I was scrolling Netflix one day looking for a horror movie. And something hit me.
The posters looked different from what I remembered.
Back when I rented horror on VHS, the boxes screamed at you. Bright colors. Masked killers. Blood red. You had two seconds to decide. The poster was the whole pitch.
But now? The new sheets look quiet. Dark. Sometimes just black with one word. Fewer faces. Less color. Nothing jumping out.
I love horror films. I've watched them my whole life. I also love data — finding patterns in things.
So I started wondering: when did horror stop shouting? Did the posters actually change, or is it just me? Because if they did, that meant something. It meant fear looked different now.
I decided to find out. I pulled 33,622 horror posters — from 1900 all the way to 2026 — and I measured them. Colors. How dark they are. What's on them. Everything. What I found surprised me. And I think it may change how you look at a horror poster.
About the data
From TMDB Horror titles I kept 33,622 English-language posters with usable art — 1900 to 2026. Filters: already released, no adult flag, no Animation, Music, or TV Movie, plus the existing exclusion lists. Then I measured every sheet the same way. The fine print is at the end, if you want it.
Exhibit A · The instruments
How a Sheet
Is Measured
Every chart that follows is built from the same instruments. Here they are, on one poster — Halloween (1978). Tap a measure to see how it is built, and how to read the number.
Tap a measure
How it's built
How to read it
Same instruments. Now the whole century.
Part I · Color River
Black Took the Sheet
Look at the colors. In the 1950s, horror posters were still warm: a quarter of the sheet was orange and amber, as it had been since the 1920s. Then the color drained out. In the '60s the sheets turned grey and washed out; from the '70s the grey gave way to black. The '80s get a quick flash of blue (neon fear). It doesn't last. Black wins. Near-black pixels go from 29% of the sheet in the 1950s to 51% in the 2020s. The warm colors never really come back.
My reading — the data show what changed, not why. The warm '50s aren't random. That was the atomic age — bright reds and radioactive greens sold "COLOSSAL TERROR!" at drive-in theaters. The blue flare in the '80s? That's VHS neon, the video-store era. And the black flood that never recedes? That's horror learning to be scary without color at all.
One thing to keep in mind while you read: not everything below belongs to horror. Thriller, mystery and sci-fi posters lost faces and text over the same century, and I measured them the same way (Part VIII). What stays specific to horror is the black: 51% of the sheet in the 2020s, against 40–44% in the others.
FULL CENTURY — n = 33,622 posters (TMDB), mean share per decade. Black/Grey counts every near-black pixel and every grey or white one (low saturation).
One real poster per hue family the chart measures — the sheet where that band actually dominates, per era.
On phones: swipe each row across the six eras. Showing Reds & Black/Grey first — the two bands that drive the story.
Selected by dominant hue-band pixel share (same six families as the river), favoring recognizable illustrated sheets over blank title cards.
Part II · Darkness
Posters Got Darker After 1970
I expected brightness to go up and down, like a normal trend. But this wasn't normal. Horror posters have kept getting darker for more than fifty years, with no real rebound. The light never came back.
It isn't a slide — it's a mountain. Brightness holds a broad plateau from the mid-'40s to the '70s, topping out around 1970, then drops in the '80s and never recovers. No rebound.
My reading — the data show what changed, not why. The 1970 peak isn't a coincidence. That was the tail of the atomic age — bright posters promised "SEE! SHOCKING!" alien invasions and mutant monsters, radiating Cold War panic in reds and yellows. Then Vietnam happened. Watergate happened. Horror stopped being about fantasy monsters and became about the real world. The lights dimmed. And they never came back.
FULL CENTURY — n = 33,622 posters (TMDB), 5-yr rolling mean.
The brightest and the darkest poster the pipeline measured in each era — the two ends of the mountain.
Selected by mean L* brightness, one extreme specimen per row per era.
Part III · Red & Blood
Red Peaked Around 2000
You'd think the '80s slashers owned the color red. They didn't. Red is horror's signature — but its real peak isn't where you'd guess. Not the slashers. The meta-horror turn of the millennium.
My reading — the data show what changed, not why. Why the delayed peak? The '80s slasher era was born from VHS. More than 1,300 horror films in this corpus came out in that decade — posters had to scream red from a video store shelf just to compete. But by the 2000s, red had become a language. Meta-horror (Scream, Saw, Final Destination) used blood as an aesthetic — not a threat.
FULL CENTURY — n = 33,622 posters (TMDB), 5-yr rolling mean.
What the reddest poster in each era actually looked like — and, just as telling, what a horror poster with essentially no red looks like in the same years.
Selected by red-pixel share (HSV red-hue mask), extremes per era.
Part VII · Two Kinds of Blood
Blood No Longer
Has to Be Red
Here's the weird part. Red and blood aren't the same thing. The chart above measured a color. This one measures a concept: does this look like blood, even if it isn't red? Through the 1990s the two measures rise together. Then they part: the color falls back to about a tenth of the sheet, while blood on the sheet peaks in the 2000s and is still well above its '80s level today.
My reading — the data show what changed, not why. The split makes sense: after the slasher boom, blood became a symbol — not just a color. It could be black, brown, dark red, or even implied. Meta-horror posters (Scream, Saw) treated blood as a language everyone already understood. You didn't need to paint it red anymore. The audience knew.
Pixel-red: FULL CENTURY, n=33,622 (TMDB). Model-detected blood: FULL RUN, n=33,622, CLIP zero-shot per 2×2 poster patch — validated against hand-inspected artwork (Jaws, Friday the 13th, Blair Witch) before trusting it. Decade means.
Where the two measures in the chart above actually disagree, poster by poster: high pixel-red with zero semantic blood detected, and the reverse.
Both rows drawn from the full-run blood model (n=33,622) as the chart above; "blood" = clip_blood zero-shot score. Bottom-row specimens hand-checked for literal blood — early high scores often fire on red cloth or type instead; the first cell is the earliest verified case shown (1959).
Part IV · Faces & Monsters
More Ghosts.
Fewer Faces.
There's something primal about a face staring back at you from a poster. It tells you what to fear. So I asked a simple question across the whole century: who looks back from the sheet? First the human face. Then the creature. Together they tell one story: horror learned to sell dread without a star close-up — while the throne of the monster kept changing hands. And the ghost, which led once before (narrowly, in the 1940s), tops the CLIP census today (a second model, Nova, puts it second to the masked killer).
The vanishing face. There's something primal about a face staring back from a poster: it tells you what to fear. In the 1940s that stare was nearly mandatory — 87% of sheets had one, the century's peak. The rolling curve hits 91% in 1945. Today, fewer than half do. Horror learned to be scarier by showing you less. What's scarier than a monster? Absence. And hiding inside the long retreat, one twist: the late '90s didn't add faces, it zoomed in on them. Face size peaks around 1998–2002 — the celebrity close-up era: on sheets that show a face, it fills about 13–14% of the sheet, against 8% in 1950–89 (and 10% in the 1930s–40s). Those zoomed-in faces had names. Scream put five of its cast on the sheet; The Faculty, six. Among well-known horror films (100+ TMDB votes), 53% of the late-'90s posters sold a star from the cast (93 posters, 95% interval 43–63%); today 36% do (800 posters). The face that vanished is mostly the small film's face: today 59% of well-known films' posters show a face, against 47% of the rest, and 36% carry a cast star, against 5%.
My reading — the data show what changed, not why. The 1940s face peak makes sense: Universal built the horror genre on painted portraits of Dracula, Frankenstein, the Mummy. The monster was the movie. But by the 2010s, elevated horror (A24, Hereditary, Midsommar) taught us the opposite: what you don't see is scarier. The face vanished because our fear turned inward — we stopped fearing monsters and started fearing our own minds.
FULL RUN — n = 33,622 posters (TMDB), YuNet face detection (validated against hand-checked artwork), 5-yr rolling mean. Counts are floors: small, profile, tilted, or heavily painted/illustrated faces can escape the detector. STARS — Rekognition celebrity recognition on the 17,018 posters with a face, kept only when the name is in the film's TMDB cast (3,161 posters); well-known = 100+ TMDB votes. A floor: lesser-known actors aren't recognized.
The two faces this chart is actually made of: the extreme close-up, and the poster with no detectable face at all.
Selected by YuNet face-detection output (n_faces, face bounding-box area) — the close-ups shown fill roughly half to all of the sheet. Zero-face row: hand-checked so the sheet really has no readable human face (silhouette, creature-from-behind, egg, light-for-a-head) — not detector misses on painted or tilted faces, which YuNet also produces.
In Asia, the face stayed. It left English-language horror posters, not Asian ones. Since 2000, 49% of English-language horror posters show a face. Japanese: 66%. Thai: 82%. Spanish, French, German and Italian horror sit right next to English. Japan dipped and came back: 75% of its posters showed a face before 1980 (112 posters), about 61% in 1980–2009, 69% today — while English kept falling, to 48%. And the title kept shouting. Put a Japanese horror poster next to an English one: the film's name is over a third bigger — 16.5% of the sheet's height against 12.0%. Korea and Thailand don't follow; their titles sit at English size or below.
OTHER LANGUAGES — horror films outside the 33,622-poster English corpus, measured with the same pipeline, face detector (YuNet) and filters. Top: 4,598 posters from 2000–2026 in the 12 languages with 150+ posters; lines are 95% intervals. Bottom: every year, grouped by era (hover a point for its n). About a third are the film's English-market poster; the gap holds for both versions. Creatures don't split the same way: 27% of English-language posters show one (CLIP census) against 16–25% in the other languages, but among films with 10+ TMDB votes the intervals overlap. TITLE SIZE — PaddleOCR on 2,508 Japanese, Chinese, Korean and Thai posters and a sample of 600 English ones; the comparison uses the 2000–2026 posters where the title was found (475 of 599 English, 638 of 829 Japanese, 169 of 246 Korean, 124 of 217 Thai); median height of the title's largest line of letters, with 95% intervals: English 12.0% (11.3–12.7), Japanese 16.5% (15.7–17.2), Korean 13.0% (11.9–14.9), Thai 10.2% (9.3–11.2). 37 of 40 boxes checked by hand measured the right line. Part of the gap is the script: a kanji title of two to four characters fits larger than an English word.
The monster census. When the human face retreats, something else often fills the stare. I asked a vision model one question about every poster: what creature is on it? The answer is a dynastic succession — and it ends with the ghost: in the 2020s it is the most common creature on this census, 23% of the posters that carry a label, more than twice the runner-up. A second reader (Nova) ranks the masked killer first and the ghost second, so the order at the top depends on the model — switch the reader above the chart to see Nova's version. Nova names a creature on more posters (54% against 27%), so its shares are higher and its curves differ most in the middle decades: it finds masked killers on 23% of the 1980s posters, CLIP on 4%. Before 1950 there are too few labeled posters to name a leader (13 in the 1920s, 27 in the 1930s). The chart keeps the ghost and the masked killer lit; the other thrones stay dim until you tap or hover a line.
FULL CENTURY — n = 33,622 posters, CLIP zero-shot vs 18-creature taxonomy (checked on 10 hand-checked posters (9/10; the miss is Annabelle) and against a second reader, Nova, which agrees on 67% of the labeled posters (82% when Nova names a creature); labels below 0.5 confidence excluded). Shares are conservative floors; pre-1950 decades carry smaller samples. About 7 in 10 posters carry no creature label (below 0.5 confidence), so the shares describe posters that show a recognizable creature; before 1950 each decade has fewer than 70 labeled posters.
Two real posters per era for whichever creature actually held the throne, per the chart above.
Selected by top CLIP-classified confidence within each era's reigning label(s) named in the chart's own copy above — where two creatures co-reign or hand off mid-era, one specimen of each.
Part V · Text & Composition
Half the Text
of 1960
Picture a 1950s shelf: painted, dimensional letters SHOUTING at you. Now lean into a 2020 lobby card — one thin word on black. You have to get closer to read it. That whisper is scarier than the shout. Three shifts, measured across every poster. First: from the silent era through 1960, text holds at about a third of the sheet — the "SEE! SHOCKING!" years. Then horror lowers its voice: to about a quarter of the sheet by the late '70s, and, after 2010, to 18% today. An OCR check supports the first drop but not the second (see the note below the chart).
My reading — the data show what changed, not why. The 1950s shout wasn't random — drive-in theaters were exploding. Posters had to grab you from your car window. By 2020, posters became thumbnails on Netflix. You browse them at arm's length, on your couch. The volume dropped because the distance did. Fear doesn't need to shout anymore. It just needs to be seen.
FULL RUN — n = 33,622 posters (TMDB), MSER text-region proxy, 5-yr rolling mean. Trend-comparable across eras; absolute level is an upper bound. OCR CHECK — EasyOCR on 800 posters (80 per decade, 1930s–2020s): readable lettering covers 18% of the sheet in the 1930s and 9% in the 1970s, then stays between 9% and 12%. Per-poster agreement with this measure is weak (Spearman 0.26) and the OCR level is about half. Read the fall to the 1970s as confirmed and the later fall as unconfirmed.
What "text-like coverage" actually flags, poster by poster: five specimens spanning the shout-to-whisper arc, from the atomic-pulp peak down to a single quiet word. The amber boxes below are the real output of the same MSER pass charted above, run on this page's reduced copy of each poster — not illustrative, detected, photographic texture and false positives included.
Attack of the Crab Monsters (1957)The title alone eats the top third of the sheet — then the detector keeps firing on every glint, shadow and film-grain fleck in the pin-up art below it. This is the "upper bound" caveat made visible: real lettering and pure texture trip the same wire.
text-like coverage 61% (among the highest 0.5% in the dataset)
Night of the Living Dead (1968)Still a shout — three stacked words command the top third, boxed almost as tightly as real lettering gets. But most of the density below comes from grain and contrast in the collaged faces, not ink.
text-like coverage 45%
Phantasm (1979)Halfway down the mountain: the logo is a fraction of the '50s peak's size, and it's nearly the only place the detector fires with any real density — the painted illustration itself stays mostly unboxed.
text-like coverage 31%
Scream (1996)One word, bottom-anchored, clean sans-serif — the detector boxes it tightly and finds almost nothing else on the sheet. The '90s "meta" look was already most of the way to today's whisper.
text-like coverage 19%
Midsommar (2019)A few stray boxes on the flowers below are nearly the whole reading — the thin serif logo barely registers as text at all. This is what the whisper looks like, pixel by pixel.
text-like coverage 9%
Boxes shown are the actual MSER glyph-candidate regions (delta=5, 15–2,000px² area, 0.1–12 aspect ratio) detected at the same 180px analysis width used across the full run — not merged or cleaned up. The scatter across photographic texture above is the method's honest failure mode: trend-comparable across a century, not a precise per-poster reading. These specimens were run on the 342-px copy shown on the page, not on the full-size image the dataset measured, and the result moves: in the dataset Night of the Living Dead reads 32% (45% here), Scream 27% (19%) and Midsommar 32% (9%). The same poster can read very differently at different resolutions, which is why the chart is a trend and not a per-poster reading.
Second: posters became more symmetric. Mirror-balance rises through 1980, holds roughly flat for three decades, then rises again after 2010 (about +0.05 on a 0–1 scale) — but not only in horror. Thriller, sci-fi and mystery posters rise by 0.04–0.06 over the same years. It's how movie posters are made now, not a signature of elevated horror.
FULL RUN — horizontal mirror similarity (0–1), 5-yr rolling mean.
What mirror similarity actually catches, poster by poster: fold each sheet on the white center line below and ask, pixel by pixel, how far the left half is from the right half's reflection. High scores mean the two halves nearly match; low scores mean they don't. I dropped the live difference heatmap — on asymmetric sheets it just looked like a second, flipped poster — so what you see is the fold the score uses, plus the number.
Braid (2019)A kaleidoscope collage built to be mirrored — skulls, faces and antlers folded around dead center. About as close to a true reflection as an illustrated poster gets.
mirror similarity 0.98
History of the Occult (2023)A face framed dead-center, hands mirrored on both temples — the centered composition that took over movie posters after 2010.
mirror similarity 0.98
Get Out (2017)The figure is dead-center — but the background is split black-left, white-right. To a person this reads as a balanced, deliberate composition. To a pixel-mirror test it's a near-total mismatch: a real gap between "centered" and "symmetric."
mirror similarity 0.45
The Hitcher (2007)A single figure planted hard on one side of the sheet, empty road on the other — fold the poster on the white line and the two halves don't match.
mirror similarity 0.60
Ghost Estate (2019)The single lowest mirror-similarity score in the entire 33,622-poster dataset — on a poster that is, thematically, about a black/white split down the middle. The design quotes symmetry; the pixels don't deliver it.
mirror similarity 0.10 (dataset minimum)
White line = the fold the score uses. The pipeline measures per-pixel |gray − mirrored gray| at 64×96 across that midline — high scores mean the two halves nearly match; low scores mean they don't. Caveat: this is pixel-value mirror-match, not human-perceived balance — a centered subject on a split-tone background can score as "asymmetric" even when it reads as balanced. n = 33,622 (TMDB), full run. History of the Occult is not part of the 33,622-poster corpus.
And third: the compositions themselves stopped leaning. I measured how much of each poster's linework — a blade, a slashed logo, a body thrown off-axis — runs on a diagonal rather than sitting flat and upright. It peaks in the atomic-pulp years, from the '40s to the '60s, steps down in the '70s, holds until the '90s and then keeps falling: today's horror poster stands up straight.
FULL RUN — n = 33,622 posters, share of Hough-detected line length at a 25–65° angle, 5-yr mean. A real share of this is stylized diagonal title lettering, not only figure pose — see Sources. The score runs about 11 points higher on TMDB images than on release-year originals, evenly across decades, so read it as a direction, not a level. On 903 films from 1930–1999 the fall shows up in both sources (about −5 points); after 1999 there is no original to compare against.
What does a high diagonal_score actually look like on the sheet? Five specimens the algorithm flagged, spanning five decades, none of them repeats from Exhibit A or B. The red dashed lines below are the real output of the same Hough-transform pass charted above — not illustrative, detected. Watch each sheet dim so the lines can take the stage.
Doctor X (1932)The title lettering isn't near a diagonal — it is one. A giant yellow "X" slashes corner to corner, doing triple duty as logo, warning sign, and composition.
diagonal score 0.65
It Came from Outer Space (1953)The fleeing cast is arranged on one long sightline toward the saucer's glow, and the stacked title climbs the same angle up the right edge.
diagonal score 0.58
The Howling (1981)Four claw-rips tear across the frame on a consistent diagonal, dragging the eye from the screaming mouth up to empty space — violence rendered as pure geometry.
diagonal score 0.64
Spawn (1997)One uninterrupted cape edge runs the full length of the sheet, corner to corner — the whole silhouette built as a single diagonal blade.
diagonal score 0.76
AVP: Alien vs. Predator (2004)Two creature heads, two studios' worth of IP, arranged along one diagonal seam — Alien upper-left, Predator lower-right, the rivalry staged as a fault line.
diagonal score 0.69
Lines shown are the actual longest Hough segments (of ≥25 total per poster on average) that fell in the 25–65° diagonal band for that specimen, at the same 180px analysis width used across the full run. Animation dims the artwork so the detected segments can be read; hover any poster to pause. Specimen scores are computed on the page's reduced copy; the dataset value for the same poster can differ (The Howling 0.57, AVP 0.39 in the dataset).
Part VI · Lettering
From Painted Letters to One Thin Word
Color wasn't the only language on the sheet — so was the title. And titles got quieter too. In the 1940s, 62% of posters wore ornate display lettering. Barely 3% were minimal. By the 2020s that flipped — ornate down to 12%, minimal up to 29%, now the single largest register. Fear stopped shouting its own name. And it isn't only horror: thrillers and mysteries made the same move earlier and went further, and sci-fi held on to ornate lettering longer than horror did.
My reading — the data show what changed, not why. The 1940s lettering was hand-painted by Universal artists like Karoly Grosz — it was craft. But by the 2010s, A24 turned typography into prestige. A single thin word on black doesn't just look modern — it says: this movie is art, not a monster movie. Ornate became "cheap." Minimal became "elevated." The letters changed because the marketing did. The critics didn't get the memo. Measured, though: across 1,714 films with a Metascore, a thin minimal title earns no better reviews than an ornate one from the same decade (critics: −0.7 points for ornate, 95% interval −3.9 to +2.3). The poster sells a mood; it doesn't predict the verdict.
FULL CENTURY — n = 33,622 posters (TMDB). Registers are quintiles of an ornate↔minimal typographic axis (CLIP zero-shot on the whole poster: similarity to "ornate hand-drawn display lettering" minus "clean minimal type"). Validated against 10 hand-ranked specimens (corr 0.85, 95% interval 0.48–0.96); across the 11 decades 1920–2020 the ornate share falls with the decade at r = −0.93 (per poster the axis correlates with year at −0.28). This measures the decorative character of the title treatment, not a specific typeface — a proxy, honest about its limits: reading the exact letterforms of stylized horror titles defeats OCR and font-ID models alike. REVIEWS — OMDb Metascore and IMDb rating (100+ votes) for 1,714 corpus films; titles compared within the same decade. GENRES — the ornate↔minimal axis is on one scale across genres, so the genre comparison uses the axis itself; the register percentages in the text are quintiles within horror.
The most- and least-ornate title treatment the model found in each era. Fear's lettering thins from painted, dimensional logos to a single quiet word.
Thirds Perfect,
Still Off Balance
Every poster is a design decision, not just a color choice. I measured three things per sheet — rule-of-thirds placement, saliency balance, and shared alignment lines between design blocks — with plain OpenCV geometry. The three specimens below are not the same overlay: Jaws and Halloween show thirds + balance; The Exorcist shows the third measure, alignment, which only reads when separate credit/title blocks share a line.
thirds + balance
Jaws (1975)The shark-and-swimmer mass (amber box) centers almost exactly on a thirds intersection (amber dot vs. white centroid dot) — and the saliency-weighted "visual weight" (red dot) lands almost dead on true center (hollow ring). Two independent measures agree: this is a textbook-balanced shot.
thirds distance 0.15 · balance offset 0.04 — both near-best among the classics I spot-checked
thirds + balance
Halloween (1978)By rule-of-thirds this is the best-placed shot I checked — the knife-and-pumpkin mass sits right on an intersection. But its visual weight (red dot) is nowhere near center: the black void below empties two-thirds of the frame on purpose. Thirds-perfect and off-balance can both be true on the same poster — and here, that imbalance is the point.
thirds distance 0.14 · balance offset 0.25 — deliberately, not accidentally
alignment lines
The Exorcist (1973)Different overlay on purpose: alignment, not thirds. Nine separate design elements — tagline, credit blocks, the silhouette. The algorithm found shared lines: these two credit blocks, on opposite sides of the sheet, sit on the exact same horizontal (highlighted). A second, unmarked match: the tagline’s left edge lines up with the illustration’s edge.
alignment score 0.056 (lower = tighter grid) · 9 elements, 2+ confirmed shared lines
Grid & alignment via connected-component clustering of MSER text regions + gradient-magnitude contours (no LayoutParser/Detectron2 — see Sources). Balance via saliency-map centroid (OpenCV StaticSaliencySpectralResidual). Caveat: for full-bleed painted posters without one isolated figure, the "main visual mass" box can span nearly the whole sheet — thirds-distance then measures the whole composition's center of gravity, not one subject's placement. n = 33,622 (TMDB), full run. The Exorcist's alignment score is 0.056 on the copy shown and 0.097 in the dataset.
One more shape, measured a different way — not the angle of a line, but the width of the composition itself. Picture slicing a poster into 8 horizontal strips and asking, strip by strip: how much of this row is visually "busy" — edges, contrast, detail — versus empty space? Below, the amber/red bands are that busy width, one per strip. A wide band low on the poster and a thin or missing one up top reads as a pyramid; the reverse reads as an inverted funnel. Five real specimens — none repeated from elsewhere in this piece.
Web of Deception (1989)Four faces lined up shoulder to shoulder across the full width of the bottom third — above them, almost nothing but blank space. About as literal a pyramid base as the dataset has to offer.
pyramid shift +0.56 (wide base)
The Chameleon Killer (2003)The killer's robe narrows toward the head; the row of tombstone silhouettes along the bottom edge is the widest, busiest strip on the whole sheet.
pyramid shift +0.50 (wide base)
Vampire Bats (2005)The inverse: a swarm spans nearly the full width across the top third, fangs bared dead-center — then the entire bottom half of the sheet is flat, empty black. Nothing for the bands to measure down there at all.
pyramid shift -0.56 (wide top)
The Returning (1983)The top third is pure flat black — zero measurable detail. All of this poster's visual "weight" sits low: a wide wash of color and the title block, spanning most of the sheet's width.
pyramid shift +0.56 (wide base)
The Nothing (2011)Barn doors thrown open, five figures inside — busy, wide, and consistent across the upper five-eighths of the sheet. Below the doorframe: nothing, literally, all black.
pyramid shift -0.55 (wide top)
Bands: energy-weighted horizontal spread of gradient magnitude per 8 vertical bins, drawn at their real measured width and position. Picked for a clean, unambiguous signal — filtered out logo-only title cards and letterboxed video-still posters, which can fake the same number without a real compositional pyramid behind it. Across the full 33,622-poster dataset this measure averages out to roughly zero by decade — top-heavy and bottom-heavy compositions largely cancel out. It's a real per-poster signature, not a century-level trend (unlike diagonal composition — see Part V). Web of Deception (a non-English original) and Vampire Bats (a TV movie) are not in the 33,622-poster corpus. The Nothing reads −0.55 on the copy shown and −0.04 in the dataset.
Part VIII · Other Genres
Is It Just Horror?
Partly. Text shrank on sci-fi, mystery and thriller posters too — about a fifth of the sheet today in all four. Faces fell everywhere; horror just started lower and stays lowest. Black is the one thing horror kept for itself: 51% of the sheet in the 2020s, against 40–44% in the others.
OTHER GENRES — sci-fi (8,338), mystery (6,895) and thriller (17,646) posters measured with the same pipeline and the corpus's filters: current TMDB genre, English original language, released, no Animation, Music or TV Movie, and no Horror tag, so no film counts as both. A film can be in two of the three. Horror is the 33,622-poster corpus. Decade means, 1930–2026, decades with 100+ posters. Near-black share as in Part II, YuNet faces as in Part IV, MSER text-like coverage as in Part V.
Act II · Autopsy
Dissect any poster
Click any sheet in the century above — or search a title. The one-sheet takes the light table. Scroll and the same measurements I ran on all 33,622 posters peel off in layers: color, faces, lettering, creature, composition, pose.
Or start with a known nightmare
A Century of Fear, Getting Quieter
If you love horror films and love finding patterns in data, this is what I found: for a century, horror posters have been dropping their faces, cutting their text and sliding into black. Fear has been getting quieter ever since we learned to listen to the dark.
The posters don't lie. They tell you everything about what scared us, when, and why. All you have to do is know how to read them.
Sources & methods
Film list and poster images from The Movie Database (TMDB). The published corpus is 33,622 English-language horror features, already released, with a servable poster, a valid year (1900–2026), no adult flag, and without Animation, Music, or TV Movie. Existing exclusion lists in pipeline/data/excluded_*.csv still apply. The id list is pipeline/data/canonical_ids.txt (see docs/CORPUS.md).
For each poster I extracted a 5-color dominant palette (k-means in CIELAB, saturation-weighted), mean brightness (L*), mean saturation, share of near-black pixels, share of blood-red pixels, and hue-family distribution. Face counts: population statistics use YuNet detection at analysis resolution (floors — small, tilted or heavily stylized faces can escape); the counts shown on individually displayed specimens are hand-verified against the artwork. Chart notes spell out each measure where it appears.
Poster-design credits from IMPAwards, verified against each film's IMDb title.
Further reading on horror poster art: seven books, checked against publisher and library records.
Base dataset by Tanya Shapiro (TidyTuesday 2022). This product uses the TMDB API but is not endorsed or certified by TMDB. Industry context: Stephen Follows' Horror Movie Report.