of the total separation is not lightness at all.

When a frame keeps several whites apart, what is doing the keeping — level, or colour?

The two ends of the corpus. Both frames are doing the same job — keeping several bright things apart — and only one of them is doing it with luminance.

Measured on: census.

This is where the curve account of highlight rendering breaks down as a description of practice. In Sisu's overcast sky the whites are ordered by level and a tone curve is a complete account of what happened to them. In the Delicatessen frame the whites are at nearly the same level and forty ΔE apart, and no operation on a luminance curve can produce that or destroy it. Both are ordinary cinema.

The corpus-wide figure — 38% of the widest separation carried by something other than lightness, at the median — sits between those two poles and is the honest general answer. Kwaidan's median is 73%: its whites are separated almost entirely by colour. Sisu's is zero.

The report set out expecting to find a luminance hierarchy among bright materials — skin sheen below white cotton below smoke below sky below the practical. It is not there.

Compact bright regions surrounded by much darker picture (which behave like sources and speculars) were separated from extended bright regions covering at least 2% of the frame (which behave like skies, windows and lit walls). In the 593 frames that contain both, the median difference in level between them is −0.96 L*: the compact source is very slightly darker than the field, and 66% of the pairs sit within three L* of each other. The difference in chroma is 0.58.

The plateau is real, and it is what makes the rest of this report necessary. Once two bright things are both in the highlight band, the display has essentially no level left to distinguish them. Whatever keeps a practical lamp reading as a lamp and a window reading as a window, it is not that one of them is brighter.

Is there anything inside a white that looks blank on a normal display?

The exploded frame is not a picture of the film; it is a picture of what the file still holds where a normal display cannot show it. The stretch is chosen per frame, so brightness between exploded pictures is not comparable — structure is.

Measured on: curated.

Match frames on how much of the picture is bright, and clipping predicts more highlight texture, not less. The folklore has the sign backwards.

The folklore says that clipping destroys highlight texture and that soft compression preserves it. Neither half survives measurement.

Pooled across the 8,537 census frames, the rank correlation between clipped area and the flat fraction of the whites is −0.21, which is already the wrong sign for the folklore. But the pooled number is confounded: frames with a lot of bright area have flatter whites almost by construction, and those are also the frames most likely to clip. The correlation between bright area and flat white is +0.47.

Matched on how much of the picture is bright, the relationship is negative in every band and reaches −0.50. Frames that clip have more textured whites than frames of the same brightness that do not.

The mechanism is not mysterious once stated. Clipping usually means a source or a specular is in shot, and sources and speculars arrive with structure: edges, flare, the objects they are reflecting off. A large flat white is a large evenly lit surface — a sky, a snowfield, a wall — and evenly lit surfaces are what actually produce featureless whites, whether or not they ever reach the ceiling. The flattest whites in the corpus, on a muzzle flash and gunsmoke in Django Unchained, do not clip a single pixel.

Can two frames clip the same amount and look nothing alike?

Two frames with almost the same clipped area and opposite highlight texture, both of which read as good photography. Clipped percentage on its own predicts nothing about appearance, which is why it should stop being quoted as though it did.

Measured on: curated.

There is a second confusion inside the word texture. The measure used above is local: the standard deviation of L* inside a seven-pixel window. By that measure a pale sky is flat, and the eye disagrees, because the sky is climbing several L* from one side of the frame to the other.

A related trap sits in the measure itself. Local texture computed over a highlight set of a few hundred pixels is arithmetically large and perceptually meaningless, because every one of those pixels is an edge: the corpus maximum on that measure is a frame of Kwaidan whose whites cover four hundredths of one per cent of the picture. Every extreme quoted here is therefore taken over frames whose whites cover at least 2% of the frame.

Measuring both scales separates them. Twenty-one per cent of frames with highlights are locally flat and still sweep more than three L* across the bright region. Only 0.9% are flat at both scales — genuinely featureless white. The Banshees sky above is one of the locally flat ones, and its exploded version is a smooth vertical gradient with a bright band along the horizon: no local texture, a great deal of shape.

So flat white almost never means empty. It means the structure is at a scale a small window cannot see, and colourists who reach for a curve to “bring back texture” in such a shot are working on the wrong quantity.

High-frequency energy in the delivered master was measured in flat midtone regions and, separately, inside the highlights. The ratio is the useful number: highlight grain runs at 46% of midtone grain at the median. Whatever else the top of the range does, it systematically carries less fine structure than the middle — which is a real effect, is consistent across films, and is the one place where the folklore's intuition about compression is straightforwardly right.

A film has a highlight colour and a highlight grain. It does not have a highlight roll-off: 95% of the variation in clipping is between shots of the same film.

The test is a variance decomposition. For each measurement, the share of its total variance that belongs to the film rather than to the shot is estimated as an intraclass correlation, with a film-level bootstrap interval — resampling shots would treat tens of thousands of correlated rows as independent and give an interval far too tight.

The ordering is the finding, and it is the same in both tiers, computed on different measurements from different data.

Everything the phrase highlight roll-off normally denotes sits at the bottom of both charts. How much of a frame is bright, how much clips, how many clipped regions there are, how far the brightest pixel reaches: these are shot properties. Attributing them to a film is attributing the weather to the country.

Everything at the top of both charts is about colour and about grain inside the whites — which white the film is heading for, how far from neutral it sits, how wide the spread of near-white hues is, and how many distinguishable steps the encode is spending up there. Those hold still across a running order.

This is not a null result dressed up. It is a specific and checkable claim: a film has a highlight colour and a highlight grain, and it does not have a highlight roll-off.

The largest axis of variation between films is how much bright material they contain — exposure, not rendering. Colour is the second. Nothing resembling a soft-versus-hard axis appears at all.

Clustering films on their raw highlight statistics recovers exposure, not strategy: the first component is dominated by how much bright material a film contains, which is a fact about what was photographed and how it was exposed.

To ask the question the prompt actually poses — given the same amount of bright material, what does this film do with it? — every shot carrying real bright material was placed in a decile of how much bright material it carries, each highlight descriptor was standardised inside that decile, and a film's profile is the mean of its shots' standardised scores. A film scores high on an axis when it treats a given quantity of bright material differently from the corpus average treatment of the same quantity.

Held low

Her, Kwaidan, Sicario, Suspiria. Given bright material, these films keep it in a narrow band well below the ceiling and almost never let it clip. Sicario's bright-zone spread is 2.7 L*, the tightest in the corpus.

Extended

Drive, Super Deluxe, O Brother, Where Art Thou?, No Country for Old Men, Gangubai Kathiawadi, Hero, Skyfall and others. Bright material is allowed to spread, reach and clip. Note that Deakins appears in this family three times and in Held low once.

Coloured whites

Amélie, Delicatessen, The City of Lost Children, The Toxic Avenger, Gangubai Kathiawadi. Whites carry strong chroma and, in several cases, more chroma than the midtones beneath them. This is the family the ICC analysis says is real.

Moderate

The largest group, eighteen films, sitting near the corpus average on both axes. Their differences from each other are mostly on the third component and are not large. A film being unremarkable on these axes does not make its highlights unremarkable to look at.

The families are honest but they are weak. The first component is about exposure discipline more than about rendering, the second is the colour signature the variance decomposition already identified, and the residual is small. There is no cluster corresponding to soft shoulder versus hard shoulder, because there is no measurement in this study that separates films that way once exposure is matched.

That is a negative result worth stating plainly rather than burying in a scatterplot: the corpus does not contain the families the folklore expects.

Three strategies, and they are not a spectrum from good to bad: films that neutralise their highlights, films that hold chroma to the ceiling, and films whose bright zone is more saturated than their midtones.

If colour is the film-level signature, the useful picture is where each film's colour goes as its pixels get brighter. Every film draws one line in the a*/b* plane, from the mean chromaticity of its midtone zone to that of its bright zone, with the head at the bright end.

The strategies the prompt anticipated are all present, and they are not a spectrum from good to bad.

Neutralising

Hero (C* 8.8 → 2.0), Oldboy (11.2 → 1.1), Bugonia, O Brother. Bright colours converge on neutral. In Hero and Oldboy the hue also swings by 45–73°, which is what channel behaviour near the ceiling looks like when one channel arrives first.

Retaining

Delicatessen (38.9 → 46.3), Amélie, A Very Long Engagement, The City of Lost Children. Chroma survives to within a few L* of the ceiling. All four are Jeunet–Khondji or Jeunet–Delbonnel pictures with a shared finishing lineage, which is a confound and also the point.

Colouring

The Toxic Avenger (15.1 → 26.8), Blade Runner 2049, Let the Right One In, Gangubai Kathiawadi. The bright zone is more saturated than the midtones, because the bright things are coloured sources rather than white surfaces catching light.

The third family is the one that breaks the standard framing. A film whose highlights are more saturated than its midtones is not preserving chroma against a neutralising transform; it is photographing coloured light. Chroma retention measured as a ratio cannot tell those apart, which is why the report reports the vector rather than the ratio: the direction says which, and the ratio does not.

The neutralising cases deserve the same care. Hue turning by tens of degrees as chroma collapses is the signature of channel-wise behaviour near the ceiling — one channel reaching its limit while the others still have room, so the surviving hue is the hue of what is left. It is visible on a parade as the channels converging at different heights, and it is not the same event as a gradual desaturation that keeps its hue.

High-frequency energy in a delivered master does not order by decade or by capture medium, so no film-versus-digital claim about highlights survives measurement on masters.

The most repeated claim in cinematography about this subject is that film has beautiful highlight roll-off and digital clips harshly. Capture medium is not recorded anywhere in this corpus, and assigning it from memory would be asserting rather than measuring. What can be measured is the high-frequency energy the delivered master carries in flat midtone areas — the quantity photochemical origination puts there, along with, indistinguishably, grain emulation, denoise, restoration and the delivery encode.

Measured, it does not order by decade and it does not order by anything a capture-medium argument would predict.

Sisu (2022) measures 0.035, the cleanest master in the corpus. The Conformist (1970) measures 0.061. Drive My Car (2021) and Portrait of a Lady on Fire (2019) sit at 0.09 and 0.10, and The Assassination of Jesse James (2007) at 0.49. Whatever this quantity is tracking, it is not the decade and it is not the origination.

The conclusion is not that film and digital highlights are identical. It is that a corpus of finished masters cannot answer the question, because the finishing is doing more to the measurable result than the capture is. Every film here has been through a digital intermediate, a display transform, a master and an eight-bit encode, and several have been through a restoration as well. Any residual capture-medium effect is downstream of all of that.

This should also be read back onto the folklore. The claim that film rolls off beautifully is, in practice, almost always a claim about a print, a scan or a master — an artefact of finishing — rather than about a negative's exposure response. The corpus is consistent with the finishing being where the effect lives.

Highlight behaviour is a shot property, but not an unstructured one — and three of the patterns it makes are different phenomena that look alike on a timeline.

If highlight behaviour is a shot property, it should still be a structured one: films are not sequences of unrelated photographs. Plotting the band from the ninetieth to the ninety-ninth percentile along each film's running order, filled with the measured colour of its bright zone, shows what kind of structure it is.

Three kinds of structure are visible and they are not the same phenomenon.

Sicario holds a constant, narrow band for two hours: the highlight behaviour is a rule the film keeps. Hero moves in blocks — the film's famous colour sections shift the whole band's hue and its height together, so the highlight treatment is one of the things the sections are made of. Amélie and Super Deluxe oscillate rapidly between interiors and exteriors with no larger arc, which is what the low intraclass correlations look like when drawn out in time: the variation is real, it is large, and it is organised by scene rather than by the film.

There is a fourth pattern worth naming because it looks like structure and is not. Enter the Void's band is dominated by isolated events — its strobes are excluded from the corpus as ineligible, but the remaining light sources still produce spikes that no amount of smoothing turns into a level. A film can have highlight drama without having highlight behaviour.

One cinematographer holds both ends of the corpus. The frame with the widest white in it is from 1967, and the most differentiated near-white is a night exterior that is 0.13% bright.

Chosen for being informative rather than famous: two statistical extremes that turn out to be about the same cinematographer, the frame with the widest white in the corpus, and the frame with the most tonal levels.

Sicario has 1,205 analysis-eligible shots. Not one of them has a ninety-ninth percentile at or above L* 90. Not one of them clips more than a tenth of a per cent of the frame. Its bright-zone spread, 2.7 L*, is the tightest in the corpus, and on the exposure-matched profile it is the most extreme film measured on that axis.

O Brother, Where Art Thou? reaches L* 90 in 43% of its shots and clips in 21% of them. It sits at the opposite end of the same corpus.

Both are Roger Deakins. Across his seven films here the share of shots reaching L* 90 runs 0.0%, 5.6%, 7.4%, 22.1%, 23.1%, 37.5%, 42.8% — the full range of the corpus, inside one filmography. The chroma of the bright zone, by contrast, stays between 12.7 and 22.1 on all seven, which is the same result the variance decomposition gave, seen from the other side.

The frame below is the case in one picture: a noon desert border crossing, one of the brightest scenes in the film, whose brightest pixel reaches L* 91 and whose ninety-ninth percentile is 75. There is not enough material above L* 85 in it to measure a highlight population at all.

What does a film look like that never goes near white?

The film routinely praised for the quality of its light is the film in this corpus that goes least near white. Whatever that praise is responding to, it is not highlight rendering in the sense the phrase usually carries.

Measured on: curated.

The widest hue spread among near-white pixels in the corpus is in a 1967 photochemical picture: 97.8° of circular spread. Looking at the frame explains it immediately. Lee Marvin's grey hair, a pale hazy sky, a light blue shirt and the sunlit water behind him are four different whites at four different hues, all within a few L* of each other, and the picture keeps them apart entirely by colour and by what surrounds them.

The exploded version shows what is living in those few L*: cloud banding across the whole sky, the city lights of Los Angeles along the horizon, glints on the reservoir, and the individual strands of the hair.

What separates four whites sitting at almost the same level?

A frame that a curve-based account of highlight rendering has no vocabulary for. Nothing here is separated by level; everything is separated by hue, texture and context.

Measured on: curated.

The frame with the most effective tonal levels inside its whites is not a bright frame. It is a night exterior in Okja with 0.13% of the picture above L* 85: dozens of small practicals, street lights, vehicle beacons and lit windows, each a separate small white with its own colour temperature, adding up to 41 effective levels in the near-white band.

This is the counterpart to the Django Unchained result. The flattest white in the corpus is a large soft smoke plume with no clipping at all; the most differentiated is a scatter of tiny sources in an otherwise dark frame. Neither has anything to do with a shoulder, and both are ordinary, competent, unremarkable cinema.

Where does the most differentiated near-white in the corpus actually live?

The two ends of near-white tonal diversity. Neither is a failure and neither is a triumph of roll-off; they are two different things to photograph.

Measured on: curated.

Three conclusions survived the frames most likely to break them, one had to be weakened, and one was abandoned.

Each of the report's conclusions was tested against the frames most likely to break it. Three survived, one had to be weakened, and one was abandoned.

Held: clipping does not predict harshness

The most clipped frame in the corpus — Hero at 80.8% — is a deliberate white ground with cloud modelling, a violet gradient and grain still living inside it. The least clipped flat white — Django at 0.000% — is the most featureless. Searching specifically for beautiful heavily-clipped frames and ugly clean ones produced plenty of both.

Held: white is not one destination

The counterexample would be a corpus where distinct white populations are always separated by level. Sisu is that film — colour share zero — and it is one film out of seven measured this way. The corpus median is 38%.

Weakened: films have a highlight colour

The colour signature is real and it is the strongest measured, but 0.42 is not 0.9. Individual shots regularly leave their film's colour entirely: a night exterior in a warm-graded film has cold whites like anyone else's. The claim is about central tendency, not about every frame.

Abandoned: a soft-versus-hard family

The report set out expecting to find films that compress early and gradually against films that compress late and steeply. Once exposure is matched, no such axis appears in the data, and the axis that does appear — extension — is about how far a film lets bright material go, not about the shape of a transition.

One more caution, on a result that looks stronger than it is. The intraclass correlations on the census tier are computed across seven films, and their bootstrap intervals are wide: nw_codes, for example, is 0.46 with an interval from 0.09 to 0.60. The ordering is stable and it agrees with the 41-film corpus tier on the axes the two have in common, which is why the report leans on it — but a reader who wants the pixel-level result at 41 films will have to wait for the stills volume to be mounted.

The quantities that carry information about a highlight are how much of the frame is bright, how many distinct whites it holds, and how much of their separation is colour. Clipped percentage carries none.

The report cannot answer this with a perceptual experiment; no observer study was run, and a claim that a metric predicts perceived quality would need one. What it can do is say which quantities carry information and which do not, and be honest about the difference.

Carries no information

• Clipped percentage, taken alone — it predicts neither texture nor appearance, and its sign flips once bright area is matched.
• The crowding of upper percentiles, which is a property of the ceiling and the subject.
• Any single roll-off score, because the axes it would average are close to orthogonal.

Carries information

• How much of the frame is bright — the strongest single predictor of anything measured here.
• How many distinct whites the frame holds and how far apart they are.
• How much of that separation is colour rather than level.
• Structure at both scales: local texture and large-scale sweep, measured separately.

Belongs to the film

• The chroma of the bright zone (ICC 0.42).
• The hue spread among near-white pixels.
• The number of distinct code values spent in the whites.
• Nothing about clipping, reach or occupancy.

The proposition the report set out to test — that good highlights preserve a deliberately constructed hierarchy of luminance, colour, texture and semantic importance — is half supported and half refuted by the corpus.

The hierarchy is real: frames overwhelmingly carry several distinct whites and keep them apart. But it is not a hierarchy of luminance. Compact sources and extended fields sit within three L* of each other two thirds of the time, and the median difference in level between them is under one L*. Whatever ordering exists up there is carried by colour, size, context and local structure — the things a luminance-first account of roll-off has no words for.

The most defensible restatement is narrower and more useful: cinema compresses a large range of scene brightness into a narrow band at the top of the display, and then does its work inside that band with everything except level.

Four consequences at the desk, and one rule about how these numbers may be quoted.

None of what follows reconstructs anyone's node tree. A finished frame does not reveal the operations that made it, and the report does not pretend otherwise. These are the practical consequences of the measurements, stated as things to look at and things to stop trusting.

Stop reading the clip meter as a quality signal

The percentage of clipped pixels predicts nothing about how the whites look, and once frames are matched on bright area it points the other way. Use it to find where clipping is, then look at the picture.

• On a waveform, the useful question is not whether the trace touches the top but how much of the trace is within a few IRE of it and what that material is.

Grade the separation, not the shoulder

If two bright things need to read as different things, level is the axis with the least room. In the corpus, colour carries a median 38% of the separation between a frame's most distant whites, and in some films almost all of it.

• A parade shows this directly: whites that converge to one point are one white, whatever the waveform says.
• A vectorscope trace that collapses to the centre above a certain luma is a film choosing neutralisation, which is a decision and not a fault.

Check texture at two scales

A pale sky can be perfectly flat inside a seven-pixel window and still climb several L* across the frame. Twenty-one per cent of frames with highlights are like this and only 0.9% are flat at both scales.

• Before reaching for a curve to bring back texture, establish which scale is missing. A local-contrast tool and an exposure ramp fix different problems.

Expect the whites to be the film's colour

The one thing that reliably persists across a film is where its whites sit relative to neutral and how wide their hue spread is. It is worth deciding deliberately and worth checking shot to shot, because it is the thing an audience will read as continuity.

• This is a gamut and hue-versus-luma question, not a curve question.
• It also travels: it is the measurement that survived both tiers of this study.

One consequence for how these numbers should be quoted. Because highlight behaviour is a shot property, a film-level average of it is an average over an intermittent event and will mislead. A film summary that says the ninety-ninth percentile is L* 71 is describing a distribution running from 35 to 90. Report the dispersion or do not report the measure.

Seven properties, each measurable on a rendered frame, and two words to retire.

The phrase is not wrong so much as overloaded. Seven separable properties are hiding inside it, and every one of them can be measured on a rendered frame without any claim about a transfer curve. Splitting them makes it possible to say which one a picture is short of.

Two words the report recommends retiring in technical contexts. Soft, unless the speaker says what is soft — the local texture, the boundary of the clipped region, the rate of chroma loss, or the large-scale gradient, which are four different things that a single frame can have in any combination. And roll-off itself, as a property attributed to a film, because the corpus says films do not have one.

As a description of a shot, roll-off is fine. As a description of a look, it should be replaced by the two properties that actually persist: the colour a film takes its whites to, and how many distinguishable steps it spends getting there.

Every figure quoted in the prose above is declared in the report's source and re-verified against the measurement files at build time; a number that drifts fails the build rather than surviving in the text. The charts read the same measurement files directly, so no chart can disagree with the sentence beside it without the build noticing.

The largest limitation is the pixel tier. Complete frame-by-frame coverage exists for seven films because their Resolve exports happen to be on local disk; for the other thirty-six, pixel access is limited to curated frames whose displacement from their films was measured in section 02 and found to be large. Mounting the stills volume and re-running tools/measure_frames.py would extend every census-tier result to the whole corpus, and the intraclass correlations in figure 11 are the numbers most likely to move.

Two further limitations are structural. No perceptual experiment was run, so nothing here establishes what observers see; the report distinguishes throughout between what is measurable and what is visible, and declines the second. And every frame is an eight-bit display-referred master, so the study describes deliveries rather than negatives, sensors or grades.

• Corpus: output_*/shots.parquet, 43 films, 52,527 analysis-eligible shots. Eligibility is analysis_eligible where the extraction carries it and the published lab's picture range otherwise; the rule used is recorded per film in data/films.json.
• Census tier: the 1920×1080 Resolve exports under stills/, cropped to each film's measured active picture, for Amélie, A Very Long Engagement, Delicatessen, Kwaidan, Sisu, The City of Lost Children and Zatoichi.
• Curated tier: the lossless frame assets published by the slide labs, byte-identical to their audited sources.
• Colour: encoded RGB decoded with each film's recorded transfer (gamma 2.4, Rec. 709 or PQ anchored at 203 cd/m²) and primaries, then CIELAB under D65. The two PQ masters are excluded from every comparison made on a lightness scale.
• Reproduce: tools/corpus_manifest.py --out white_report/data (shared with the black report, at the repository root), then white_report/tools/corpus_scalars.py, measure_frames.py, white_census.py, texture_material.py, analysis.py and render_frames.py, then white_report/build.py.