Prompt Lab

AI人物画像のプロンプトを実験で検証するブログ。seed固定・条件統制で「本当に効くプロンプト」だけを報告します。

[Tested] 15 Ero-Horror Combo Recipes — Do the Best Elements From 5 Articles Stack?

[Tested] 15 Ero-Horror Combo Recipes — Do the Best Elements From 5 Articles Stack?

目次

TL;DR

  • Combinations stack by default. All 15 combos delivered their main elements simultaneously (zero images lost an entire element out of 45). Location, pose, and styling behave as independent axes.
  • But “dilution” happens. Busy scenes weaken small-detail anomalies: clouded eyes came out fainter in the flooded basement than in the solo test, and doll joints nearly vanished in front of the lit carousel.
  • And “amplification” happens too. Horror locations escalate styling on their own — subtle zombie makeup rendered noticeably heavier inside the tunnel and in front of the CRT than in the solo test.
  • Mirrors are combination amplifiers. Vampire × mirror gave us “her back to the camera, fangs and crimson eyes only in the glass” 3 out of 3; black tears × cracked mirror gave “only the reflection is crying.”
  • Best fear-allure crossover: clouded eyes × flooded basement, and soaked-sheer × night well. Pure fear: zombie × VHS. Pure bewitchment: fox mask × midnight torii corridor.

This is the compilation finale of our ero-horror series. For the individual element tests, see 24 locations, 18 mirrors, 16 poses, 15 inhuman stylings, and fear escalation.

Experiment Design

ItemValue
Modelz-image-turbo (6B, photorealistic distilled)
Steps / CFG / size8 / 1.0 / 1024x1024
Seeds42, 43, 44 fixed
Conditions15 combos × 3 images = 45 images
IngredientsOnly elements that scored 3/3 in the previous five articles

Crossover Recipes: Scary AND Alluring

Clouded eyes × flooded basement

c01
1girl, 32yo japanese actress, pale cloudy white eyes without visible pupils, white silk slip dress soaked from the knees down, standing knee-deep in murky water in a flooded dark concrete basement at night, water dripping from above, long black hair, looking at viewer
seed 42seed 43seed 44
Clouded-eyed woman standing in a flooded basementKnee-deep murky water and a sheer wet dressBasement with water dripping from the ceiling

Knee-deep murky water, drips from above, and a wet sheer dress landed in all three. The clouded eyes appeared 3/3 too — but fainter than in the solo test (fully pupil-less only on seed 43). The busier the scene, the more a fine-detail anomaly like the eyes gets watered down — the first discovery of this finale. Still one of the strongest crossover images in the set.

Drowned look × night well

c03
1girl, 32yo japanese actress, soaked waterlogged white dress clinging to her body, dripping wet long black hair, pale bluish skin, standing beside an old moss-covered stone well in a dark forest at night, moonlight through trees, looking at viewer
NSFW - クリックで表示
seed 42seed 43seed 44
Soaked woman standing by a moonlit wellHands on the well’s edge in the rainNight forest well and a clinging sheer dress

Night well, moon, and a fully soaked translucent dress in all three. On seed 43, unrequested rain began to fall, and she leans on the well’s edge. The sheerness ranks highest of all 45 images. The fear is quiet: nothing is happening — there’s just a woman standing by a well who has no reason to exist.

Torn dress × foggy mountain path

c04
1girl, 32yo japanese actress, torn tattered white slip dress, dirt smudges on her skin, long black hair, standing barefoot on a foggy mountain path at night, thick fog, cold bluish moonlight, looking at viewer
seed 42seed 43seed 44
Barefoot woman in a torn dress on a foggy pathMoon, fog, and a mud-smudged survivorMountain path in cold blue moonlight

Fog, moon, tears, dirt, bare feet — all three delivered, in naturally distant full-body compositions. One frame tells the whole story: something happened on that mountain, and she came back. The balanced recipe.

Clouded eyes × door-gap peek

c05
1girl, 32yo japanese actress, pale cloudy white eyes without visible pupils, peeking through the gap of a sliding paper door, only half of her body visible, white silk slip dress, long black hair, dim tatami room at night, moonlight, looking at viewer
seed 42seed 43seed 44
Clouded-eyed woman leaning out of a doorwayPale-blue eyes peeking past a screenWatching from deep in the shoji gap

We layered clouded eyes onto the pose test’s champion, the door-gap peek. Both landed in all three images — pose anomalies and eye anomalies don’t interfere. Seed 44, watching motionlessly from the dark beyond the shoji gap with a moon glowing through the paper, has the strongest quiet pressure.

Mirrors Are Combination Amplifiers

Vampire × candle × antique mirror

c11
1girl, 32yo japanese actress, vampire, small fangs, crimson eyes, pale skin, elegant, holding a lit candle in front of an antique mirror in a dark room at night, her reflection visible in the mirror, white silk slip dress, long black hair
seed 42seed 43seed 44
The vampire face appears only inside the gilded mirrorRaising a candle to look into the mirrorCrimson eyes and fangs in the glass

The masterpiece of this finale. In all three images, the subject shows us her back or profile while the crimson eyes and fangs appear only on the face in the glass. Candle positions disagree between subject and reflection (seeds 42, 43). The mirror’s self-breaking tendency from the mirror test combines with the vampire into an actual narrative: her true face only shows in the mirror.

Black tears × cracked mirror

c06
1girl, 32yo japanese actress, black tears streaming from her eyes, standing before a cracked full-length mirror, her reflection visible in the shattered glass, white silk slip dress, long black hair, dim abandoned room at night
seed 42seed 43seed 44
Tear streaks only on the face inside the cracked mirrorRadial cracks and two reflectionsThe reflection crying beyond the cracks

Beyond the radial shatter, the tear streaks are visible only on the reflected face — the subject’s own face ends up hidden by hair or turned away in all three images, so “only the me in the mirror is crying” composes itself. Seed 43 adds a second reflection at the right edge, free of charge.

Light zombie × dead CRT

c07
1girl, 32yo japanese actress, subtle zombie makeup, pale skin, sitting on the floor in front of an old CRT television that is turned off, her pale reflection visible on the dark screen, white silk slip dress, long black hair, dim room at night
seed 42seed 43seed 44
A different woman appears inside the CRT screenZombie-makeup woman sitting before a dead TVPale reflection and blood-smudged makeup

We layered zombie makeup onto the CRT — the “shortest route to Sadako” from the mirror test. In all three, the image in the glass reads less like a reflection and more like a separate woman living inside the screen (seed 42 looks outright broadcast). The TV-duplication glitch recurred too. And notice: the subtle zombie makeup came out much heavier than in the solo test — see the next section.

Discovery: Context Amplifies Styling

Light zombie × tunnel interior

c02
1girl, 32yo japanese actress, subtle zombie makeup, pale skin, faint dark veins on her neck, white silk slip dress, long black hair, standing inside an old abandoned tunnel lit by green fluorescent lights at night, damp concrete walls, looking at viewer
seed 42seed 43seed 44
Heavier zombie makeup inside the tunnelGreen fluorescents and blood-tinged eyesDarkness pooling deep in the tunnel

The prompt uses the exact same subtle zombie makeup as solo test z01 — yet inside the tunnel, dark red smudges and blood-tear streaks appear around the eyes (heavier than solo in 3 of 3). Makeup that stopped at “sick beauty” in the abandoned room escalates on its own in a scarier location. Location changes the intensity of styling — the styling-side version of the “context decides interpretation” law from the pose test.

Doll joints × abandoned amusement park (dilution example)

c10
1girl, 32yo japanese actress, looking like a broken ball-jointed doll, visible doll joint lines on her arms, white silk slip dress, long black hair, standing in front of an old carousel in an abandoned amusement park at night, carousel lights glowing, looking at viewer
seed 42seed 43seed 44
Standing before the lit carouselSmall joint hardware visible at the elbowsNight carousel and a woman in a white slip

The lit carousel is gorgeous in all three — but the doll joints that consumed entire arms in the solo test shrank to a bit of elbow hardware on seed 43. Flashy backgrounds eat fine-detail anomalies. Amplification (c02) and dilution (c01, c10) are two sides of the same coin: plain backgrounds to showcase detail anomalies, rich backgrounds when atmosphere is doing the scaring.

Still Fear: The Waiting Woman, the Upward Stare

Pale bluish skin × seiza × night shrine

c08
1girl, 32yo japanese actress, pale bluish skin, sitting seiza-style on the stone ground in front of an abandoned shrine at night, stone lanterns dimly glowing, white silk slip dress, long black hair, looking at viewer
seed 42seed 43seed 44
Sitting and waiting on the lantern-lit approachSeated on the shrine’s stone path at nightBlue-gray skin on the shrine approach

We combined the “lit-lantern night shrine” established in the location test’s fixed version with seiza and pale bluish skin. Lanterns glow in all three, with a seated waiting figure (strict seiza tends to slip). On seed 44 the skin is distinctly blue-gray: a woman the wrong color to be sitting at the end of that path.

Pale bluish skin × ceiling stare × hotel corridor

c09
1girl, 32yo japanese actress, pale bluish skin, standing rigidly in an old hotel corridor with stained red carpet, her head tilted back staring at the ceiling, white silk slip dress, long black hair, dim lights at night
seed 42seed 43seed 44
Staring at the ceiling in a stained-carpet corridorRigid upward stare with pale bluish skinBlue-graded corridor

Stained red carpet, symmetrical corridor, head thrown back — all three delivered. Seeds 43 and 44 show clearly bluish skin. “A woman alone in an empty hotel corridor, looking at something on the ceiling” — a complete one-scene horror.

Maxed-Out Recipes

Heavy zombie × VHS (maximum fear achieved)

c13
1girl, 32yo japanese actress, zombie, gray decayed skin, dark veins, cloudy eyes, torn dirty dress, disheveled long black hair, standing in a dim abandoned room at night, looking at viewer, old VHS footage, analog video noise, scan lines, washed-out colors
seed 42seed 43seed 44
VHS-textured zombie with vignette frameFilm noise and gray-white skinHollow-eyed zombie as VHS footage

The inhuman test’s heavy zombie plus the escalation test’s VHS texture. Film noise, washed-out color, and a dark vignette frame (seed 42) — the maximum fear reached anywhere in this 300-plus-image series. Almost no allure survives. Included as the marker for where the “pure horror” endpoint of ero-horror lies.

Fox mask × midnight torii corridor (maximum bewitchment)

c14
1girl, 32yo japanese actress, white fox mask tilted on the side of her head, white silk slip dress, long black hair, standing in a corridor of red torii gates at midnight, pitch dark night sky, dim stone lantern light, looking at viewer
seed 42seed 43seed 44
Fox mask in the midnight torii corridorFox mask on her head among the red gatesMask hiding her eyes, only red lips showing

The torii corridor that the location test taught us to force into night with at midnight phrasing — now with a fox mask. The midnight vermilion-and-darkness held 3/3 again. Seed 44 became the mask hiding only her eyes, red lips exposed below — the series’ peak of bewitchment. Use this recipe when you want not fear, but the allure of something that isn’t human.

Drowned look × night sea

c12
1girl, 32yo japanese actress, soaked waterlogged white dress clinging to her body, dripping wet long black hair, pale bluish skin, standing in shallow dark waves on a beach at midnight, black sea behind her, looking at viewer
NSFW - クリックで表示
seed 42seed 43seed 44
Soaked woman at the night waterlineDusk sea and a clinging sheer dressStanding against the black sea

Black waves, soaked sheer fabric, pale bluish skin — 3/3 (seed 43 leans dusk). Sheerness ties c03 for the series maximum. “The woman who walked into the sea is standing at the waterline again.” A still image that carries its own backstory.

All-In: The Series Finale

c15
1girl, 32yo japanese actress, pale cloudy white eyes without visible pupils, pale bluish skin, soaked white slip dress clinging to her body, dripping wet long black hair, crouching on all fours in shallow murky water in a flooded dark concrete basement at night, water dripping from above, looking at viewer
NSFW - クリックで表示
seed 42seed 43seed 44
Blue-skinned, white-eyed apparition crawling through murky waterCrouching woman with blank white eyesCrawling through the flooded basement in a sheer wet slip

Clouded eyes + pale bluish skin + soaked sheer fabric + crawl pose + flooded basement: every winning element in the series, stacked. All of it landed in all three images. The highlight: the crawl finally became an apparition — the same family of pose that turned into gravure in the moonlit tatami room of the pose test becomes something crawling through murky water in this context. Final confirmation of “the pose is the vessel; context is the contents.” One caution: seed 42’s skin came out aggressively blue — stacking pale-bluish skin AND clouded eyes can overshoot.

Summary: The Ero-Horror Recipe Table

GoalRecipeReference
Best crossoverClouded eyes × flooded basement / soaked-sheer × night wellc01, c03
StorytellingTorn dress × foggy path / drowned × night seac04, c12
Mirror anomaliesVampire × mirror / black tears × cracked mirror / zombie × CRTc06, c07, c11
Still fearPale skin × seiza × night shrine / ceiling stare × hotel corridorc08, c09
Pure fearHeavy zombie × VHSc13
Pure bewitchmentFox mask × midnight toriic14
EverythingClouded eyes + pale skin + soaked + crawl + flooded basementc15

The Three Laws of Combination

  1. Stacking is the default. Location, pose, and styling act independently; all 15 combos delivered their main elements together.
  2. Watch for dilution. Information-rich backgrounds weaken fine-detail anomalies (clouded eyes, doll joints). Keep the background plain when the detail is the point.
  3. Exploit amplification. Horror locations deepen makeup on their own, and mirrors invent subject-reflection mismatches for free. Choosing the right location adjusts your styling intensity without touching the prompt.

Results are for z-image-turbo (CFG=1.0, 8 steps); other models may differ.

[Tested] 24 Horror Locations — Bamboo Grove, Tunnels, Flooded Basement… Which Backdrop Is Scariest Without Killing the Allure?
[Tested] 24 Horror Locations — Bamboo Grove, Tunnels, Flooded Basement… Which Backdrop Is Scariest Without Killing the Allure?
[Tested] 15 Ghost & Zombie Styling Prompts — Which Ones Turn a Beauty Inhuman Without Breaking Her?
[Tested] 15 Ghost & Zombie Styling Prompts — Which Ones Turn a Beauty Inhuman Without Breaking Her?
[Verified] Image Generation Prompt Best Practices
[Verified] Image Generation Prompt Best Practices

The Setup Behind This Series

This series generated over 300 images. Without a local GPU, a cloud GPU gets tests of this scale done in realistic time.

PR RunPod クラウドGPUでAI画像生成 RunPodを始める →