strongly warm-led
Film-balanced mean share of shots whose chromatic pixels are more than 80% warm.
Colour field study · 5,288 shots · 1991–2004
The warm
axis
A shot-by-shot study of Yvan Lucas’s colour work across four Jean-Pierre Jeunet films asks what survives when photochemical timing gives way to digital grading. The answer is not a fixed palette. It is a direction.
Four aggregate palettes
Eight colours per film · matte bars excluded
The finding in one sentence
A warm chromatic grammar survives the medium change; digital grading expands its range.
In all four films, roughly four shots in five are strongly warm-led. The digital pair is nevertheless brighter, more yellow and more contrasty, with almost no neutral pixel field left. Continuity and transformation are both present—and neither is well described by calling every film simply “sepia”.
L* in the digital pair
The film-balanced mean key rises from 15.20 to 25.15—an expansion, not a small trim.
a* across the transition
The red–green centre barely moves, even as mean b* rises by 7.59 toward yellow.
Finding 01 · Direction
Warmth is the shared rule, not a digital novelty
Share of shots by chromatic warm/cool balance
Read it: all four films make cool-led shots exceptional—only 5–7% of each corpus. A Very Long Engagement is the least uniformly warm, yet 79% of its shots still clear the strong-warm threshold.
What “signature” can mean here
Look for a stable relationship among colours, not a repeated LUT.
The cinematographers differ and the medium changes. A Lucas-associated signature therefore need not reproduce the same mean values. A more defensible test is relational: whether warm and cool are assigned similar jobs, whether faces remain integrated into the dominant field, and whether excursions away from the baseline behave as deliberate punctuation.
01
1991 · Photochemical
Delicatessen
1,116 shots
18.75 mean L*
0.62 mean S
The first film establishes warmth as atmosphere.
Delicatessen is dark, but not as dark as its successor. Its key resides in brown-black, ochre and tarnished green; red and yellow are allowed to model flesh, rooms and exterior air at once. The result is a world with very little colour-neutral distance between character and set. Coolness is most legible when it is technologically mediated.
Amber becomes atmosphere
Ruined façades dissolve into mustard-brown fog. Warmth is the air holding the image together.
Flesh inside the same field
Face, hair and background occupy neighbouring ochres; value and texture do the separating.
Blue is assigned to mediation
The coldest selection arrives through a degraded electronic image: cyan as interruption.
The same amber, opened up
The key rises dramatically without leaving the yellow-brown axis.
Darkness studyFour ways the film gives black dramatic work
Recognition begins with the eyes
Almost every pixel stays below L* 20. Tiny highlights make darkness an act of looking.
A keyhole turns black into point of view
A tiny opening carries the scene; black becomes enclosure, obstruction and access.
Amber remains alive below the key
Low luminance does not become neutral: brown-red chroma remains embedded in shadow.
The tunnel is organised around one glow
A small opening supplies direction and depth rather than broad illumination.
02
1995 · Photochemical
The City of Lost Children
1,252 shots
11.65 mean L*
0.55 mean S
The signature survives by being compressed into darkness.
This is the outlier in key and neutrality: 78% of shots fall below L* 15, while 58% contain more than a quarter neutral pixels. Yet the remaining chromatic pixels still lean warm in 84% of shots. The palette is not less designed; it is allocated more sparingly, often as small fires, red objects and green machinery suspended in black.
Colour survives at the edge of black
Small red and green accents carry a frame almost swallowed by darkness.
Fire becomes a local colour engine
Orange flame and skin punch through the low-key green-brown world.
Green replaces neutral underwater
The submerged machinery is cool yet insistently chromatic—uncanny, never clinical.
A face breaks the low-key rule
Cream highlights stay yellow-side, preserving the logic at a radically higher key.
Darkness studyHere black becomes environment, duration and threat
Figures become inhabitants of black
The dark determines the figures’ scale, vulnerability and relation to the world.
A fragment of light activates the void
At mean L* 3.2, a tiny streak makes black suspenseful negative space.
Green-black still retains atmosphere
Fog and structure remain legible through colour separation inside the shadow band.
A child is modelled by one local source
One warm source makes attention, innocence and danger share the same construction.
Finding 02 · The analogue pair is not one look
The largest early change is from dark amber abundance to colour rationed inside black.
Delicatessen and Lost Children share a direction, not a histogram. Their mean keys differ by 7.10 L*, and the second film has more than twice the high-neutral share. Treating “photochemical” as a single visual condition would erase the strongest within-era distinction.
Shot-key distributionsMean CIELAB L* per shot
Saturation distributionsMean HSV saturation per shot
03
2001 · Digital
Amélie
1,532 shots
26.11 mean L*
0.82 mean S
Digital control turns the warm bias into a nearly continuous chromatic field.
Amélie is the saturation outlier: 83% of shots exceed S 0.70 and the median is 0.86. Neutral pixels nearly disappear. The famous red–green complement matters, but the balance is asymmetric: yellow-red warmth occupies most of the frame while green operates as a designed opponent, especially in walls, night exteriors and memory-like inserts.
Skin, olive and red negotiate the frame
Warm skin is held against olive green, with almost no genuinely neutral field.
Paris is pushed toward recollection
Street and afternoon collapse into a single golden register; warmth is also saturation.
Coolness arrives as drained memory
The green-grey insert punctuates a film in which highly saturated shots are normal.
The palette can carry daylight
Window light lifts the café without neutralising its red–green design.
Darkness studyA rare punctuation inside the abundant chromatic field
The cinema makes darkness communal
Faces and seats arrive as fragments; colour is rationed and attention becomes collective.
Window light cuts intimacy into silhouette
The hot curtain becomes a boundary between private gesture and exterior space.
Even the dark register stays chromatic
Amélie’s darkness is rarer, but red and green remain dense when the key falls.
A lit window survives inside the night wall
Darkness interrupts abundance and turns habitation into the chromatic accent.
04
2004 · Digital
A Very Long Engagement
1,388 shots
24.20 mean L*
0.58 mean S
The digital palette relaxes saturation without surrendering yellow light.
Mean saturation falls close to Delicatessen, making this film a useful rebuke to a simple “digital equals vivid” story. Its mean key and contrast remain near Amélie, while yellow b* stays elevated. The signature appears here as a flexible scaffold: golden recollection and landscape can coexist with cyan-grey war material without requiring one invariant intensity.
Amber becomes archival
Paper, wardrobe and light meet in a restrained brown-gold field.
The field turns yellow into destiny
The digital grade makes warmth spatially immense rather than merely dense.
War is permitted a cyan-grey register
Brown skin and earth keep the restrained cool field from becoming clinically blue.
Distance is graded as gold
The open road raises the key and scale while remaining part of the same colour world.
Darkness studyMemory, war and distance organised by limited light
The lighthouse becomes a vertical memory
The whole frame stays below L* 20; shape and atmosphere replace conventional detail.
Blue-black isolates the wounded body
A small cyan figure sharpens vulnerability without opening the surrounding space.
Gold outlines thought without filling the room
Warm light supplies contour and memory while the interior retains its weight.
Architecture funnels the eye toward one figure
The black-gold corridor turns repetition into emotional distance.
Finding 03 · Darkness as character
Lost Children does not invent a different kind of black. It lets darkness occupy the film.
Seventy-eight percent of its shots sit below mean L* 15, and 54% fall below L* 10. More importantly, one low-key passage persists for 87 consecutive shots—about eight minutes. The longest comparable runs are 29 shots in Delicatessen, eight in Amélie, and twenty in Engagement. Darkness becomes duration: a condition the audience inhabits rather than an occasional lighting event.
The surprise comes inside the dark shots. Their average anatomy is broadly similar across all four films: roughly 58–65% of the active picture falls below L* 5, while 81–84% remains below L* 20. Lost Children is not simply crushing each dark frame more aggressively. Its distinction is editorial and cumulative—how often darkness returns, how long it refuses relief, and how small islands of light become events within it.
03A · Occupancy
How much of each film lives below the key
Shot mean L* bands; one midpoint still per detected shot
Read it: almost nine shots in ten in Lost Children remain below mean L* 20. Its darkest half alone exceeds the complete sub-L* 20 share of either digital film.
Persistence, not just prevalenceConsecutive shots below mean L* 15
Read it: the 90th-percentile run in Lost Children is already seventeen shots. Its 87-shot maximum is not a momentary blackout—it is a sustained dramatic register.
Anatomy of a dark shotOnly shots with mean L* below 15
Read it: the four dark-shot structures are unexpectedly close. Amélie even retains the most chromatic shadows; what makes Lost Children singular is recurrence and duration.
What this means at the grading desk
A useful low-key review should therefore ask three separate questions. First: floor—how much of the frame is allowed below L* 5? Second: legibility—does the remaining shadow band preserve texture and colour separation? Third: rhythm—how many cuts can remain under the same perceptual adaptation before a highlight resets the eye? A waveform can answer the first two. Only the timeline answers the third.
Finding 04 · The transition
Digital grading lifts the image, strengthens yellow and removes neutral territory.
Balancing the two films within each era, the digital pair gains 9.95 L* in mean key, 4.71 L* in contrast and 7.59 b* toward yellow. Saturation rises by 0.109. The striking non-movement is mean a*: only +0.11. That pattern suggests a transformed dynamic range around a relatively stable red–green centre—not a wholesale replacement of the colour grammar.
Finding 04 · Era comparison
What moves, and what stays put
Two film means per era; scales are metric-specific
Read it: the comparison is descriptive. With only two films in each era, it cannot isolate medium from production design, cinematography, story or restoration.
Finding 05 · Tonal architecture
The warmth is built upward through the tonal scale—not poured over the whole image.
In every film the shadows remain comparatively restrained, averaging only C* 7–11, before chroma jumps to C* 34–43 in the midtones and C* 38–48 in the highlights. Yellow b* follows the same climb. That is a grading strategy: preserve weight and separation below, then let colour become expressive as luminance opens.
The red–green axis does something subtler. Red a* generally peaks in the midtones—where faces, wood, brick and costume live—then retreats in the highlights. In Engagement, the mean highlight reaches b* 42.5 while a* returns almost exactly to neutral. The result is gold light rather than indiscriminate orange contamination.
The tonal ladderMean within-shot L* percentiles
Read it: Lost Children compresses the typical image into the bottom third; the digital films preserve a much longer P10–P90 span.
Tonal-zone vectorsShadow → midtone → highlight in CIELAB a*b*
Read it: all four vectors rise strongly toward yellow; none simply pushes red and yellow together at every exposure.
Scope pressure
Where the signal is being asked to work
Near-zero RGB occupancy versus high-C* occupancy; circle area = channel-ceiling contact
Practical reading: Lost Children parks one third of active-picture pixels at RGB code values 0–2, while only 7% exceed C* 40. Amélie reverses that pressure: 31% high-chroma pixels and the corpus’s highest single-channel ceiling contact. These are encoded-signal warnings—not proof of illegal levels or creative clipping.
Finding 06 · The colour space
The corpus forms overlapping neighbourhoods, not two era islands.
A principal-component map condenses nine shot features into two axes. The first axis is driven chiefly by warm fraction, yellow b*, neutral absence and saturation. The digital films move rightward, but their shot clouds overlap the photochemical films. Most tellingly, the closest pair of mean feature centres is cross-era: Delicatessen and A Very Long Engagement.
Shot-feature mapStandardised PCA; 260 deterministic samples per film
Distribution distanceMean difference across seven quantiles × nine features
The full distributions—not just their centres—put the two digital films closest together (distance 0.356), with Delicatessen and Engagement next (0.409). That dual result is exactly what a persistent but adaptable signature should look like: medium matters, while a cross-era kinship remains measurable.
Finding 07 · Selective shaping
We cannot see Lucas’s nodes. We can test for the rendered fingerprints of isolation.
A single finished frame cannot disclose a tracked Power Window, key, or node tree. Lighting, lens falloff, blocking and production design can make the same shape. But two proxies are still worth testing across thousands of shots: whether the central image is repeatedly lifted against its edges, and whether coherent skin-coloured regions are held apart from their immediate—not global—surround.
Skin versus near-surroundMedian local differences; bars show interquartile range
Read it: coherent candidate regions are locally brighter by a median 19–25 L* in every film. Amélie separates them far less by chroma—consistent with keeping flesh inside an already saturated world.
Centre versus edgeFixed radial proxy; median and interquartile range
Read it: the median centre is modestly brighter and more chromatic in all four films, strongest in Lost Children. Wide intervals forbid calling this a detected vignette or window.
Context, not a beauty key
Heuristic skin-tone centres in CIELAB
Circle area = estimated fraction of frame
The broad detector identifies roughly 8–12% of each frame. Mean candidate skin moves from L* 36–39 photochemically to L* 47–57 digitally, but costume, wood and warm walls can pass the same HSV/YCrCb gate. The local chart above first removes tiny components and is the more useful—but still heuristic—separation measure.
Finding 08 · Cut matching
Shot-to-shot continuity leaves a measurable fingerprint.
For every cut, the five-colour palettes on either side were compared with an area-weighted Earth Mover distance in CIELAB. Adjacent shots are then tested against random pairs from the same film. Every corpus exhibits continuity: its actual cuts are 30–40% closer at the median than arbitrary pairings. That is compatible with deliberate match work, but scene, production-design, camera and editorial continuity all contribute.
Absolute distance answers a different question from relative matching. Lost Children has the quietest cuts because its entire palette is compressed into darkness. Amélie makes the largest typical jumps, yet still pulls adjacent shots 30% closer than its own broad look range. Engagement shows the strongest relative match—40%—while retaining conspicuous accent cuts in its upper tail.
The cut desk
Actual neighbours versus random shots
Palette EMD in ΔE76-like Lab distance; lower = closer
Use it: the 90th-percentile actual cut marks each film’s candidate-discontinuity territory—about 20 for Lost Children, 32 for Delicatessen, 33 for Amélie, and 37 for Engagement. Those are useful review thresholds, not universal grading tolerances.
Finding 09 · Colour movements
A grade has movements, reprises and dramatic weather—not just a house look.
Colour alone cannot classify happiness, sadness or humour. It can measure some of the conditions through which those feelings are staged: whether the image opens or encloses, how much chromatic energy it carries, how long a register persists, and how sharply it changes. On those terms the four films have radically different emotional geometries.
The analysis below detects 22–30 contiguous colour movements per film. These are not screenplay scenes. A movement is a run whose palette, key, contrast, warmth, shadow share and chromatic pressure are relatively stable. Five recurring look families then reveal when a film returns to a visual world after leaving it.
09A · The colour score
Dominant palettes through actual runtime
Width = duration · each vertical stack = one detected movement
Read it: Amélie sustains dense amber, yellow and red across long stretches. Lost Children makes near-black the ground on which brief chromatic openings appear. Delicatessen alternates compressed brown-black and more open amber; Engagement ranges between restrained earth, gold release and blue-grey war.
High key, saturation, highlight access and high-chroma occupancy remain jointly elevated for three quarters of runtime.
Its sadness is compatible with a visual world that withholds release; only 0.4% of runtime enters the radiant range.
The fastest significant turn rate and the most evenly shared look repertoire support tonal elasticity—an opening for the film’s macabre silliness.
A restrained opening expands through the middle and returns to restraint: less a fixed mood than memory, ordeal and provisional release.
Visual-affect trajectoryRadiance and chromatic energy · 0–100 descriptive indices
Read it: Amélie operates high on both axes but still falls sharply near the end. Lost Children remains low even when its colour briefly intensifies. Engagement contains the clearest broad middle expansion.
Recurring look repertoireFive film-relative families; width = runtime share
Read it: a family may return in several non-adjacent episodes. Delicatessen distributes time most evenly across its five families; the other films possess more dominant home registers.
What this gives a colourist
Treat the boundary list as a review index, not an automatic grade. It can expose an unintended look drift within a sequence, show where a new visual chapter begins before anyone names it, and test whether a late reprise truly returns to the earlier palette or merely recalls it narratively. The recurring families are a data-backed starting point for a look bible: choose hero stills inside each family, inspect scopes and node intent, then accept, merge or reject the machine’s grouping. The generated CSV and Parquet tables preserve every movement’s shot range, timecodes, palette, family and diagnostic values.
The practical synthesis
Four grading fingerprints a colourist can actually interrogate.
These are not reconstruction recipes. They are compact scope-reading hypotheses: where each film puts tonal mass, how colour climbs through exposure, how a likely subject is separated, and how closely neighbouring palettes land. Every value below is computed from the active picture.
01 · 1991
Delicatessen
Dense blacks support an amber scale that becomes more chromatic as it rises. Coherent candidate flesh regions separate by both lift and chroma.
02 · 1995
Lost Children
The lowest, narrowest tonal field; colour is rationed into local objects and likely flesh regions. Its rendered separation proxies are strongest here.
03 · 2001
Amélie
The widest chromatic field and greatest channel pressure. Candidate flesh regions are lifted, but not chromatically detached from their saturated surrounds.
04 · 2004
Engagement
A more restrained digital chroma field with an open floor, broad highlights and the strongest relative cut-to-cut palette continuity.
Conclusion
The evidence supports continuity with transformation.
Lucas’s most persistent signature in this corpus is a warm direction with purposeful cool exceptions, not a fixed level of saturation or exposure. Photochemical timing accommodates both Delicatessen’s pervasive amber and Lost Children’s rationed colour. Digital grading then broadens the available key and yellow range: Amélie intensifies almost everything; Engagement retains the expanded light while relaxing saturation.
That is a stronger answer than “the films look alike.” It identifies what is stable, what changes with the toolset, and where the comparison fails to separate Lucas from Jeunet, the cinematographers, production design and the source editions. The signature is visible—but it is collaborative and adaptive.
The temporal evidence sharpens the distinction. Amélie makes radiance a durable home state; Lost Children makes darkness an environment; Delicatessen turns between compressed and open amber with comic elasticity; Engagement builds a broad middle expansion before returning to restraint. A shared warm axis can therefore carry joy, sorrow, grotesque play and remembrance without collapsing those films into the same mood.
Method & limits
How to read the numbers
Corpus
Resolve Studio scene detection produced one timeline item per detected shot. The pipeline exported the midpoint frame of each item and analysed a 512-pixel-long-edge derivative, excluding detected letterbox/pillarbox bars. The final corpus contains 5,288 stills: 1,116, 1,252, 1,532 and 1,388 shots in chronological film order. Unreadable-file count: zero.
Measurements
Per-shot features include five-colour CIELAB k-means palettes; mean and standard deviation in HSV and CIELAB; warm, cool and neutral pixel shares; L* key and contrast; and a skin-tone estimate. The grading diagnostics add L* percentiles; shadow, midtone and highlight occupancy and Lab vectors; C* pressure; encoded near-black and channel-ceiling shares; fixed radial centre/edge differences; coherent skin/near-surround differences; and adjacent-palette distance. Film palettes re-cluster all shot palette centres using their pixel proportions as weights. Temporal analysis adds duration-scaled colour movements and film-relative recurring look families.
- Warm
- Chromatic HSV hue 300°–360° or 0°–90°.
- Cool
- Chromatic HSV hue >90° and <300°.
- Neutral
- Pixels below the analysis saturation threshold (S < 0.08).
- Strongly warm-led shot
- Warm / (warm + cool) > 0.80.
- Low key
- Per-shot mean L* < 15.
- Highly saturated
- Per-shot mean HSV saturation > 0.70.
- Tonal zones
- Shadow L* < 20; midtone L* 20–65; highlight L* ≥ 65.
- Deep shadow
- Active-picture pixels below rendered CIELAB L* 5.
- Shadow span
- P90 minus P10 computed only among pixels below L* 20; a code-space differentiation measure.
- Light islands
- Scale-qualified connected L* ≥ 20 regions after a small opening; rendered topology, not detected sources or masks.
- High chroma
- Pixel C*ab = √(a*²+b*²) ≥ 40.
- Near black
- All encoded RGB channels ≤ 2/255; a signal-pressure flag, not proof of clipped grading.
- Channel ceiling
- Any encoded RGB channel ≥ 253/255; may be saturated colour rather than white clipping.
- Local skin separation
- Mean coherent candidate-region Lab versus a nearby dilated ring; ΔE is Euclidean CIELAB 1976.
- Centre/edge proxy
- Normalised radial centre r ≤ 0.50 minus outer field r ≥ 0.85.
- Cut distance
- Area-weighted Earth Mover distance between adjacent five-colour Lab palettes.
- Colour movement
- A contiguous, minimum-ten-shot span selected by multivariate variance reduction over palette vocabulary, key, saturation, contrast, Lab centre, warmth, shadow share and high-chroma share; it is not an editorial scene.
- Recurring look family
- One of five film-relative k-means groups over movement signatures; non-contiguous movements may belong to the same family.
- Visual radiance
- A transparent 0–100 descriptive index: 36% mean key, 22% saturation, 18% high-chroma occupancy, 12% highlight occupancy and 12% inverse shadow occupancy. “Radiant” means index ≥ 55. It is not a happiness score.
- Chromatic energy
- A 0–100 index combining mean saturation (45%), high-chroma occupancy (35%) and mean C* (20%).
- Enveloping darkness
- A 0–100 index combining deep-shadow occupancy (45%), shadow occupancy (35%) and sub-L* 15 key pressure (20%); the headline share uses index ≥ 55.
- Radiance reversal
- A change in direction after a movement-to-movement radiance shift of at least five index points; expressed per hour of actual runtime.
Frame curation
The first four candidates per film were surfaced computationally in broad look strata—baseline, strongly warm, cool-led and high-key—then reviewed visually. A second quartet per film samples four uses of darkness: total field, aperture, coloured shadow and pool of light. The final thirty-two frames maximise evidentiary range and compositional readability. They are lossless 1280-pixel PNG derivatives cropped to the same source-level active-picture bounds used by analysis; no creative colour transform, filter or CSS effect is applied. Metrics always come from the 512-pixel analysis derivatives, not these display files.
Limits
- Four films are four observational units—not 5,288 independent experiments.
- Scene detection and one midpoint frame can miss internal transitions, dissolves and shot-scale variation.
- Colour-movement boundaries are candidates for human review. They do not recover screenplay scenes, locations, narrative beats or emotional intent.
- Radiance, darkness and chromatic-energy indices are declared descriptive composites. Emotion also depends on narrative, performance, framing, motion and sound.
- The input transfer was provisionally treated as gamma 2.4. Absolute Lab claims should be revalidated with known colour patches or colour-managed reference exports.
- The source editions encode acquisition, timing or grading, mastering, restoration and display transforms together.
- The skin mask is an HSV/YCrCb heuristic, not face detection, and will admit similarly coloured materials.
- Centre/edge and local skin separation describe rendered geometry. They cannot prove a vignette, key, tracked window, relight, or any hidden Resolve operation.
- Adjacent-palette distance mixes grading continuity with scene, production-design, camera and editorial continuity; random-pair comparisons are an internal baseline, not a control group.
- The design cannot causally assign a look to one collaborator. It describes a Lucas-associated Jeunet corpus.