클로드 오퍼스 5.5와 피시 크리에이티브 MCP 기반 실사 AI 비디오 연출 워크플로우

AI 크리에이터 압둘 샤쿠르(Abdul Shakoor, @abxxai)가 Claude Opus 5.5를 총괄 디렉터로 삼고 Fish Creative MCP를 연결해 실사급 일관성을 구현하는 AI 비디오 연출 파이프라인을 공개했습니다.

이미지 출처: X @abxxai

영상 생성 모델의 성능 향상으로 단순 영상 렌더링보다 '의도한 장면과 디테일을 정확히 제어하는 프롬프트 설계'가 핵심 병목이 되고 있습니다. 샤쿠르는 단순히 "시네마틱(cinematic)"이나 "사실적(realistic)"이라는 모호한 수식어를 피하고, Claude Opus 5.5에게 카메라 결함, 피부 질감, 불안정한 프레이밍, 광원 위치를 구체적으로 지정하도록 지시하는 3단계 레퍼런스 시트 구축 및 시댄스 2.5(Seedance 2.5) 비디오 프롬프트 워크플로우를 정리했습니다.

핵심 구조: Claude Opus 5.5 = 크리에이티브 디렉터

이 파이프라인의 핵심은 역할 분담입니다.

• 기획 및 디렉팅 (Claude Opus 5.5): 샷의 구도, 렌즈 스펙, 결함 요소, 인물 묘사, 환경 시트, 타임라인 비트 등 모든 프롬프트와 세부 명세를 설계합니다.
• 실행 및 생성 (영상 모델 & Fish Creative MCP): Opus 5.5가 생성한 정밀 명세에 따라 비디오 모델이 렌더링을 수행합니다. Claude 내부에서 Fish Creative MCP를 통해 전체 워크플로우가 직접 구동됩니다.

1단계: 캐릭터 레퍼런스 시트 (3x3 그리드)

영상 전반에서 인물의 얼굴과 신체 일관성을 영구적으로 고정하기 위해 한 번 생성해 재사용하는 캐릭터 레퍼런스 시트 프롬프트입니다.

Character reference sheet of the exact woman in the attached photo. Keep her identity 100% identical: same face, bone structure, freckles, lips, eyes and hair. Photoreal and documentary, not a render.

SUBJECT: woman, early 20s, sun-kissed light olive-tan skin. Dense freckles across her nose, cheeks and upper shoulders, with naturally flushed cheeks and a faint sunburn on the nose bridge. Warm hazel-amber eyes with long natural lashes. Full soft pink lips. Very long, damp, salt-wavy hair, mid-brown with honey-blonde sun streaks, parted slightly off-centre, strands sticking to her shoulders.

WARDROBE: cream linen sundress with a blue floral toile print, thin tie-straps knotted into bows on each shoulder, and a drawstring tie at the square neckline. Two thin gold chains: a short one with a tiny cross, and a longer one with a small oval coin medallion. Barefoot. No other jewellery.

LAYOUT (3x3 grid on a plain warm-grey wall, soft overcast daylight from the left, same light in every panel):
Row 1: full body front / full body 3/4 left / full body back.
Row 2: head-and-shoulders front / 3/4 right / pure profile left.
Row 3:
- extreme macro of her cheek and nose: every pore, freckle and fine peach fuzz, dried salt-water droplets, a stray hair
- macro of her eyes: hazel iris detail, wet lashes, tiny under-eye texture
- close-up of the necklaces on her collarbone, with tan lines from the dress straps

SKIN IS NEVER SMOOTH: visible pores, uneven tone, tiny blemishes, natural oil sheen on the T-zone, fine vellus hair. No retouching, no beauty filter, no plastic or waxy skin, no airbrushing.

Shot on a 50mm lens, f/4, natural colour, light film grain. No text, no labels, no watermark.

• 핵심 원칙: 피부를 절대로 매끄럽게(smooth) 처리하지 않도록 지시합니다. 모공, 고르지 않은 피부톤, T존의 유분기, 솜털을 명시하고 뷰티 필터나 에어브러싱을 금지합니다.

2단계: 로케이션 레퍼런스 시트 (2x3 그리드)

배경 환경만 독립적으로 제어하기 위해 메인 인물을 배제한 상태로 장소 시트를 생성합니다.

Location reference sheet of the exact beach club in the attached photo, environment only. NO main character in any panel.

PLACE: tropical Caribbean / Tulum-style beach club on a bright, overcast late afternoon. Fine powder-white sand with footprints and small shell fragments. Teal-turquoise terry towels and teal lounger cushions on low dark-wood daybeds. Large cream canvas umbrellas with dark wooden poles and ribs. Tall coconut palms leaning in the wind. A thatched-roof cabana bar in the background. Dense green tropical bushes along the back. Small, out-of-focus beachgoers on loungers, with no clear faces. The sea is a soft turquoise line beyond the sand. The sky is milky grey-white with soft diffuse light and no hard shadows.

LAYOUT (2x3 grid, same weather and same time of day in every panel):
1. Wide establishing shot from towel height, eye level ~40 cm off the sand, looking toward the umbrellas and palms.
2. Reverse angle: looking toward the sea, waterline, low surf.
3. Medium shot of the row of daybeds with teal cushions under cream umbrellas.
4. Low upward angle at the palm canopy against the overcast sky, fronds moving.
5. Ground detail: a teal towel crumpled on the sand, sand grains on the terry loops, a wet patch.
6. Detail of an umbrella underside and pole, canvas texture, slight wind ripple.

Photoreal and documentary, as if shot on an iPhone main camera. Natural colour, slight compression, light grain. No cinematic grade, no people in focus, no text, no watermark.

• 촬영 감각: 아이폰 메인 카메라로 직접 찍은 다큐멘터리 톤을 요구하며, 인위적인 시네마틱 색보정을 배제하고 자연스러운 압축감과 입자감을 부여합니다.

3단계: 첫 번째 프레임(Frame 1) 확정

캐릭터 시트와 로케이션 시트를 모두 첨부해 비디오의 시작 기준이 되는 1번 프레임을 생성합니다. 이 첫 프레임이 전체 영상의 구도, 포즈, 조명을 결정합니다.

A tightly cropped, slightly upward-tilted close-up photograph centers on a young Caucasian woman in her early twenties with striking, model-like facial features. She has long, damp, wavy dark brown hair that frames her face and falls over her shoulders, and she wears a muted animal-print swimsuit top with thin straps. Her complexion is fair, with a subtle natural blush on her cheeks, glossed lips, and distinctive dark eyebrows with feathery brows. Her upward gaze directly into the camera exhibits a gentle, relaxed expression, and her skin shows subsurface scattering from diffuse sunlight. The beach setting in the background features cream sand, lounge chairs draped with turquoise towels, a scattering of vacationers, and large white sun umbrellas, with tall palm trees swaying gently overhead. Lighting is soft and overcast, eliminating harsh shadows and bathing the scene in subtle, cool daylight. The dominant color palette includes sandy beige, deep brown, turquoise, white, and hints of green. It appears to be captured with a mobile phone camera in high resolution with a moderate wide-angle lens, producing slight perspective distortion at the edges and a crisp, shallow depth of field focused on the subject's face. The overall mood is casual, summery, and inviting, evoking the relaxed ambiance of a beach holiday. HEX VALUES: ["#332a2a", "#1b1417", "#b98a76", "#f2f8fb", "#dcbdb1", "#51423d", "#e4efc3", "#d1dfee", "#725b51", "#9e6f5e", "#cbc6bd", "#cca994", "#858079", "#a4a8a4"]

• HEX 색상값 명시: 장면의 톤 앤 매너가 튀지 않도록 주요 색상 팔레트의 16진수(HEX) 코드를 프롬프트 말미에 직접 주입해 일관성을 강제합니다.

4단계: 시댄스 2.5(Seedance 2.5) 비디오 프롬프트 설계

완성된 레퍼런스 이미지들을 바탕으로 시댄스 2.5 모델에 입력하는 22초 원테이크 영상 프롬프트 구조입니다.

=== REFERENCES ===
Image 1 → the exact first frame, with no settle-in. A woman sitting on a teal towel on white sand, filming herself with an iPhone front camera at arm's length. Her phone arm reaches toward the lens in the lower left foreground, her free arm reaches down toward the sand at lower right. Head tilted slightly, eyes on the lens, long wet hair falling over her right shoulder. Cream umbrellas at right, rows of teal-towelled loungers and beachgoers behind her, leaning coconut palms, milky overcast sky. Image 1 sets the framing, light, colour and camera texture for the entire video.
Image 2 → the woman. Face, bone structure, freckles, skin texture, eyes, lips, hair, body, dress and jewellery always match Image 2. Her cheek close-up matches the macro skin panels of Image 2.
Image 3 → the beach club environment only: white powder sand with footprints, teal-cushioned dark-wood daybeds in long rows, cream canvas umbrellas on wooden poles, leaning coconut palms, a thatched cabana bar with hanging lights, turquoise sea with low surf, dense green bushes, milky overcast sky.
Image 4 → the man. Face, eyes, mustache, hair, skin, body and clothes always match Image 4. His close-up matches the eye and mustache macro panels of Image 4.
Nothing about any face, hairstyle, mustache, eye colour, outfit or necklace ever changes during the video.

=== FORMAT ===
22 seconds. 1080p. 9:16 vertical. One continuous real-time take filmed on her phone. The only cut is the hard cut on the very last frame.

=== CHARACTERS ===
THE WOMAN → early 20s. Sun-kissed light olive-tan skin. Dense freckles across her nose, cheeks and shoulders, flushed cheeks, faint sunburn on the bridge of her nose. Hazel-green eyes, long natural lashes, full soft pink lips. Very long damp salt-wavy brown hair with honey-blonde streaks, parted slightly off-centre. Cream linen sundress with a blue floral print, thin bow-tied shoulder straps, gathered drawstring neckline, tiered skirt. Two thin gold necklaces: a short one with a tiny cross, a longer one with a small oval coin. Barefoot. Dried salt droplets and a few sand grains on her shoulders and arms. She films herself the whole time. Her phone arm always holds the phone. Her free hand does every gesture. The two hands never swap.
THE MAN → lanky and thin, late 20s. An enormous pale-blonde mushroom bowl-cut: heavy blunt bangs cut straight across his eyebrows, puffed wide at the sides like a helmet, frizzy flyaway strands at the edges. A thick waxed pale-blonde handlebar mustache with both ends curled into tight spirals. Giant round bulging pale-blue eyes with white visible all around the iris and fine red veins in the whites. He never blinks while on screen. Pale pink sunburned skin with visible pores, a red nose tip, freckles on his forehead, a long thin neck with a visible Adam's apple. Wrinkled beige textured-linen shirt with rolled sleeves, open over a plain white crew-neck tee, a thin silver chain, tortoiseshell sunglasses hooked on the shirt placket. He never touches the phone. He only enters by sliding into her selfie frame from behind her right shoulder.

=== EMOTIONAL CORE & SPEECH RULES ===
There are exactly seven spoken lines in the whole video: six from the woman, one from the man. Each quoted line is spoken exactly one time, word for word, in the order written. No line, word or phrase is ever repeated, echoed, restarted or said twice. Every camera move happens while its line is being spoken, perfectly in sync with the voice.

=== TIMELINE (22s) ===
BEAT 1 — FRAME 0 (0.0–0.3s): Frame 0.0 is exactly Image 1. Handheld drift begins.
BEAT 2 — THE HOOK (0.3–3.2s): Stands up, camera tilts and dim auto-exposure recovers. Line: "Everyone's posting these realistic AI videos now..."
BEAT 3 — CHEEK CLOSE-UP (3.2–5.2s): Physical move-in to 15cm. Autofocus hunts and locks. Line: "The skin."
BEAT 4 — SKY TILT (5.2–7.4s): Pull-back, tilt up to sky. Line: "The lighting."
BEAT 5 — WIDE-LENS SPIN (7.4–10.4s): Wide-view tap, 120-degree spin around beach club. Line: "The whole environment."
BEAT 6 — THE REVEAL (10.4–13.4s): Camera still. Line: "All of it is AI. Even me."
BEAT 7 — THE PROBLEM (13.4–17.2s): Wide-view tap back to selfie. Line: "And nobody... is teaching you the actual steps behind them."
BEAT 8 — THE INTERRUPTION (17.2–18.2s): Blonde mushroom hair enters frame.
BEAT 9 — HIS LINE (18.2–20.4s): Face slides in. Line: "But... I wiiill."
BEAT 10 — DEADPAN + HARD OUT (20.4–22.0s): Eye contact, laughter break, hard cut mid-laugh.

=== REALISM & AUDIO RULES ===
Faces are never smooth. Real iPhone front-camera sensor noise, handheld drift, soft overcast daylight with zero lighting pops, raw phone mic audio with natural handling noise and background ambience.

커뮤니티 피드백 및 실전 검증 포인트

공개 직후 커뮤니티에서는 프롬프트 엔지니어링의 정밀도에 대해 다양한 피드백과 검증 의견이 이어졌습니다.

• 프레이밍 결함의 역설: 사용자 @Life__Mastery는 완벽한 영화적 구도보다 카메라 결함과 어색한 프레이밍을 허용하는 것이 실사 체감도를 높인다고 평가했습니다.
• 영상 모델과 프롬프트의 관계: 크리에이터 @catmanyau는 "비디오 모델이 카메라 크루라면, 프롬프트는 모호한 아이디어를 모델이 실행할 수 있는 장면으로 구체화하는 감독의 촬영 목록(shot list)"이라고 분석했습니다.
• 연속성 사전 테스트 권장: 조나타 나시멘토(Jonatha Nascimento, @jhowolfx)는 전체 비디오를 렌더링하기 전에 얼굴, 의상, 소품 위치 등 3가지 핵심 평면에서 연속성 테스트를 선행할 것을 권장하며, 레퍼런스 시트가 존재하더라도 출력 결과의 지속적인 크로스체크가 필수적이라고 짚었습니다.

원문 출처

• Abdul Shakoor (@abxxai) X 스레드: https://x.com/abxxai/status/2111105993384022253
• Fish Creative MCP: @FishCreativeHQ