← 컷 프롬프트 해부도로 돌아가기
Hamlog · Character Sheet Bible

캐릭터 시트 바이블

AI 영상에서 같은 얼굴을 끝까지 지키는 이미지 시트(캐릭터 레퍼런스) 제작 공식. Higgsfield 장편 「Hell Grind」(칸 2026 마르셰 상영) 프로덕션 검증 기법과 힉스필드 공식 캐릭터 시트 워크플로를 하나의 규격으로 합쳤다.

시트는 캐릭터의 유일한 기억이다 — "예쁜 얼굴"이 아니라 "구별되는 얼굴"을 한 장에 잠가라.
제1장

대원칙 — 모델은 아무것도 기억하지 않는다

보이스 바이블과 같은 출발점이다. 영상 모델은 매 컷을 독립적으로 새로 생성한다. 컷 1의 얼굴을 컷 2가 기억할 방법은 없다. 유일한 기억 장치가 시트를 Element로 등록하고 모든 프롬프트에서 참조하는 것이다.

캐릭터 태그가 얼굴을 지키고, 프로필 문단이 목소리를 지킨다 — 이 문서는 그 태그에 들어갈 "얼굴의 원본"을 만드는 법이다.
제2장

시트 규격 — 3패널 합본, 그리고 머리 없는 전신

시트는 낱장 여러 장이 아니라 16:9 한 장 안에 3패널이다. Higgsfield 장편 프로덕션이 검증한 구성:

SHEET LAYOUT3-panel · one image · 16:9
┌─────────────┬─────────────┬──────────────────┐
│  전신 앞모습    │  전신 뒷모습    │  정면 대형          │
│  (머리 없음,    │  (뒷머리는     │  얼굴 클로즈업        │
│   HEADLESS)   │   보임)       │  (카메라 직시)       │
└─────────────┴─────────────┴──────────────────┘
— neutral grey seamless background, flat even lighting
— "Consistent across all three panels: [의상·체형·소품]" 문구 포함

왜 전신에 머리가 없나

미친 소리 같지만 이게 와이드 샷 얼굴 붕괴를 통째로 고친 핵이다. 와이드 샷을 생성할 때 모델은 시트에서 얼굴을 참조하는데, 전신 패널의 작고 흐릿한 얼굴을 참조해 버리면 뭉개진 얼굴이 복제된다. 전신의 머리를 지워버리면 모델이 얼굴을 가져올 곳이 대형 클로즈업 하나뿐이 된다. "시트 안에 얼굴은 정확히 하나" 규칙의 구현법이다.

제3장

얼굴 앵커 공식 — 예쁘게가 아니라 구별되게

시트 프롬프트의 캐릭터 묘사는 아래 슬롯을 순서대로 채운다. 모든 슬롯이 구체적이어야 한다 — "nice", "pretty" 같은 형용사는 슬롯을 채운 게 아니다.

나이 + 국적IDENTITY
나이대 + 국적/인종을 명시. 한국인이면 반드시 "Korean", 사극이면 시대까지. a Korean man in his late 50s, Joseon-era …
얼굴형 + 골격FACE
얼굴형·턱선·광대. 성인 캐릭터는 성숙한 골격을 명시해 동안화(babyface)를 막는다. round weathered face / oval face with a defined jawline, mature bone structure
눈 + 눈썹EYES
모양·색 + 반드시 캐치라이트. 홍채가 빛나는(glowing) 눈은 금지. natural catchlights in both eyes, muted iris color
헤어HAIR
색 + 길이 + 스타일 + 마감(finish) + 가르마. 마감이 리얼리즘을 판다. black hair in a low ponytail with curtain bangs, natural matte finish
구별 특징 2~3개DISTINCTIVE MARKS
위치가 특정되는 것: 점·흉터·안경·수염·보조개·눈썹 모양. 이게 캐릭터끼리 닮는 것을 막는 핵심 앵커다. a small mole under his left eye, a thin grey goatee, heavy eyebrows
체형BODY
체형 + 비율. 전신 패널 두 개가 이걸 잠근다. stocky build with broad shoulders / slender, balanced proportions
의상 + 소품WARDROBE
머리→발끝, 항목마다 색 + 소재 + 특이점 1개. 시그니처 컬러는 화면 라벨 역할도 한다. a jade-green silk robe with a gold crane crest, black leather boots

"beautiful / handsome / attractive"는 쓰지 마라. 모델마다 기본값 미인상으로 수렴해서 캐릭터끼리 전부 닮아버린다. 목표는 distinctive다.

제4장

시트는 일부러 심심하게 — 룩은 영상에, 피부는 진짜로

시트의 배경·조명·질감은 의도적으로 밋밋해야 한다. 뉴트럴 그레이 배경 + 플랫 라이트 + 무보정 피부.

REALISM MODULEphotoreal 시트 공용 · 복붙
visible fine skin texture with natural pores, fine lines,
subtle asymmetries, natural visible makeup with slightly uneven
blending rather than flawless coverage, slight natural sheen
rather than glossy retouched finish, matte-to-natural complexion,
no digital smoothing, no beauty filter, no AI-airbrushed look
// 애니·3D 스타일 시트에서는 이 모듈을 빼고 해당 렌더 스타일만 명시
캐치라이트 검사 — 동공에 작은 빛 반사가 없는 얼굴은 탈락이다. 죽은 얼굴로는 어떤 비디오 모델도 연기를 못 시킨다. "예쁜 얼굴"이 아니라 "믿기는 얼굴"을 골라라.
제5장

부정문의 함정 — 이미지 프롬프트는 목소리와 반대다

보이스 바이블의 절반은 네거티브 잠금("하지 마라")이었다. 이미지 프롬프트에서는 정확히 반대다. 이미지 모델은 언급된 것을 그린다 — "no hat"이라고 쓰면 모자가 나올 확률이 오히려 오른다.

제6장

변형 시트 — 상태마다 별도 자산

같은 캐릭터의 의상·상태 변화(평상복→드레스, 마른→비 맞은)는 변형마다 시트를 따로 만들어 별도 Element로 등록한다.

VARIANT ELEMENTS상태 분리 등록
@ella          // 베이스 — 평상복
@ella_dress    // 드레스 버전 — 베이스 시트를 참조로 걸고 의상만 변경 생성
@ella_wet      // 비 맞은 버전 — 동일
// 하나의 시트에 텍스트로 "드레스 버전으로"를 시키면 얼굴이 드리프트한다.
// 변형 시트는 베이스를 참조하므로 얼굴이 유지된다.
제7장

스트레스 테스트 — 락 전에 부숴봐라

시트가 마음에 든다고 바로 쓰지 마라. 등록(락) 전에 검증한다:

  1. 다른 포즈 · 다른 조명으로 3~5회 생성 (장편이면 10회). 전부 같은 사람으로 알아볼 수 있어야 통과.
  2. 혼자가 아니라 다른 자산 옆에서, 실제 씬 조명 아래서 테스트하라. 혼자 안정적인 캐릭터가 프레임을 공유하는 순간 깨지는 경우가 흔하다.
  3. 실패하면 문제는 모델이 아니라 묘사다 — 워딩을 고치고 재테스트.
통과 기준은 "알아본다"이지 "똑같다"가 아니다. 조명이 바뀌면 얼굴은 당연히 달라 보인다. 구별 특징(점·수염·눈썹·헤어 실루엣)이 살아있고 같은 사람으로 읽히면 통과다.
제8장

복붙용 템플릿 + 예시 2종

TEMPLATE3패널 시트 · 슬롯 채우기
A film character sheet: three studio photographs of the same
person arranged side by side on a neutral grey seamless
background, flat even studio lighting.
LEFT panel: full-body front view, HEADLESS — cropped above the
shoulders so no head is visible.
CENTER panel: full-body back view.
RIGHT panel: a large front-facing close-up portrait of the
face, looking straight into the camera, head level and not
tilted, both eyes level and clearly visible.
The character: [나이+국적], [얼굴형+골격], [눈+눈썹+캐치라이트],
[헤어], [구별 특징 2~3개]; wearing [의상: 색+소재+특이점, 머리→발끝];
[체형].
Consistent across all three panels: [의상 요약], [체형], [소품].
[REALISM MODULE — 제4장]. 16:9.
EXAMPLE 1사극 · 50대 임금
A film character sheet: three studio photographs of the same
person arranged side by side on a neutral grey seamless
background, flat even studio lighting. LEFT panel: full-body
front view, HEADLESS — cropped above the shoulders so no head
is visible. CENTER panel: full-body back view. RIGHT panel: a
large front-facing close-up portrait of his face, looking
straight into the camera, head level and not tilted, both eyes
level and clearly visible.
The character: a Korean man in his late 50s, Joseon-era king,
round weathered face with mature bone structure, heavy dark
eyebrows, a thin grey goatee, a small mole under his left eye,
deep laugh lines, natural catchlights in both eyes; grey-streaked
hair tied under a black ikseongwan royal hat; wearing a crimson
silk gonryongpo royal robe with a gold dragon crest on the chest,
a jade belt, white beoseon socks and black silk shoes; stocky
build with broad shoulders.
Consistent across all three panels: the crimson robe, jade belt,
black hat, stocky build.
Visible fine skin texture with natural pores, weathered sun
lines, matte-to-natural complexion, no digital smoothing,
no beauty filter. 16:9.
EXAMPLE 2현대 · 20대 여성 — 실증 완료
A film character sheet: three studio photographs of the same person
arranged side by side on a neutral grey seamless background, flat even
studio lighting. LEFT panel: full-body front view, HEADLESS — cropped
above the shoulders so no head is visible. CENTER panel: full-body back
view. RIGHT panel: a large front-facing close-up portrait of her face,
looking straight into the camera, head level and not tilted, both eyes
level and clearly visible.
The character: a Korean woman in her mid-20s, oval face with a defined
jawline, straight dark eyebrows, a small beauty mark near the right corner
of her mouth, monolid eyes with natural catchlights; black hair in a low
ponytail with curtain bangs; wearing an oversized ivory cable-knit sweater,
straight-leg light-wash jeans, white canvas sneakers, thin silver hoop
earrings, no bag; slender build with balanced proportions.
Consistent across all three panels: the ivory sweater, light-wash jeans,
white sneakers, low ponytail silhouette.
Visible fine skin texture with natural pores, natural visible makeup with
slightly uneven blending, slight natural sheen rather than glossy retouched
finish, no digital smoothing, no beauty filter, no AI-airbrushed look. 16:9.
EXAMPLE 2 프롬프트의 실제 생성 결과 — headless 전신 앞모습, 전신 뒷모습, 정면 얼굴 클로즈업 3패널 시트
위 프롬프트의 실제 생성 결과 (GPT Image 2 · 4K · 16:9, 2026-08-21). 체크 통과: 왼쪽 전신 머리 없음 · 클로즈업 정면 직시 + 캐치라이트 · 세 패널 의상 일치 · 구별 특징(입가 점) 재현. 이대로 Element 등록 가능.

배경 플레이트·소품 시트도 같은 정신이다 — 배경은 3/4 앵글 + 캐릭터 들어갈 공간 + 시간대 조명, 소품은 단일 오브젝트 중앙 + 재질·라벨·마모 상태까지. 재질 묘사가 없으면 컷마다 재질이 바뀐다.

제9장

나쁜 예 → 좋은 예

BAD — 형용사로 채운 슬롯 a beautiful young Korean woman with a pretty face, nice casual outfit
모델의 기본값 미인상으로 수렴한다. 다음 캐릭터도 같은 프롬프트를 쓰면 둘이 자매처럼 나온다.
GOOD — 위치 특정 앵커 a Korean woman in her mid-20s, oval face with a defined jawline, a small beauty mark near the right corner of her mouth, straight dark eyebrows, black low ponytail with curtain bangs
점 하나, 눈썹 모양 하나가 "이 사람"을 만든다.
BAD — 시트에 시네 룩을 구움 cinematic film still, 35mm anamorphic, moody teal-orange grade, dramatic rim light, character sheet
이 캐릭터는 모든 씬에 틸오렌지를 끌고 다니고, 새 조명에 반응하지 않는다.
GOOD — 심심한 시트 neutral grey seamless background, flat even studio lighting, raw unretouched skin with visible pores
룩은 비디오 프롬프트의 통일 스타일 문자열이 담당한다.
BAD — 텍스트로 변형 지시 @ella but wearing a blue ball gown
참조 이미지와 텍스트가 싸우고, 얼굴이 드리프트한다.
GOOD — 변형 시트 별도 생성 베이스 시트를 참조로 걸고 드레스 버전 시트를 새로 생성 → @ella_dress로 별도 등록 → 해당 컷은 @ella_dress만 참조.
제10장

사고 체크리스트 — 무너지는 일곱 가지 방식

F1와이드 샷에서 얼굴이 뭉개진다

모델이 시트의 작고 흐린 전신 얼굴을 참조했다.

전신 앞모습을 headless로 — 얼굴 출처를 대형 클로즈업 하나로 강제한다 (제2장).

F2캐릭터들이 서로 닮아간다

"beautiful/handsome"이 기본값 미인상으로 수렴시켰다.

형용사를 빼고 위치 특정 구별 특징 2~3개를 캐릭터마다 다르게 박는다 (제3장).

F3연기를 시켜도 얼굴이 죽어있다

동공에 캐치라이트가 없는 시트를 락했다.

캐치라이트 검사에서 탈락시키고 재생성 — 영상 단계에서는 못 고친다 (제4장).

F4얼굴이 점점 플라스틱이 된다

수정한다고 이미지를 모델에 통째로 여러 번 돌렸다.

포인트 변경은 편집 모델 + 마스킹 합성으로, 원본 텍스처를 지킨다 (제6장).

F5국적 불명의 얼굴이 나온다

"Korean"과 시대 디테일을 안 썼다.

국적 + 시대("Joseon-era") + 복식 이름(갓·관모·한복 종류)까지 명시 (제3장).

F6모든 씬이 같은 룩으로 오염된다

필름 그레인·시네 렌즈·그레이드를 시트에 구웠다.

시트는 심심하게, 룩은 비디오 프롬프트와 로케이션에 (제4장).

F7컷마다 의상 디테일이 바뀐다

의상을 "casual outfit"처럼 뭉뚱그렸거나, 변형을 텍스트로 시켰다.

항목마다 색+소재+특이점, "Consistent across all three panels" 문구, 변형은 별도 시트 (제3·6장).