The Shape of a Palette #4

A harmony template is a set of hue anchors at fixed relative angles. Fitting one means finding the rotation that best explains where a shot's chromatic weight actually sits. This fourth edition starts with 55,852 shot rows from 42 films. Their photographed picture ranges and eligibility flags leave 48,847; the duration, light, chroma and tonal-structure gates leave 19,325; and stage two decodes and rescores 2,052 shortlisted frames on a continuous hue distribution.

1 of the 8 exemplars changed: a frame the previous edition had never seen now fits one of the templates better than the one it chose. What is new is not the corpus. It is that a numerical leader is no longer automatically a picture in the report. Every retained frame was inspected at full resolution and again in the rendered layouts, and 8 high-ranking candidates were set aside with their reasons kept in the payload. The report shows 15 specimens across eight sections, from one to three each, and says in every case why that number and not another.

The difference matters most where the two answers disagree. The mathematical square-tetradic leader of this corpus is a Django Unchained action insert; it is set aside, and the frame the section shows instead is labelled a limit rather than an exemplar. An honest report says which frame won the arithmetic and which frame it chose to print.

Each specimen is shown the same way: the picture, its authored reading, the fit it clears and the margin it clears it by, where its colour actually sits on the wheel with the fitted template drawn over it, and — for the frame that leads each section — the same frame as a vectorscope.

Edition 3 ended at 35 films and 16,051 eligible candidates. This one adds 7 films — Okja, The Toxic Avenger, Parasite, Django Unchained, The Assassination of Jesse James by the Coward Robert Ford, Her, and First Cow — which bring 8,250 more photographed shots and 3,274 more eligible candidates.

The corpus-wide shape barely notices. The largest movement in any share is monochromatic at +0.9 percentage points, and one hue family or two adjacent ones still explain 97% of eligible shots. A finding that survives a corpus growing by 20% is a stronger finding than one measured once.

Not every new film earns a profile: First Cow has 9 eligible shots, below the thirty-shot floor, so its departure is withheld rather than reported from an unstable denominator.

Two stages, because the two instruments cost differently. Stage one fits each shot's five-colour CIELAB palette. It is cheap and coarse, so it only shortlists. Stage two decodes 2,052 stills and scores them on a continuous saturation-weighted hue distribution measured from the active picture at 512 pixels, in 72 bins of 5°. Every winner comes from stage two.

The photographed picture is the first gate. The source runs contain 55,852 shot rows. Per-film picture ranges remove distributor idents, titles and credits from 38 of 42 films; the 4 older ungated films enter whole and remain named in the payload. Upstream analysis_eligible flags also remove detected burnt-in subtitles and the collapsed strobe passages of Enter the Void. That leaves 48,847 photographed shots before the ordinary duration, light and chroma candidate rules.

A field of colour is not a photograph of a scene. A flat screen fits a one-leg template perfectly by construction, so this edition also requires σL* ≥ 4.0 and p90−p10 ≥ 8.0. The rule sets aside 48 otherwise-eligible flat frames. After every gate, 19,325 shots can supply candidates.

A frame must beat every rival template on its own pixels. Templates nest, so raw fit alone would hand too much material to monochromatic. The exemplar and per-film profile both use the winning margin. 8 further frames were set aside by hand, with their reason preserved in the frozen payload. An analysis source is deliberately outside this corpus: sampled on a clock rather than cut into shots — it has no shot durations, and the observational unit of this study is the shot. A ranked frame is a candidate, not a picture. The shortlist carried into stage two rose to 360 candidates per template with the corpus; holding it where the previous edition left it would quietly have handed a larger share of the result to the coarse instrument. Every retained frame was then inspected at full resolution and again in the rendered layouts, and each section keeps between 1 and 3 specimens according to what it has to show rather than to fill a row. 15 frames are printed here out of 2,052 decoded.

One hue family, varied only in lightness and chroma. One anchor: every chromatic pixel sits in a single hue family.

The best-fitting frame for this template in 19,325 eligible shots is A Very Long Engagement shot 0366, at 01:31:42:08, held for 5.8 seconds. It fits: 1.000 against the template, and the nearest rival reading (analogous) reaches only 0.000.

Amber without an opponent. Curtains, paper, timber and skin remain inside one amber family; the room separates itself through lightness and texture rather than a second hue. Kept as the purest fit, stage two's 1st ranked candidate for this template, clear of its nearest rival by +1.000.

Its hue spread is 4°, its mean chroma C* 18.4, its mean key L* 14.5, and 32% of its pixels carry enough colour to vote on hue at all. This section keeps 2 reviewed specimens — purest fit and forceful portrait — because that is what the corpus supports for this shape, not because a layout wanted a fixed number.

Forceful portrait. Warm skin, auburn hair and the green background compress into a narrow yellow-to-green run, giving the close-up much more chromatic force than the numerical winner.

It fits the template at 0.778 against 1.000 for the frame above, clearing its nearest rival — analogous at 0.766 — by +0.012. Its own photography is a different proposition: mean chroma C* 34.1 against 18.4, mean key L* 25.2 against 14.5, and a visual force of 0.501 against 0.176. It is stage two's 434th ranked candidate for this template.

Neighbouring hues, no opposition. 3 anchors, 32°, 32° apart, each with a ±22° window.

The best-fitting frame for this template in 19,325 eligible shots is A Very Long Engagement shot 0036, at 01:05:25:23, held for 1.5 seconds. It fits: 0.943 against the template, and the nearest rival reading (monochromatic) reaches only 0.476.

Mud, wool and wet timber. The soldier, rain-darkened uniform and trench sit across neighbouring olive, green and cyan-grey families with no opposing accent. Kept as the purest fit, stage two's 1st ranked candidate for this template, clear of its nearest rival by +0.466.

Its hue spread is 32°, its mean chroma C* 10.5, its mean key L* 19.3, and 11% of its pixels carry enough colour to vote on hue at all. This section keeps 3 reviewed specimens — purest fit, forceful colour block, and lighting contrast — because that is what the corpus supports for this shape, not because a layout wanted a fixed number.

Forceful colour block. The cardigan's green and the cafe's yellow-orange light occupy a broad but connected arc, showing analogous colour as bold production design rather than camouflage.

It fits the template at 0.842 against 0.943 for the frame above, clearing its nearest rival — monochromatic at 0.614 — by +0.229. Its own photography is a different proposition: mean chroma C* 43.3 against 10.5, mean key L* 36.1 against 19.3, and a visual force of 0.514 against 0.033. It is stage two's 40th ranked candidate for this template.

Lighting contrast. Magenta light on the balcony glass turns through red into the amber interior, a high-chroma neighbouring-hue run made almost entirely by practical-looking scene light.

It fits the template at 0.913 against 0.943 for the frame above, clearing its nearest rival — monochromatic at 0.562 — by +0.351. Its own photography is a different proposition: mean chroma C* 31.4 against 10.5, mean key L* 22.1 against 19.3, and a visual force of 0.333 against 0.033. It is stage two's 11th ranked candidate for this template.

Two hues directly opposite. 2 anchors, 180° apart, each with a ±24° window.

The best-fitting frame for this template in 19,325 eligible shots is Drive shot 0717, at 02:17:12:22, held for 1.8 seconds. It fits: 0.991 against the template, and the nearest rival reading (monochromatic) reaches only 0.502.

Warm face, cyan night. Warm skin occupies one side of the axis while the cyan background and rim light occupy the other, with almost equal measured weight on both legs. Kept as the purest fit, stage two's 1st ranked candidate for this template, clear of its nearest rival by +0.489.

Its hue spread is 118°, its mean chroma C* 10.3, its mean key L* 26.7, and 16% of its pixels carry enough colour to vote on hue at all. This section keeps 2 reviewed specimens — purest fit and landscape scale — because that is what the corpus supports for this shape, not because a layout wanted a fixed number.

Landscape scale. The wide frame divides cleanly between a blue sky and ochre-green land, turning a near-perfect two-leg fit into an immediately readable landscape.

It fits the template at 0.970 against 0.991 for the frame above, clearing its nearest rival — monochromatic at 0.529 — by +0.441. Its own photography is a different proposition: mean chroma C* 34.9 against 10.3, mean key L* 51.8 against 26.7, and a visual force of 0.428 against 0.049. It is stage two's 7th ranked candidate for this template.

One hue against the two neighbours of its opposite. 3 anchors, 150°, 60° apart, each with a ±22° window.

The best-fitting frame for this template in 19,325 eligible shots is Moonlight shot 0075, at 01:13:48:16, held for 4.4 seconds. It fits: 0.932 against the template, and the nearest rival reading (complementary) reaches only 0.517.

Blue jerseys across grass and skin. Deep blue clothing is opposed by two warmer families in the green-yellow field and the players' skin, leaving all three template legs visibly occupied. Kept as the purest fit, stage two's 1st ranked candidate for this template, clear of its nearest rival by +0.415.

Its hue spread is 102°, its mean chroma C* 18.8, its mean key L* 38.5, and 22% of its pixels carry enough colour to vote on hue at all. This section keeps 2 reviewed specimens — purest fit and depth and colour blocks — because that is what the corpus supports for this shape, not because a layout wanted a fixed number.

Depth and colour blocks. A red tail-light and cyan vehicle split the sharp foreground while warm walls sit behind the soft figure, carrying the three-way relation through focus and depth.

It fits the template at 0.757 against 0.932 for the frame above, clearing its nearest rival — complementary at 0.588 — by +0.169. Its own photography is a different proposition: mean chroma C* 21.4 against 18.8, mean key L* 36.5 against 38.5, and a visual force of 0.167 against 0.121. It is stage two's 13th ranked candidate for this template.

Three hues evenly spaced around the wheel. 3 anchors, 120°, 120° apart, each with a ±24° window.

The best-fitting frame for this template in 19,325 eligible shots is Oldboy shot 0722, at 02:33:48:18, held for 6.5 seconds. It fits: 0.628 against the template, and the nearest rival reading (monochromatic) reaches only 0.376.

Teal dress, violet wall, warm skin. Teal cloth, violet decor and the smaller warm faces and red buttons populate three separated hue regions without any one disappearing. Kept as the strongest complete triad, stage two's 1st ranked candidate for this template, clear of its nearest rival by +0.252.

Its hue spread is 104°, its mean chroma C* 13.5, its mean key L* 21.8, and 16% of its pixels carry enough colour to vote on hue at all. This section keeps 2 reviewed specimens — strongest complete triad and geometric variation — because that is what the corpus supports for this shape, not because a layout wanted a fixed number.

Geometric variation. The overhead composition assigns blue clothing, burgundy cloth and yellow wood to distinct structural zones, producing a less balanced but still winning triad.

It fits the template at 0.545 against 0.628 for the frame above, clearing its nearest rival — split-complementary at 0.436 — by +0.108. Its own photography is a different proposition: mean chroma C* 13.1 against 13.5, mean key L* 26.0 against 21.8, and a visual force of 0.067 against 0.064. It is stage two's 3rd ranked candidate for this template.

Four hues at right angles. 4 anchors, 90°, 90°, 90° apart, each with a ±22° window.

The frame this section leads with is In the Mood for Love shot 0264, at 01:46:24:03, held for 14.7 seconds. It is stage two's 2nd ranked candidate here, and the first one the visual review kept. It does not fit. This is the closest the corpus comes, and even here the template scores 0.364 while a monochromatic reading scores 0.458. The frame is better explained as something simpler.

Four hints, not four legs. Green curtains, red cloth, a violet-blue foreground and pale lamp suggest four regions, but two remain too slight for the square to beat its simpler monochromatic reading. Kept as the best naturalistic limit, stage two's 2nd ranked candidate for this template, beaten on its own frame by 0.094.

Its hue spread is 74°, its mean chroma C* 11.7, its mean key L* 14.0, and 14% of its pixels carry enough colour to vote on hue at all. This section keeps 1 reviewed specimen — best naturalistic limit — because that is what the corpus supports for this shape, not because a layout wanted a fixed number. The frame that leads this template numerically is Django Unchained shot 357, and it is not shown here: an optically extreme action insert — it is the mathematical square-tetradic leader, but not a clean cinematographic specimen.

Two complementary pairs, unevenly spaced. 4 anchors, 60°, 120°, 60° apart, each with a ±22° window.

The best-fitting frame for this template in 19,325 eligible shots is Super Deluxe shot 0596, at 02:11:46:05, held for 3.8 seconds. It fits: 0.518 against the template, and the nearest rival reading (split-complementary) reaches only 0.392.

Red cloth in a cyan car. Red clothing and warm skin sit against cyan upholstery and a yellow-green exterior, the corpus's only rectangular specimen that beats every simpler rival. Kept as the only clean winner, stage two's 1st ranked candidate for this template, clear of its nearest rival by +0.127.

Its hue spread is 91°, its mean chroma C* 22.4, its mean key L* 40.4, and 28% of its pixels carry enough colour to vote on hue at all. This section keeps 1 reviewed specimen — only clean winner — because that is what the corpus supports for this shape, not because a layout wanted a fixed number.

No hue survives the chromatic threshold. Nothing here clears the chromatic floor, so the frame has no hue to be harmonious with.

The best-fitting frame for this template in 19,325 eligible shots is Roma shot 0145, at 01:52:31:15, held for 42.7 seconds. There is no template to fit. Only 0.00% of the frame carries enough colour to be placed on the wheel at all, which is what makes it the corpus's clearest achromatic frame.

Rain, asphalt, silver air. Wet road, dark trees and pale sky distribute contrast across a wide exterior while the saturation-weighted instrument finds no surviving hue. Kept as the tonal landscape, stage two's 1st ranked candidate for this template.

Its hue spread is 180°, its mean chroma C* 0.0, its mean key L* 60.1, and 0.00% of its pixels carry enough colour to vote on hue at all. This section keeps 2 reviewed specimens — tonal landscape and high-contrast interior — because that is what the corpus supports for this shape, not because a layout wanted a fixed number.

High-contrast interior. A hard white flare behind the dark suit turns the hearing room into layered silhouettes and faces, showing achromatic structure at a very different scale from the landscape.

There is no template to fit. 0.00% of it carries enough colour to have a hue at all, against 0.00% for the frame above. Its own photography is a different proposition: mean chroma C* 3.4 against 0.0, mean key L* 53.1 against 60.1, and a visual force of 0.000 against 0.000. It is stage two's 92nd ranked candidate for this template.

Square tetradic remains a visible absence. The frame this report prints for it scores 0.364 against the square template and 0.458 as a monochromatic reading, so it is shown as the corpus's naturalistic limit and labelled a loss. The frame that does lead the square numerically comes from Django Unchained and is set aside for the reason recorded beside it: it is an optically extreme action insert, and it loses to a simpler reading on its own pixels anyway. Neither route produces an exemplar.

Rectangular tetradic is the opposite case and the rarest clean success in the study. Its frame scores 0.518 against its own template and beats the nearest split-complementary reading at 0.392. One four-leg shape occurs; the other does not.

That pair is the finding. Growing a corpus can turn an absence into an occurrence, but only where the chosen frame beats every simpler explanation of the same pixels — and no amount of growth entitles a category to an illustration.

Ranking is arithmetic; being a specimen is a judgement, and the judgement is recorded here rather than applied silently upstream. Each of these frames scored well enough to be shown and was declined for a stated reason — most often because the colour is carried by an effect, a graphic or archival material rather than by the film's own photography.

• Amélie shot 60 — editorial review. Set aside: carries burnt-in main titles — part of the measured colour is typography, not photography.
• Hero shot 674 — monochromatic · fit 0.9697 · margin +0.267. Set aside: an effects-led wire-action insert — the colour is legible, but the frame is organised around stunt spectacle rather than ordinary photographed action.
• Hero shot 507 — complementary · fit 0.9856 · margin +0.473. Set aside: an effects-led arrow-storm insert — a mathematically excellent opposition whose spectacle would overstate the role of cinematography.
• Suspiria shot 178 — split-complementary · fit 0.5207 · margin -0.226. Set aside: a practical-effects gore insert — the colour relationship is carried by the effect rather than by an ordinary scene.
• Django Unchained shot 357 — square tetradic · fit 0.2890 · margin -0.061. Set aside: an optically extreme action insert — it is the mathematical square-tetradic leader, but not a clean cinematographic specimen.
• Amélie shot 237 — rectangular tetradic · fit 0.4742 · margin -0.177. Set aside: a diegetic television graphic — most of the qualifying colour belongs to the screen image, not the room's photographed light.
• Another Round shot 402 — achromatic. Set aside: archival news footage inside the film rather than the film's own cinematography.
• Sisu shot 958 — achromatic. Set aside: a motion-blurred transitional action frame that is not legible enough to teach the category.

An omission a reader can see is a finding. One they cannot is a hole.

Monochromatic leads almost everywhere, so the useful comparison is how much material escapes it. A film's departure is the share of eligible shots assigned to anything wider than one hue family.

The upstream package reports 40 film rows with enough eligible observations. Super Deluxe leads at 54%; A Very Long Engagement sits at 0.10%. That describes these measured editions under the same gates. It does not identify who caused the pattern, and it does not turn shots nested inside one film into independent films.

97% of eligible shots are still explained by one hue family or two adjacent ones. Triadic and both tetradics together account for 0.10%. Seven new films, 3,274 more eligible candidates, and the shape of the answer did not move.

The additions changed the search without forcing themselves into it. They supply the rejected numerical square leader, reach the reviewed triadic comparisons without winning them, and alter every per-film and corpus profile on this page. Being in the mathematics is not the same as being entitled to an illustration, and this edition is the first that can show the difference.

The per-film profile stays descriptive. Super Deluxe has the largest measured departure from one hue family at 54%; A Very Long Engagement sits at 0.10%. Those are profiles of these editions under these gates, not signatures of a director, cinematographer or colourist — and a film with too few eligible shots to be stable has no row at all rather than an unstable one.

The specimen count is the other thing worth carrying away. Three analogous frames are here because environmental colour, wardrobe colour and coloured scene light are genuinely different cases; one rectangular-tetradic frame is here because the corpus has exactly one. A report that printed three of each would be saying something about its layout rather than about its corpus.