Claude Opus 5.5와 Seedance 2.5로 만드는 극사실적 AI 비디오 워크플로우

압둘 샤쿠르(Abdul Shakoor, @abxxai)가 Claude Opus 5.5, GPT Image 2.5, Seedance 2.5를 결합하여 기존 AI 영상 특유의 부자연스러움을 걷어내고 실사 홈비디오 질감을 구현하는 실전 프로덕션 워크플로우와 프롬프트 전문을 공개했습니다.

이미지 출처: @abxxai (X)

많은 AI 영상 제작자가 '초현실적(photorealistic)', '8K 시네마틱' 같은 모호한 수식어를 입력해 인위적인 플라스틱 질감의 실패를 겪습니다. 반면 이번 워크플로우는 연출 총괄(Claude Opus 5.5), 마스터 캐릭터 시트 및 시작 프레임 고정(GPT Image 2.5), 비디오 렌더링(Seedance 2.5)으로 역할을 분리하여 인물 정체성과 카메라 물리 법칙을 정밀하게 제어하는 접근법을 취합니다.

1. Claude Opus 5.5의 총괄 연출(Director)과 카메라 물리 프롬프팅

전체 파이프라인에서 인간 작업자는 단 하나의 핵심 콘셉트 아이디어와 원하는 시각적 레퍼런스 스크린샷 몇 장만을 제공합니다. 구체적인 프롬프트 작성은 전적으로 Claude Opus 5.5가 감독(Director)으로서 전담합니다.

• 총괄 기획 위임: 캐릭터 설정 시트, 시작 프레임 구도, 카메라 무빙, 인물의 감정 변화선, 대사 타이밍, 현장 음향(사운드스케이프) 및 배경 음악까지 Opus 5.5가 한 번에 구조화합니다.
• 핵심 트릭 (카메라 오퍼레이터의 물리적 실체 명시): Opus 5.5에 단순히 "사실적인 영상을 만들어달라"고 요청하지 않는 것이 핵심입니다. 대신 **"누가 카메라를 쥐고 있는가, 그 사람이 어디에 앉아 있으며, 어떻게 몸을 움직이는가"**를 프롬프트에 구체적으로 지정해야 합니다.
• 물리적 오차의 사실성 구현: "내 친구가 뒷좌석에서 촬영하고 있다(my friend is filming from the back seat)"는 한 문장만으로도, 모델은 기계적인 카메라 무빙 대신 미세한 손떨림(micro shake), 반응 지연(late reactions), 초점 탐색(autofocus hunting), 창밖 빛에 따른 노출 출렁임(exposure shifts)을 자연스럽게 묘사해 AI 특유의 이질감을 지워냅니다.

2. GPT Image 2.5 기반 마스터 캐릭터 시트와 시작 프레임 생성

영상 클립 간 얼굴이나 의상이 변형되는 드리프트(drift)를 원천 차단하기 위해, Seedance 2.5 생성 전 GPT Image 2.5로 정체성 고정 시트와 첫 프레임 레퍼런스를 먼저 렌더링합니다.

마스터 캐릭터 시트는 16:9 화면에 전신 턴어라운드 패널과 다각도 표정 클로즈업, 장신구 세부 디테일을 배치하여 비디오 모델에 서술적·공간적 기준을 단일 이미지로 각인시킵니다.

마스터 캐릭터 시트 프롬프트 (GPT Image 2.5)

Create ONE master character + frame reference sheet for the woman in <<<image_1>>>. She must be the exact same person in every panel: same face, bone structure, brown eyes, freckles across nose and cheeks, natural full lips, long sun-streaked light-brown hair parted in the middle with caramel highlights and loose flyaways. Same thin gold chain with small cross pendant, brown cord bracelets on right wrist, faded cropped white baby tee with pale blue and coral sun-and-waves print, light-wash frayed denim cutoff shorts. Identity is locked, no drift between panels.

<<<image_2>>> and <<<image_3>>> are STYLE ONLY: copy their early-2000s compact camera home-video grade, not their person, not their text overlays.

LAYOUT: one wide 16:9 sheet on a warm off-white paper background, thin grey panel borders, small clean black labels under each panel in a simple sans-serif. Three rows.

ROW 1 — TURNAROUND (neutral studio, soft even daylight, plain light-grey backdrop, full body, relaxed standing A-pose, arms slightly away from body, bare feet or white sneakers):
FRONT / 3/4 LEFT / LEFT PROFILE / BACK / RIGHT PROFILE / 3/4 RIGHT
Clean true-to-life color here so wardrobe and proportions read clearly. Consistent height and scale across all six.

ROW 2 — FACE + DETAIL (tight head-and-shoulders close-ups, same soft daylight, real skin with pores, freckles, peach fuzz, slight sheen, no retouching):
NEUTRAL / SOFT PRIVATE SMILE / EXCITED DISBELIEF (brows up, eyes wide) / BIG GENUINE LAUGH (teeth showing, cheeks lifted) / PLAYFUL "I TOLD YOU" (head tilt, closed-mouth smile) / DETAIL PANEL split in four: eye macro, necklace and cross pendant, cord bracelets on wrist, tee print and denim fray.

REALISM EVERYWHERE: real skin texture, natural asymmetry, individual hair strands and flyaways, real fabric creases, correct hands with five fingers, no beauty filter, no plastic skin, no CGI look. Aspect ratio 16:9. No text, no watermark, no logos.

시작 프레임(Start Frame) 프롬프트 (GPT Image 2.5)

오래된 필름 룩은 단순 후반 필터가 아니라 Opus 5.5가 프롬프트 단계에서 35mm 똑딱이 카메라의 광학적 특성, 28mm 왜곡, 한낮의 직사광선 노출을 직접 기술하여 형성됩니다.

Use the uploaded character sheet as the strict identity reference. Keep her exact face, bone structure, brown eyes, light freckles across nose and cheeks, natural full lips, and long straight light-brown hair with sun-kissed caramel highlights parted in the middle. Keep the thin gold chain necklace with the small cross pendant.

SCENE: A candid photo taken from inside a parked silver sedan, looking out through the open rear passenger door. The door frame, the grey fabric headliner, the black door panel with its chrome handle, and part of the seat belt frame the shot on both sides. She stands just outside the open door on a rocky clifftop road above the ocean, both arms raised, one hand gripping the top of the door frame and the other resting on the roof edge, leaning slightly toward the camera. Relaxed, unposed, calm expression, looking straight into the lens.

WARDROBE: Fitted cropped white cotton baby tee with a faded, sun-washed vintage graphic print in pale blue and coral red, slightly worn and soft. Light-wash low-rise denim cutoff shorts with frayed hems. Thin brown leather cord bracelets on one wrist. Hair slightly messy and windblown, a few strands across her face.

BACKGROUND: Deep blue Mediterranean sea filling the frame behind her, sparkling sun glints on the water, a flat horizon line at chest height, pale hazy sky, dark rocks and sandy stones at the cliff edge below.

LIGHT: Harsh direct midday summer sun, bright and slightly overexposed. Sun-warmed skin with a natural sheen and slight glow on the cheeks and nose, soft shadows under the chin, a faint reflection of light off the white car interior.

CAMERA AND LOOK: Early 2000s point-and-shoot film photo, 35mm film, 28mm wide lens with slight barrel distortion at the edges, shot from chest height inside the car. Warm, faded, sun-bleached color grade, lifted blacks, low contrast, creamy highlights that gently blow out in the sky, slightly muted blues and greens. Natural 35mm grain throughout, subtle lens flare from the sun in the upper frame, genuine consumer compact camera flash look mixed with natural daylight. No bokeh, wide depth of field so both the car interior and background sea are recognizable.

3. Seedance 2.5 멀티모달 씬 렌더링과 카메라 제어 기법

준비된 마스터 시트(<<<image_1>>>)와 시작 프레임(<<<image_2>>>)을 멀티모달 레퍼런스로 주입한 후, Seedance 2.5에 시간대별 샷과 카메라 물리 제약을 설정합니다.

씬 1 프롬프트: 차량 조수석 10초 연속 롱테이크 (Seedance 2.5)

=== REFERENCES ===

<<<image_2>>> is the exact START FRAME and master reference for the scene, car, location, framing, light and film grade. Frame one must match it pixel for pixel: camera inside a parked white sedan looking out through the open front passenger door, grey fabric headliner along the top, grey seat belt and B-pillar at screen-left, the open door with black panel and chrome handle at screen-right, rocky dry clifftop, glittering deep-blue sea, flat horizon, hazy pale bright sky, harsh midday sun.

<<<image_1>>> is the IDENTITY LOCK for the woman. She must be this exact person in every frame: same face and bone structure, warm brown eyes, freckles across nose and cheeks, soft flush on cheeks, natural full lips, long sun-streaked light-brown hair with caramel highlights, middle part, loose flyaways. Jewelry: thin gold chain with a small cross pendant and a small round gold coin pendant; brown multi-strand cord bracelets on her left wrist. Wardrobe: faded cropped white baby tee with a vintage print (pale blue stripe band, coral-red sun disc with a small seagull, blue waves), light-wash frayed denim cutoff shorts. Use the sheet's expression panels (Soft Private Smile, Excited Disbelief, Big Genuine Laugh, Playful "I Told You") as her emotional range. Identity, hair, jewelry and wardrobe never change or morph.

=== CAMERA LOCK (MOST IMPORTANT) ===

One continuous 10-second take, 1080p. Zero cuts, zero camera switches, zero time jumps.

ONE camera angle and ONE framing for the entire clip, exactly the composition of <<<image_2>>>. The camera does NOT rotate, pan, tilt, zoom, dolly, reframe or change position at any moment. The headliner, B-pillar, seat belt, open door, horizon line and sea stay in the same place in the frame from first frame to last. Only the woman moves inside this fixed frame.

The camera is held by a friend sitting in the driver's side of the car, filming through the open passenger door on a 2003-era compact digital camera held steady with both hands. Very subtle natural human breathing micro-shake only: tiny sub-pixel drift that stays centered. No sudden pans, no zooms, no robotic motions.

=== 0.0-3.0s: STANDING, FIRST LINE ===

Frame one is identical to <<<image_2>>>. She hangs from the door frame and roof edge with both arms raised, leaning slightly toward the lens, calm direct gaze. Half a second of stillness, looking into the lens. Then a slow private smile starts at one corner of her mouth, like she has been dying to tell someone this.

A small audible breath, then she speaks, soft and warm, relaxed American English, slightly raspy, intimate, like telling her best friend a secret:

"I've been testing out the new Opus 5.5 model..."

(pronounce "Opus five point five")

On "testing out": eyebrows lift slightly, eyes brighten and widen a touch. On "Opus 5.5": a tiny proud head tilt, quick glance down and back to the lens. Eyes glossy and alive, one natural blink.

Sea breeze pushes loose strands across her face, the tee flutters, sun glitter shimmers on the water.

=== 3.0-5.0s: SITTING DOWN INTO THE FRONT SEAT (SAME FRAME) ===

Still smiling, she lowers her arms from the door frame, turns slightly and sits down sideways onto the front passenger seat, right in front of the lens, one hand on the seat edge for balance. She swings her legs in and settles, then turns her upper body toward the camera, smiling into the lens as the take finishes.

씬 2 프롬프트: 해안 도로 6단 비트 컷 시퀀스 및 오프닝 샷 (Seedance 2.5)

=== REFERENCES ===

<<<image_2>>> is the exact START FRAME: aerial view of the winding coastal cliff road, the older white 2000s sedan driving along it, dry golden hillside at left, green coastal scrub, sandy beach, rocks and rolling white surf at right, beach houses and palms on the far hills, sun low over the sea with a glittering sun path. Frame one matches it pixel for pixel.

<<<image_1>>> is the master reference for the car interior, both identities, wardrobe and interior light. Woman: long sun-streaked light-brown hair, freckles, brown eyes, thin gold chain with small cross and coin pendant, brown cord bracelets on her left wrist, white baby tee with a coral sun, seagull and blue waves print, seat belt across her chest, passenger seat. Man: short dark brown hair, tanned olive skin, short full beard and mustache, faded teal-blue palm-leaf print shirt open at the collar, light-wash jeans, silver watch, driving. Dark grey fabric headliner, small sunroof, black door panels, grey seats, simple older dashboard. The white sedan is the same car in every shot. Identities, wardrobe and car never change or morph.

=== FORMAT ===

10 seconds, 1080p, six shots cut on the drum hits of the music, like a real summer road-trip film shot by a friend with a drone and a handheld camera. Every shot has real camera life: natural sway, small corrections, organic motion, never CGI-smooth.

=== SHOT 1 — 0.0-2.0s: AERIAL FOLLOW ===

Starts exactly on <<<image_2>>>. The drone glides forward and slowly descends behind the white sedan as it cruises along the curve. Wheels turning, soft shadow sliding over cracked asphalt and the faded double yellow line. Waves roll in and break along the beach, the sun path shimmers, scrub sways in the sea wind. Slight real drone float and drift, a soft flare passing as the angle shifts.

4. 실전 제작 체크포인트와 조명 일관성 관리

실제 제작 파이프라인 적용 시 반드시 준수해야 할 핵심 기술 규준은 다음과 같습니다.

• 카메라 락과 비트 기반 컷 제어: 단일 씬 렌더링 시에는 카메라 앵글 전환이나 타임 점프를 엄격히 금지(Zero cuts, zero camera switches, zero time jumps)하여 정체성 드리프트를 차단하고, 다중 샷 컷 시퀀스는 프롬프트 내에 비트 단위 타임라인(Shot 1: 0.0-2.0s 등)을 명시해 모델이 음악 박자에 맞추어 컷을 전환하도록 설계합니다.
• 조명 일관성(Lighting Consistency) 통제: 압둘 샤쿠르는 답글을 통해 "조명 일관성 유지가 AI 비디오 제작에서 사실성을 유지하기 가장 어려운 부분 중 하나(lighting consistency is one of the hardest parts to keep realistic in AI videos)"라고 설명했습니다. 프롬프트 전반에 걸쳐 직사광선(harsh midday sun)과 그림자 각도, 필름 톤을 통일해야 여러 클립을 이어붙였을 때 시각적 이질감이 생기지 않습니다.
• 네거티브 제약(Negative Constraints) 명시: CGI 질감, 뷰티 필터, 플라스틱 피부, 인공적인 스무딩을 프롬프트 본문에서 완전히 배제하여 모델이 과도한 후보정을 적용하지 못하도록 물리적 피부 결(pores)과 피부 잔털(peach fuzz), 손떨림을 보존합니다.

원문 출처

이 가이드는 압둘 샤쿠르(@abxxai)가 공유한 제작 워크플로우 전문을 바탕으로 구성되었습니다.