[Tested] Can You Generate Kuchisake-onna? Mask On/Off Pairs and 3 Slit-Mouth Phrasings on the Same Seed

Key Takeaways

• slit mouth, mouth corners cut open to her ears was completely ignored in 3 out of 3 images. The mouth stays identical to the control’s normal lips.
• glasgow smile drew red scar lines on the cheeks in 3 out of 3 images. The closest output to Kuchisake-onna among the three main phrasings. However, an actually split-open mouth never appeared — the scars read as lines on the skin, like special-effects makeup.
• impossibly wide grinning mouth produced a cheerful laugh in 3 out of 3 images. Mouth width stays within normal anatomy. Zero horror.
• Mask on/off pairs work reliably on the same seed. The white surgical mask appeared in 3 out of 3 images, with eye area and composition mostly preserved.
• [Follow-up] deep lacerations at the corners of her mouth is the current best answer. It drew bloody wounds at the mouth corners in 3 out of 3 images, without glasgow smile’s side effects (forced smile, brighter scene).
• [Follow-up 2] Wound symmetry can be specified with on both cheeks + symmetrically (3 of 3). But a wound reaching the ears never appeared in any of the 21 wound-type outputs — mid-cheek is the practical ceiling.
• [Follow-up 3] snake-like mouth turns into “a woman holding an actual snake in her mouth” (3 of 3). The metaphor gets interpreted literally.
• [Follow-up 3] mouth line extending far toward her ears produced the series’ first ear-reaching mouth line (1 of 3). Specifying a mouth line rather than a wound or slit is what breaks the mid-cheek ceiling.

Second entry in our yokai series. See also the Hachishaku-sama edition and the Yuki-onna edition.

The Kuchisake-onna Motif

Kuchisake-onna is the urban legend of a masked woman who asks “Am I pretty?” — and when she removes the mask, her mouth is slit from ear to ear. As an image-generation subject, the motif breaks down into two technical elements:

1. Mask on/off differential — can you generate a matched pair (masked beauty / unmasked face) by toggling only the mask on the same seed?
2. The slit-mouth expression — which phrasing, if any, can produce an anatomically impossible mouth?

Element 2 deliberately asks for a broken anatomy, and it’s an open question how far a photorealism-focused model will comply. We compare three phrasings.

Experiment Setup

Base Prompt

Conditions

The follow-up group (C) digs further into wound vocabulary, following the main result that “describing it as a wound is the best move.” Same seeds 42, 43, 44.

The third round (D) aims for wounds running symmetrically from both mouth corners to the ears. Same seeds 42, 43, 44.

The fourth round (E) borrows animal metaphors — can a “snake-like mouth” widen the mouth’s structure itself? Same seeds 42, 43, 44.

Results by Condition

A00: Control (Unmasked)

seed 42	seed 43	seed 44

Observations: All three show a night residential street, a streetlight, and a trench coat with a smartphone-snapshot texture. Seeds 42 and 43 are frontal; seed 44 is a turned, over-the-shoulder composition. Note that suburban street rendered as a Western-style suburb in all three images, not a Japanese neighborhood.

A01: White Surgical Mask

seed 42	seed 43	seed 44

Observations: The white surgical mask appears in 3 out of 3 images. Seeds 42 and 43 keep nearly the same stance, composition, and eye area as the control — they work as a “same person, now masked” pair. Seed 44 flipped from the control’s turned pose to a frontal one and gained bangs, weakening the pair’s consistency. The mask’s reliability matches the surgical mask results in our mask type comparison.

B01: slit mouth (Direct)

seed 42	seed 43	seed 44

Observations: 0 out of 3. The mouth is essentially identical to the control — no slit, no wound, nothing. Aside from minor composition and expression drift, the output looks as if the keywords weren’t there.

B02: glasgow smile (As Scars)

seed 42	seed 43	seed 44

Observations: Red scar lines appear on the cheeks in 3 out of 3 images. Seed 43 has symmetric scar lines running toward the ears on both cheeks — the closest to Kuchisake-onna of the five main conditions. Seed 42 puts scars on the left cheek and neck; seed 44 scatters them across cheeks, chin, and neck. However, an actually split-open mouth appears in 0 out of 3 images — the scars stay on the skin’s surface. As side effects, all three faces smile (likely pulled by the word smile), and the scene shifted brighter, toward dusk.

B03: impossibly wide grin (Exaggeration)

seed 42	seed 43	seed 44

Observations: All three produced a wide-open laughing mouth — but the width never exceeds normal anatomy, resulting in cheerful, head-tilted snapshot photos. The “impossibly wide” nuance is not reflected, and the images don’t function as horror.

Follow-up: Digging into Wound Vocabulary

C01: deep lacerations

seed 42	seed 43	seed 44

Observations: Raw, bloody wounds appear at the mouth corners in 3 out of 3 images. Seed 42 — stitch-like wounds running from both mouth corners onto the cheeks with dripping blood — is the closest image to Kuchisake-onna in the entire series. None of the glasgow smile’s forced smiles appear; all three keep a neutral expression, and the dark night scene is preserved. The clinical term lacerations respected the position directive precisely, with no side effects observed.

C02: open bleeding wounds

seed 42	seed 43	seed 44

Observations: Bloody wounds again in 3 out of 3 images, but position control is looser than C01. At seed 43 the wounds run down to the jaw and neck; at seed 44 they drift from the right cheek toward the ear and neck — away from the specified “mouth corners to ears” path. There’s more blood than C01, raising the visceral horror, but C01 reproduces Kuchisake-onna’s symmetric mouth-corner iconography more reliably.

D01: Symmetric, Ear-to-Ear (Explicit)

seed 42	seed 43	seed 44

Observations: With on both cheeks + symmetrically made explicit, all 3 images produce left-right symmetric wounds (C01 had one-sided wounds in 2 of 3). But “all the way to her ears” appears in 0 of 3 — the wounds stop around mid-cheek. At seeds 42 and 43 the wound direction also drifts vertically, falling from below the eyes. Seed 44 — crack-like symmetric wounds radiating from around the mouth — is the best static horror image of the entire series.

D02: Adding the Direct Slit-Open Phrasing

seed 42	seed 43	seed 44

Observations: The words slit open get interpreted as “an open mouth”: all 3 images combine a wide-open laughing mouth (teeth visible) with symmetric cheek scar lines — B03’s forced-smile effect returning with wounds attached. The mouth itself splits in 0 of 3. Still, a scarred face laughing open-mouthed on a deserted midnight street works as the dynamic “after she removes the mask” cut.

E01: snake-like mouth (The Snake Metaphor)

seed 42	seed 43	seed 44

Observations: An unexpected interpretation. In 3 of 3 images the mouth’s shape is unchanged — instead, the woman holds an actual snake in her mouth. Seed 43 has a green snake; seed 44 holds one sideways. The simile was read not as shape similarity but as summoning the real thing. Useless for slit mouths, but a blank-faced woman holding a snake in her teeth is a horror image in its own right.

E02: Reptilian Mouth + Explicit Line Reach

seed 42	seed 43	seed 44

Observations: The “to the ears” reach that never appeared in 27 wound/slit images finally shows up here. At seed 42, thin lines extend horizontally from both corners of the open mouth all the way to the ears. Seed 44 also grows stitch-like black lines from both mouth corners across the cheeks — close to the Kuchisake-onna icon. Seed 43 misfires: reptilian becomes pointed elf ears and the mouth stays normal. All three open the mouth as a side effect. Specifying a “mouth line” rather than a wound (laceration) is the key that breaks the reach ceiling.

Cross-Condition Comparison (seed 43)

Unmasked	Mask	glasgow smile	wide grin

Even on the same seed, glasgow smile and wide grin change the expression to a smile and alter the background brightness. The most stable differential is the mask on/off pair.

Discussion

Anatomical breakage doesn’t come out when asked directly

The direct phrasing (B01) was ignored outright, and the exaggeration (B03) was rounded down to a normal laugh. On this photorealism-focused model, asking head-on for anatomically impossible faces simply doesn’t get through.

Specifying it as a wound is the current best move

The only main-group phrase that worked was glasgow smile, and it worked by converting “a slit mouth” into “scar marks on the cheeks” — a skin-surface depiction. If you want a Kuchisake-onna-style image, aim for wounds and special-effects makeup rather than mouth deformation.

The follow-up pushed this direction further: deep lacerations at the corners of her mouth achieved “bloody wounds + neutral expression + preserved night scene” in 3 out of 3 images, with neither of glasgow smile’s weaknesses (forced smile, brightened scene). This is the current best answer for Kuchisake-onna imagery. Choose lacerations (C01) for positional precision, or open bleeding wounds (C02) for more visceral bleeding.

Wounds have a ceiling; lines break it

Round 3 succeeded in making the wounds symmetric with explicit on both cheeks + symmetrically (3 of 3). But a wound actually reaching the ears never appeared across all 21 wound-type images — mid-cheek is the effective limit of wound reach.

Round 4’s E02 broke that ceiling by changing the framing from a wound to a facial feature: “mouth line extending far toward her ears” reached the ears in 1 of 3 images. Skin wounds and facial contour lines apparently pass through different constraints. Current picks: D01 for a quiet symmetric-wound face, D02 for the laughing Kuchisake-onna, E02 to attempt the true ear-to-ear line.

Metaphors materialize

snake-like mouth generated “a woman with a real snake in her mouth” in 3 of 3 images, not a snake-shaped mouth. “B like A” similes risk being read as summoning A itself rather than transferring A’s attributes. If you want to borrow an animal’s facial structure, specify concrete visual attributes (line position, length) instead of the metaphor.

For differentials, the mask is the right tool

Same-seed mask toggling kept high consistency in 2 out of 3 seeds. Combined with the scar approach, a two-stage set of “masked” → “unmasked + glasgow smile” is the practical way to build the reveal within what the model can currently do.

Note: The b03 “laughing woman” images feel strangely off when combined with the deserted midnight street — a bright expression that doesn’t belong. The horror prompt getting rounded into a smile lands the result in the uncanny valley by accident.

Summary

Related Articles

[Tested] Can You Generate a Hachishaku-sama-Style Tall Woman? Testing extremely tall, 8 feet tall, and Low Angle

[Experiment] Mask Type Comparison — Output Differences Across 5 Types and Overriding Outfit Side Effects

[Verified] Image Generation Prompt Best Practices