영상 속 사람·제품 바꾸기 (힉스필드 젠츄츠 재현) — 넘김 문서
작성 2026-11-16 · 엔진: 크리에이젠 경유 ByteDance Seedance 2.5 reference-to-video · 이 문서 하나로 공정·지시문·재료·호출값·결과를 모두 볼 수 있게 썼습니다. LLM 에게 그대로 넘겨도 됩니다.
0. 결론 (먼저 읽기)
- 힉스필드 젠츄츠(Genjutsu)와 같은 결과를 씨댄스 2.5 참고 영상 방식(r2v)으로 만들 수 있다. 힉스필드 공개 예시를 같은 재료로 다시 돌려 사람·동물·옷·차·제품 바꾸기가 모두 됐다.
- 성패를 가르는 건 지시문이다. 규칙 두 개:
- 길이 일치 — 지시문이 설명하는 길이 = 넣는 영상 길이 = 뽑는 길이. 20초 지시문 + 5초 영상이면 뒷장면을 5초 안으로 당겨 «없던 장면»을 만든다.
- 카메라·동선을 초 단위로 — «카메라가 따라간다» 한 줄이면 바짝 붙어 찍고 배경을 새로 그린다. 카메라 위치·거리·인물 크기·누가 언제 지나가는지를 0.25~0.5초 단위로 쓰면 원본 구도가 남는다(축구 예시로 확인).
- 사람 바꾸기 = 영상 + 사람 그림(전신·옆·앞 시트) + «원본 옷·머리 금지»까지 적은 지시문. 제품 바꾸기 = 영상 + 제품 그림 + «어디에 어떤 크기로, 무엇이 아니어야 하는지» 지시문. 실제 상표(코카콜라·레드불)가 결과에 나와도 막히지 않았다.
- 막히는 것: 유명 TV 광고 원본을 넣으면 입력 단계 «저작권 내용» 거절(들쭉날쭉하지만 세 광고 모두 걸림) 또는 결과 «안전 기준» 거절. 거절은 0크레딧. 해법 = 권리 있는 우리 영상, 또는 잘 나온 광고 장면을 본떠 우리가 만든 장면(GPT 그림 → 미니맥스 5초)을 원본으로 쓴다 → 3/3 성공.
- 힉스필드와 남는 차이: 힉스필드 «물건 바꾸기» 메뉴는 바꿀 부분만 새로 그리고 나머지는 원본 화소를 그대로 둔다(셋이 춤 예시 원본 유사도 0.92 vs 우리 0.34). 씨댄스는 화면 전체를 새로 그린다. 같은 효과를 내려면 «오려 붙이기(마스크 합성)» 단계를 따로 만들어야 한다(미구현·미확인).
1. 공정 (한 줄씩)
[재료] 원본 영상(3~30초, 소리 뺌) + 바꿀 대상 그림 1~10장(사람은 전신·옆·앞 3면 시트가 좋음 · 제품은 단색 배경 제품 사진) ↓ ① 자르기: 쓸 구간을 N초로 자른다(시험은 5초). ffmpeg -ss <시작> -t N -an -c:v libx264 -crf 18 ↓ ② 보기: 0.25~0.5초마다 장면을 뽑아(ffmpeg fps=4,tile) 컷 시각·카메라·인물 위치·누가 언제 무엇을 하는지 적는다 ↓ ③ 지시문 쓰기: §2 시스템 프롬프트(LLM) 또는 §3 틀. N초 안에 일어나는 일만, 초 단위로 ↓ ④ 호출: 크리에이젠 submissions → model fal-ai/bytedance/seedance-2.5/reference-to-video (§4 호출값) ↓ ⑤ 검수: 원본·결과를 같은 간격으로 뽑아 위아래로 놓고 본다 + 원본 유사도(SSIM) 숫자 ↓ (선택) ⑥ 오려 붙이기: 바꾼 대상 자리만 원본 위에 합성 — 힉스필드 «물건 바꾸기»와 같은 화면 보존 (미구현)
값: 480p 기준 크레딧 ≈ 20.6 × (결과 초 + 참고 영상 초 × 0.6). 5초 넣고 5초 뽑기 = 165크레딧, 21~24초 = 690~790크레딧. 걸리는 시간 5초판 약 3~8분. 거절(입력 저작권·결과 안전)은 0크레딧(크리에이젠 receipts 의 used_credit=0 확인).
힉스필드 내부에서 본 것(공개 JS·예시 페이지): 작업 종류 hf_mult_motion_control(동작 따라 하기) · hf_mult_replace_object(물건 바꾸기) · hf_mult_replace_object_flash · 입력 = 영상 1 + 그림 최대 10 + 지시문 · 결과 해상도·초당 24장이 씨댄스 규격과 같음 · 씨댄스 전처리 옵션 use_blur·use_eye_mask·use_eye_mask_sam 존재. 빈 지시문은 씨댄스가 거절하므로(«at least 1 character») 힉스필드는 지시문을 안에서 채우는 것으로 보임.
2. 시스템 프롬프트 — 5초 장면 설명을 받아 바꾸기 지시문을 쓰는 LLM 용
아래를 그대로 시스템 프롬프트로 쓰고, 사용자 메시지에 ① 장면 메모(0.25~0.5초 단위 관찰) ② 바꿀 대상(사람/제품)과 그림 설명 ③ 길이(초)를 넣는다.
You write prompts for ByteDance Seedance 2.5 reference-to-video that REPLACE one subject in a source video while keeping everything else identical. Inputs you get: (1) shot notes of the source clip with timestamps, (2) what to replace and a description of each reference image, (3) the clip length N seconds. Write ONE English prompt with these blocks, in this order: 1. "Source video: @Video1." then one line per reference: "@Image1 shows ..." (describe the reference concretely: shape, colors, logo text, outfit). 2. REPLACE block: "Replace ONLY <the source subject, described by what it looks like in the source> with <the subject from @ImageK>." Copy-list: for a person = face, hairstyle, hair color, skin, age, body type, height, proportions, headwear, full outfit and shoes; for a product = shape, size relative to hands/scene, colors, label, logo text, cap/lid. Then a NEGATIVE list of everything that must NOT carry over from the source subject (source clothes, numbers, hair, bottle color, label, cap...). If several people: "@Image1 = THE MAN, @Image2 = THE WOMAN", fix left/right roles by appearance, not only by side. 3. KEEP block: "Keep everything else in @Video1 exactly as it is: the same N seconds, the same shots and hard cuts at the same moments, the other people, camera moves, framing, lens, lighting, colors, on-screen text. Normal real-time speed, no slow motion, no new shots." 4. CAMERA AND FRAMING block (always): camera position and height, lens feel, how big the subject is in frame (e.g. full body ~2/3 of frame height), what is NOT visible (e.g. no stands, no sky). Add "Do NOT move the camera closer, no close-ups" when the source is wide. 5. SHOTS / WHAT HAPPENS block: a timeline that covers exactly 0.0–N s and nothing after N. One line per cut or action change, with times to 0.25 s, who enters/exits/passes and from which side, when the subject is picked up or shown. Never describe events that happen after N seconds in the original. 6. OTHER PEOPLE line: they stay exactly as in the source and never take the new look. Rules: plain concrete language, no style words, no story. Mention @ImageK where the subject appears. If the product should be a different size than the source object, say where and how big. Do not ask to remove watermarks or text unless the user wants it.
3. 지시문 틀 (사람 · 제품)
사람 바꾸기 틀 (축구 예시에서 구도까지 맞춘 판의 뼈대)
Source video: @Video1. Replace ONLY the <main player wearing number 10> with the person from @Image1. Copy that person's face, hairstyle, hair color, facial hair, skin tone, age, body type, height, body proportions, headwear and clothing exactly as in their photo, in every frame. ... CLOTHES AND HAIR COME ONLY FROM THE PHOTO, NEVER FROM THE SOURCE PLAYER: ... Keep the source video's motion, timing, camera moves, zooms, framing and lighting exactly. Normal real-time speed ... ONE continuous shot, no cuts. CAMERA AND FRAMING (very important ...): <camera position, lens, what is visible / not visible, subject size>. Do NOT move the camera closer ... WHAT HAPPENS (N seconds ...): - 0.0–1.0 s: ... - 1.0–1.6 s: ... WHO IS WHO: PERSON A = @Image1: ... PERSON A must NOT look like the source player: no <source outfit items>, no <source hair> ... OTHER PEOPLE stay exactly as in the source ...
제품 바꾸기 틀 (코카콜라·레드불·립스틱 성공판의 뼈대)
Source video: @Video1. Product reference: @Image1 shows one <product: shape, colors, logo text, cap>. Replace the <source product described by look> in @Video1 with the <product> from @Image1, everywhere it appears, at the same size. Copy its <shape, colors, label, logo, cap> exactly. No <source colors>, no <source cap>, no <source label>. Keep everything else in @Video1 exactly as it is: the same N seconds, the same <people/hands> and every motion, <background>, camera position, framing, lighting and colors. Normal speed, <one continuous shot / same cuts>, no new shots. WHAT HAPPENS / SHOTS (N seconds): - 0.0–1.0 s: ... THE PRODUCT stands ... in the left foreground. - 1.0–2.0 s: she grabs THE PRODUCT ... - 2.0–N s: she holds THE PRODUCT up beside her face, label facing the camera ...
4. 호출값 (크리에이젠 → 씨댄스 2.5 r2v)
크리에이젠 API(로그인 쿠키 · 조직 헤더)로 제출하고 receipts 를 4~6초마다 조회, 성공하면 public-result URL 을 받는다. 같은 submission_key 는 다시 과금하지 않는다(멱등). 영상·그림은 먼저 공개 URL 로 올린다(fal 저장소 · 기록 안 남김 설정).
{
"submission_key": "<sha1(gen_type, model, inputs)>",
"gen_type": "i2v",
"model_input": {
"gen_model": "fal-ai/bytedance/seedance-2.5/reference-to-video",
"gen_count": 1,
"inputs": {
"prompt": "<§2/§3 지시문 — @Video1 · @Image1.. 로 재료를 가리킴>",
"video_urls": [
"<원본 N초 mp4 · 소리 뺌>"
],
"image_urls": [
"<바꿀 대상 그림 1~10>"
],
"resolution": "480p",
"duration": "5",
"aspect_ratio": "16:9 | 9:16 | 4:3 | 3:4 | 1:1 | 21:9 중 원본에 가장 가까운 것",
"generate_audio": false,
"task": "reference",
"draft": false
}
}
}
duration은 문자열. 원본 길이와 같게(초 반올림). 빈 지시문은 거절 → 최소@Video1 @Image1.- 원본 소리는 뺀다(소리 때문에 저작권 거절된 일 있음).
- 힉스필드 지시문의
<<<image_1>>>·IMAGE 1·img 1표기는@Image1로,<<<video_1>>>은@Video1로 바꿔 넣었다. - 제품 그림은 크리에이젠
openai/gpt-image-2글→그림(quality high · 1024² · 장당 35크레딧)으로 만들었다. 우리 원본 장면은 같은 모델(1536×864) →fal-ai/minimax/h3-max-turbo/image-to-video(duration 5 · 약 25크레딧).
5. 시험 결과 한눈에
| # | 묶음 | 무엇을 | 길이 | 결과 | 크레딧 |
|---|---|---|---|---|---|
| A1 | 사람·동물 | 힉스필드 예시 3편(강아지 · 고양이 둘 → 개 둘 · 축구 호날두) 원본 길이 | 8/22/8초 | 셋 다 바뀜 · 축구 구도만 다름 | 1,250 |
| A2 | 사람·동물·옷·차 | 힉스필드 예시 10편 앞 5초 + 원래 20초 지시문 | 5초 | 10/10 바뀜 · 4편이 뒷장면을 당겨 «없던 장면» (설계 실수) | 1,650 |
| A3 | 사람·동물 | 그중 3편을 원본 전체 길이로 | 21/24/24초 | 원본 장면 순서 그대로 · «없던 장면» 0 | 2,260 |
| A4 | 사람·동물·옷·차 | 5초 영상 + 5초만 적은 지시문 5편 | 5초 | 4편 원본 그대로 · 축구 구도 빗나감 | 825 |
| A5 | 사람 | 축구를 카메라·동선 초 단위로 다시 | 5초 | 힉스필드와 거의 같음 | 165 |
| B1 | 제품 | 향수 광고 편집본 → 캔 / 햄버거 · 드리프트 차 → 쇼핑 카트 | 5초 | 캔·카트 됨 · 햄버거 결과 안전 거절 | 330 |
| B2 | 제품 | 실제 TV 광고 3편(토레타 → 코카콜라 · 스프라이트 → 립스틱 · 나랑드사이다 → 레드불 · 사람까지 바꾼 판 · 상표 없는 캔 판) | 5초 | 모두 거절(입력 저작권 / 결과 안전) | 0 |
| B3 | 제품 | 우리가 만든 광고 장면 3편 → 코카콜라 · 레드불 · 립스틱 | 5초 | 3/3 됨 | 495 (+장면 약 180) |
원본 유사도(SSIM, 앞 5초, 1 = 원본과 똑같음): 셋이 춤 — 힉스필드 0.92 / 우리 0.34(힉스필드는 가운데 한 명만 바꾸고 나머지 화소 보존). 그 밖 예시는 힉스필드·우리 모두 0.1~0.65 로 비슷(둘 다 화면 전체를 다시 그림).
6. 사람 바꾸기 예시
A1. 힉스필드 예시 3편, 원본 길이 그대로 (지시문 = 힉스필드 공개 지시문 변환)
각 비교 영상: 왼쪽 원본 · 가운데 힉스필드 결과 · 오른쪽 우리 씨댄스.
강아지 (지시문 없이 그림만 → 최소 지시문)

결과: 됨 · 265크레딧(8초). 힉스필드 지시문이 비어 있어 @Video1 @Image1 만 보냄.
보낸 지시문 원문 (15자)
@Video1 @Image1
고양이 둘 → 개 둘 (동물 둘 각각)

결과: 됨 · 720크레딧(22초).
보낸 지시문 원문 (81자)
Source video: @Video1. Replace 1 cat with @Image1 and the other cat with @Image2
축구 선수 한 명만 → 다른 사람

결과: 사람은 바뀜 · 구도가 더 가까움 · 원본 소리 때문에 처음 저작권 거절 → 소리 빼고 통과 · 265크레딧.
보낸 지시문 원문 (687자)
Repeat the entire supplied video @Video1 exactly. Replace only the jumping player originally wearing the black-and-white jersey with the man from @Image1: his face, ginger hair and beard, body and reference clothing. Keep the blue-shirted defender and all other people unchanged. Precisely reproduce the jump height, airborne trajectory, header, landing, limb poses and original slow-motion timing. Preserve correct overlaps with the defender and ball contact. Keep all original camera movements, framing, vertical orientation, stadium, lighting, cuts, soundtrack and full 8.12-second duration. Consistent replacement identity throughout, realistic anatomy, no duplicates or extra limbs.
A2. 힉스필드 예시 10편 — 앞 5초 + 원래 지시문 (설계 실수 포함 기록)
영상은 앞 5초만 넣고 지시문은 원본 전체(19~24초)를 설명하는 원문을 그대로 넣었다. 그래서 고양이·황실·축구·엘리베이터는 뒷장면을 5초 안으로 당겨 왔다. 바꾸기 자체는 10편 모두 됐다. 165크레딧 × 10. 비교 영상: 왼쪽 원본 5초 · 가운데 힉스필드 · 오른쪽 우리.
A2-trio · 셋이 춤 (물건 바꾸기 메뉴 · 사람 3) — 힉스필드 메뉴 hf_mult_replace_object

보낸 지시문(힉스필드 원문 변환) (682자)
Source video: @Video1. Make all three boys perform exactly the same dance choreography as the people in the original reference video. Preserve the original dance moves, body movements, rhythm, gestures, timing and synchronization. They must dance together at the same time, with natural and realistic movements. Keep their faces recognizable and consistent with their individual reference photos throughout the entire video. Do not add any accessories. They must keep the outfits shown in their respective reference photos. Make the result look realistic, smooth and entertaining. Do not change the original background, camera angle or choreography. Do not add any extra people.
A2-catpigeon · 정원 둘 (고양이·비둘기) — 힉스필드 메뉴 hf_mult_motion_control

보낸 지시문(힉스필드 원문 변환) (9907자)
[EDIT GOAL] Recreate @Video1 with exactly two new animal characters: the ginger Maine Coon cat from @Image1 and the grey pigeon from @Image2. CRITICAL PRIORITY: Reproduce the original movement choreography from @Video1 as accurately as possible. Preserve the original rhythm, movement mechanics, body gestures, timing, pauses, trajectories, character interactions and comedic performance. Change ONLY the characters and environment. Do not invent a new dance or reinterpret the original choreography. [REFERENCE ROLES] @Video1 — SOLE MOTION AND CHOREOGRAPHY MASTER. @Video1 defines the complete sequence of movements, body poses, dance rhythm, limb gestures, character positions, relative distances, forward movement, timing, speed changes, pauses, camera behavior and final action. Use ONLY its choreography and camera mechanics. Do not copy the original rats’ appearance, alley, concrete walls, rubbish, lighting or background. @Image1 — CHARACTER A: GINGER MAINE COON CAT. Use @Image1 as the strict identity reference for Character A. Preserve its recognizable appearance: large ginger Maine Coon, thick reddish-orange fur, long pointed ears with prominent lynx-like tufts, broad muzzle, long white whiskers, large fluffy paws, expressive amber-green eyes and distinctive serious facial expression. Maintain the same fur color, facial structure, proportions and recognizable identity throughout the entire video. Do not inherit the indoor background, furniture, lighting, pose or camera framing from @Image1. @Image2 — CHARACTER B: GREY PIGEON. Use @Image2 as the strict identity reference for Character B. Preserve its grey plumage, dark grey head, iridescent green-purple neck feathers, orange eyes, black wing markings, pink feet, realistic beak and natural pigeon anatomy. Do not inherit the dark studio background, lighting, pose or framing from @Image2. [CHARACTER MAPPING] Replace the foreground dancing rat from @Video1 with Character A, the ginger Maine Coon cat from @Image1. Replace the background dancing rat from @Video1 with Character B, the grey pigeon from @Image2. Each new character inherits ONLY the movement choreography of its corresponding original rat. Maintain the same foreground/background relationships, movement paths and relative character positions. Exactly ONE cat and ONE pigeon. [LOCATION AND ENVIRONMENT] A beautiful, realistic green lawn in a peaceful countryside garden. Fresh, dense, naturally growing green grass, subtle uneven ground, soft daylight, distant green bushes and slightly blurred trees. Both animals are physically present on the same lawn. Natural contact shadows beneath their feet and paws. Grass bends naturally under their weight. The environment remains completely consistent throughout the video. [VISUAL STYLE] ULTRA-PHOTOREALISTIC LIVE-ACTION VIDEO. The entire scene must look like a genuine viral smartphone video filmed spontaneously by a real person. Natural handheld smartphone footage, authentic daylight, realistic exposure, natural motion blur, subtle handheld camera shake and realistic autofocus. Highly detailed individual fur strands, realistic feather textures, physically accurate animal anatomy and natural lighting. No CGI appearance, no cartoon rendering, no stylized animation, no artificial cinematic effects. The animals must look completely real, even while performing the unusual dance choreography. [MOTION INHERITANCE — ABSOLUTE PRIORITY] Reproduce the complete dance choreography from @Video1. Every major movement must correspond to the original reference video. Preserve: * The exact order of body movements. * The original alternating stepping rhythm. * The original side-to-side body swaying. * The original forward and backward weight transfers. * The original rhythmic bouncing. * The original torso rotations. * The original alternating front-limb gestures. * The original movement acceleration and deceleration. * The original comedic pauses. * The original forward approach toward the camera. * The original ending movement. CHARACTER A — MAINE COON: The cat stands upright on its hind legs, just as the original foreground rat does. Its hind paws reproduce the same alternating dance steps. Its front paws reproduce the original rat’s arm gestures, including the same raising, lowering, swinging and rhythmic movements. Its shoulders, torso, hips and head follow the original rat’s body movement pattern. The cat maintains the same dance tempo, body swaying and forward momentum. Its thick fur, whiskers and ear tufts react naturally to movement and acceleration. Preserve the cat’s recognizable serious facial expression while allowing natural head movement and eye tracking. CHARACTER B — PIGEON: The pigeon stands upright on its two natural legs. Its feet reproduce the original background rat’s stepping rhythm and movement trajectory. Its body follows the same rhythmic swaying, bouncing and directional changes. Its wings reproduce the original rat’s front-limb gestures as closely as anatomically possible. The pigeon rhythmically raises, lowers and moves its wings in synchronization with the source choreography. The pigeon must remain grounded throughout the entire dance. No flying, hovering or unnecessary wing flapping. IMPORTANT: Adapt movements only where required by the animals’ natural anatomy. Preserve the recognizable choreography, rhythm and comedic synchronization. [CAMERA AND FRAMING] Use @Video1 as the strict camera movement reference. Preserve the original camera perspective, shot progression, framing changes, relative character scale and camera timing. The camera behaves like a real handheld smartphone recording. Maintain the original spatial relationship between the foreground and background characters. As the cat approaches the camera, reproduce the same increase in apparent size and the same energetic close-up effect as the original foreground rat. Use only natural handheld movement. Do not introduce new camera angles, cinematic orbit shots, dramatic zooms, slow motion, artificial transitions or additional cuts. [SCENE STRUCTURE] STAGE 1 — OPENING MOVEMENT Both animals begin in the same relative positions and movement phase as the original two rats in @Video1. Immediately reproduce the opening rhythmic stepping, upright body postures and alternating limb gestures. The cat performs the foreground rat’s opening choreography. The pigeon simultaneously performs the background rat’s opening choreography. Maintain the same movement rhythm and initial spacing. END STATE: Both animals continue the original dance sequence in their corresponding positions. STAGE 2 — SYNCHRONIZED DANCE Continue the choreography directly from the previous stage without resetting the movements. The cat reproduces the original foreground rat’s rhythmic side-to-side movements, alternating paw gestures, body sways and stepping patterns. The pigeon reproduces the original background rat’s corresponding movements using its feet, body and wings. Maintain the same synchronization, movement speed and playful body dynamics. Do not add new dance steps. END STATE: Both animals maintain the original movement rhythm and relative positions. STAGE 3 — APPROACH TO CAMERA Reproduce the original foreground rat’s movement toward the camera using the Maine Coon cat. Maintain the same forward stepping rhythm, body swaying, front-paw gestures and acceleration. The pigeon continues the original background rat’s choreography. Preserve the original camera relationship and character spacing. The cat gradually occupies more of the frame as it approaches the lens. END STATE: The cat reaches the same relative foreground position as the original rat. STAGE 4 — FINAL CLOSE-UP Reproduce the original ending choreography from @Video1. The Maine Coon moves close to the camera, matching the foreground rat’s final position, body orientation, movement rhythm and front-limb gestures. The pigeon remains behind it, continuing the corresponding background choreography. Preserve the same spontaneous, funny and energetic ending. END STATE: The cat is close to the lens in the same final movement phase as the foreground rat, with the pigeon visible behind it. [SOUND DESIGN] Natural outdoor garden ambience. Soft wind, distant birds, subtle grass rustling and realistic paw and foot contact with the lawn. No dialogue. No added music. No cartoon sound effects. [PHYSICAL REALISM] Both animals must retain their natural recognizable anatomy. The Maine Coon has four anatomically correct legs and paws, with its front paws used as dancing arms while standing on its hind legs. The pigeon has exactly two legs and two wings. No human hands, human arms or human body proportions. Realistic balance adjustments, body weight, natural inertia, realistic foot placement, fur dynamics and feather movement. All movements must appear physically continuous. [CONSISTENCY LOCKS] Exactly one Maine Coon cat from @Image1. Exactly one grey pigeon from @Image2. No additional animals. No duplicated limbs. No identity swapping. No character morphing. No environment changes. No teleportation. No random actions. No invented choreography. No changes to the original dance rhythm. No subtitles, logos or text. [FINAL PRIORITY RULE] @Video1 controls ALL movement choreography, timing, rhythm, body gestures, action order and camera mechanics. @Image1 controls ONLY the Maine Coon’s visual identity. @Image2 controls ONLY the pigeon’s visual identity. The new green lawn controls ONLY the environment. The result must look like a genuine real-life viral smartphone recording of a Maine Coon cat and a grey pigeon performing the exact same recognizable dance choreography as the two original rats in @Video1. Movement fidelity is more important than cinematic beauty. Photorealism is mandatory.
A2-cat · 우주 고양이 — 힉스필드 메뉴 hf_mult_motion_control

보낸 지시문(힉스필드 원문 변환) (3708자)
Source video: @Video1. Replace the cat in the source video with the cat from the reference image. The reference image shows one cat: copy that cat's breed, head shape, face, ear shape, fur length, fur color, markings, face mask, tail and eye color exactly as in its photo, in every shot. Keep its breed, fur color and markings exactly as in the photo. Do not invent or change its look. The pose and the room in the photo do not matter: take the pose and motion only from the video, and the backgrounds only from the video. Keep the source video's motion, timing, camera moves, zooms, framing, lens, cuts, dissolves and lighting exactly. Normal real-time speed, exactly matching the source video timing, no slow motion, no speed-up, no time compression. THE VIDEO HAS 4 SHOTS. Keep every cut and dissolve exactly there, add no new cuts: - SHOT 1 (0.00–5.875 s): THE CAT stands upright on its hind legs in deep blue starry space above a glowing blue planet horizon, front paws held in front of its chest, looking at the camera; the camera pulls back, the cat starts to dance; from about 4.8 s the stars turn into a white hyperspace warp and the cat floats in the middle with its legs spread and front paws raised. - SHOT 2 (5.875–13.29 s, hard cut): warm orange glow with large floating soap bubbles and glass cubes; THE CAT dances upright on its hind legs, raises both front paws high and wide, sways, turns to the side and stretches one front paw forward; the camera pushes in from a full-body shot to a side medium shot. - SHOT 3 (13.29–15.5 s, hard cut): close-up of THE CAT's head and chest on a vivid purple backdrop with floating violet glass cubes; it looks at the camera, sings, raises a front paw at the left edge, turns its head to the right. - 15.5–15.8 s: dissolve into SHOT 4. - SHOT 4 (15.8–21.04 s): close-up of THE CAT's head and chest on a shiny gold backdrop with glowing golden glass blocks and sparkles; it turns its head, sings, raises a front paw in the lower foreground, then calmly looks into the camera and sings again. SINGING: THE CAT sings the song like a singing animal in a music video: whenever the source cat opens its mouth, the cat's lower jaw visibly drops open and snaps shut on every sung syllable, EXACTLY in time — about 13.9–15.3 s, 17.1–17.4 s, 17.9–18.1 s, 18.6–19.6 s and 20.4–21.0 s. Real cat mouth, small pink tongue inside, no panting, no human lips. When the source cat's mouth is closed, the mouth stays closed. THE CAT must NOT look like the cat in the source video: no grey-and-white bicolor British longhair coat, no solid grey cap and back with a white face blaze, no white chest bib and white belly, no round grey-and-white face with copper-orange eyes — unless the same thing is visible in its own reference photo. The cat stands and moves on its hind legs exactly like the source cat. Real cat legs ending in round fluffy cat paws with toe beans and fur — NO fingers, no thumb, no human hands; when it raises a paw, it lifts the whole round paw, toes together, no finger pointing up. Belly and groin covered with dense fur, anatomically neutral like a plush toy, no nipples, no genitals — also when it dances with its legs spread. Realistic real cat, natural fur only, no clothes, no collar, no bow, no accessories. BACKGROUNDS (stay exactly as in the source, nothing redesigned, never the room from the photo): starry space with the blue planet horizon and the hyperspace warp, the orange bubbles and glass cubes, the purple backdrop with violet cubes, the gold backdrop with golden blocks. NO OTHER ANIMALS OR PEOPLE: this video contains only this one cat. No text, no subtitles, no Chinese characters, no watermark. Photorealistic music video.
A2-money · 돈 흔드는 사람 — 힉스필드 메뉴 hf_mult_motion_control

보낸 지시문(힉스필드 원문 변환) (15자)
@Video1 @Image1
A2-car · 드리프트 차 — 힉스필드 메뉴 hf_mult_motion_control

보낸 지시문(힉스필드 원문 변환) (23자)
@Video1 @Image1 @Image2
A2-suit · 정장 인물 — 힉스필드 메뉴 hf_mult_replace_object_flash

보낸 지시문(힉스필드 원문 변환) (1095자)
Source video: @Video1. Replace the face of the main male character in the source video with my face from @Image1. The attached collage shows the same person from multiple angles; use all views to accurately preserve my identity. Match my facial proportions, eyes, nose, lips, jawline, skin tone, and distinctive facial features. My face must remain recognizable and consistent throughout the entire video, including during head turns, speech, and fast movements. Preserve the original character’s body, clothing, actions, head movements, gaze direction, facial expressions, and lip-sync timing. Keep the original camera movement, framing, background, other people, editing, and audio unchanged. Use Image 1 only as a facial identity reference; do not copy the suit or background from the photos. Blend my face naturally into the scene with matching lighting, shadows, perspective, and motion blur. Preserve realistic skin texture and natural occlusions when objects or hands pass in front of the face. No identity drift, flickering, warped features, beauty filters, or visible face-swap seams.
A2-dress · 웨딩드레스(옷만) — 힉스필드 메뉴 hf_mult_replace_object_flash

보낸 지시문(힉스필드 원문 변환) (15자)
@Video1 @Image1
A2-elevator · 엘리베이터 두 사람 — 힉스필드 메뉴 hf_mult_motion_control

보낸 지시문(힉스필드 원문 변환) (3667자)
Source video: @Video1. IDENTITY / OUTFIT LOCK @Image1 = THE MAN. @Image2 = THE WOMAN. THE MAN is exactly the character from @Image1 in every frame: same face, hair, skin, body, height, proportions, full outfit and accessories, including his headphones from @Image1. His complete look from @Image1 stays identical from the first frame to the last. THE WOMAN is exactly the character from @Image2 in every frame: same face, same hairstyle and hair color, same body shape, height and proportions, same outfit as in @Image2. Her face, hair and body stay identical from the first frame to the last. Both characters keep the render style of their own reference images. ROLES AND BLOCKING (two people only, the whole video) 0.0–19.9 s, medium two-shot inside the elevator: THE MAN stands on the left side of the frame, THE WOMAN stands on the right side of the frame, both framed from the chest up, side by side. - 0.0–2.0 s: THE MAN wears his @Image1 headphones on his head and looks ahead; THE WOMAN glances at him and smiles. - 2.0–3.5 s: THE MAN raises his hand and pulls his @Image1 headphones off his head down around his neck. - 3.5–19.9 s: THE MAN's headphones rest around his neck. The two turn to each other and talk; THE WOMAN speaks to him in profile (about 3–11 s), then turns toward the camera, smiles, looks up and sings (about 12–18 s); THE MAN listens in profile and smiles. 19.9 s: hard cut to a wide frontal shot of the closed elevator doors. - 19.9–21.0 s: the doors slide open; THE MAN stands on the left inside the elevator, THE WOMAN stands next to him on the right, both seen full body. - 21.0–22.5 s: THE WOMAN walks forward out of the elevator and exits the frame on the right. - 22.5–24.1 s: THE MAN stands alone in the elevator, full body, headphones around his neck, looking ahead, and speaks a short line to himself: his lips clearly move in sync with his voice in the original audio, even though his face is small in this wide shot. LIP SYNC (TOP PRIORITY) Lip sync is the most important part of this video. Both characters keep the original lip sync: THE MAN's and THE WOMAN's mouths open, close and shape every word exactly in sync with the speech and singing of the original audio, frame by frame, with the same timing as the original performers. When the original mouth is closed, the character's mouth is closed. Facial expressions, smiles, eye direction and head turns follow the original. FINAL LINE: at the end of the video (22.5–24.1 s), alone in the wide shot, THE MAN must visibly say his line: his mouth opens and closes on every syllable, exactly in sync with his voice in the original audio. His mouth is clearly visible and clearly moving while he speaks. LOCATION Same elevator as the input: warm dark-brown panelled walls with brushed-steel vertical strips, soft warm overhead light; in the wide shot, brushed-steel sliding doors and a metal handrail on the back wall. Same lighting and color as the input. HARD CHECKS - Exactly two people in every frame where people are visible: THE MAN (@Image1) and THE WOMAN (@Image2); in the last 1.6 s only THE MAN. - THE MAN looks exactly like @Image1 in every frame, including both close shot and wide shot; his @Image1 headphones are on his head 0–2 s and around his neck from 3.5 s to the end. - THE WOMAN looks exactly like @Image2 in every frame, including her walk-out in the wide shot: same face, hairstyle and body. - Lip movements of both characters match the original audio, including THE MAN's final spoken line alone in the wide shot (22.5–24.1 s). - One hard cut at 19.9 s, no other cuts, no transitions, no morphing. Keep motion, camera, timing, background unchanged.
A2-imperial · 트럭 앞 인물 — 힉스필드 메뉴 hf_mult_replace_object_flash

보낸 지시문(힉스필드 원문 변환) (576자)
Source video: @Video1. Use the uploaded video as the strict reference for the complete action, timing, dance movements, gestures, camera movement, framing, location, cars and sequence of events. Replace the first person with the character from @Image1 , the person filming in selfie mode with the character from @Image2 , and the person appearing between the cars with the character from @Image3 . Preserve the recognizable face, hairstyle and identity of each character throughout the entire video. Each image controls only its assigned person. Do not mix or swap identities.
A2-athlete · 축구 선수 한 명 — 힉스필드 메뉴 hf_mult_motion_control

보낸 지시문(힉스필드 원문 변환) (3262자)
Source video: @Video1. Replace ONLY the main player (the one wearing number 10 in the source) with the person from @Image1. Copy that person's face, hairstyle, hair color, facial hair, skin tone, age, body type, height, body proportions, headwear and clothing exactly as in their photo, in every frame. If the person in the photo is short, old, bearded or wears a hat, keep exactly that: same height and proportions, same hat on the head the whole time, also while sprinting, falling and rolling on the grass. Do not invent or change their look. Everybody else stays as in the source. CLOTHES AND HAIR COME ONLY FROM THE PHOTO, NEVER FROM THE SOURCE PLAYER: the person wears exactly the outfit, socks and shoes from @Image1 from the first to the last frame, even when seen from behind, in motion blur, lying on the ground and in the close-up at the end. Keep the source video's motion, timing, camera moves, zooms, framing and lighting exactly. Normal real-time speed, exactly matching the source video timing, no slow motion, no speed-up, no time compression. ONE continuous shot, no cuts. WHAT HAPPENS (one continuous handheld broadcast shot on a sunny football pitch): - 0.0–1.5 s: PERSON A stands on the grass with the back to the camera, then starts to move. - 1.5–8.5 s: PERSON A sprints forward across the pitch past opponents, then dribbles one football with the feet; the camera pans and follows. - 8.5–11.0 s: an opponent in a white-and-sky-blue striped shirt runs into PERSON A from the left, they collide and BOTH fall and roll on the grass. - 11.5–16.0 s: PERSON A sits on the grass with legs bent, arms out to the sides, looking around and gesturing with both hands, complaining. - 16.0–20.5 s: a tall teammate (navy blue shirt with number 9, white shorts, red socks) comes in from the left, leans over and pulls PERSON A up by the arm; for a moment the teammate's back covers PERSON A. - 20.5–23.6 s: close-up, PERSON A stands up bent forward, hands near the knees, grimacing and then laughing, looking around. WHO IS WHO: - PERSON A = @Image1: the main player in the centre of the frame who runs, gets fouled, falls, sits and is lifted up. PERSON A must NOT look like the source player: no navy blue football shirt, no number 10 or any number or name on the back, no white football shorts, no red knee socks, no football boots, no white tape or wristbands on the wrists, no short black buzz-cut hair — unless the same thing is visible in their own photo. OTHER PEOPLE stay exactly as in the source and never take the look or clothes of PERSON A: opponents in white-and-sky-blue vertical striped shirts with black shorts; the tall teammate with number 9 in a navy blue shirt, white shorts and red socks, short dark hair and dark beard; the referee in a bright neon-green shirt at the very start. PERSON A never wears their kit. LOCATION AND OBJECTS stay exactly as in the source: real stadium grass pitch with mowing stripes and white lines, bright sunshine with hard shadows. Exactly ONE football (white with red pattern), it stays at the feet of PERSON A while dribbling, never multiplies. Normal human anatomy, five fingers on each hand, two legs, no extra limbs during the fall. No text on screen, no subtitles, no watermark, no logos.
A3. 같은 지시문, 원본 전체 길이로 — «없던 장면»은 길이 불일치 탓임을 확인
위에서부터 원본 · 힉스필드 · 우리. 비교 영상은 원본 소리 포함.
고양이 21초 · 690크레딧

보낸 지시문 원문 (3708자)
Source video: @Video1. Replace the cat in the source video with the cat from the reference image. The reference image shows one cat: copy that cat's breed, head shape, face, ear shape, fur length, fur color, markings, face mask, tail and eye color exactly as in its photo, in every shot. Keep its breed, fur color and markings exactly as in the photo. Do not invent or change its look. The pose and the room in the photo do not matter: take the pose and motion only from the video, and the backgrounds only from the video. Keep the source video's motion, timing, camera moves, zooms, framing, lens, cuts, dissolves and lighting exactly. Normal real-time speed, exactly matching the source video timing, no slow motion, no speed-up, no time compression. THE VIDEO HAS 4 SHOTS. Keep every cut and dissolve exactly there, add no new cuts: - SHOT 1 (0.00–5.875 s): THE CAT stands upright on its hind legs in deep blue starry space above a glowing blue planet horizon, front paws held in front of its chest, looking at the camera; the camera pulls back, the cat starts to dance; from about 4.8 s the stars turn into a white hyperspace warp and the cat floats in the middle with its legs spread and front paws raised. - SHOT 2 (5.875–13.29 s, hard cut): warm orange glow with large floating soap bubbles and glass cubes; THE CAT dances upright on its hind legs, raises both front paws high and wide, sways, turns to the side and stretches one front paw forward; the camera pushes in from a full-body shot to a side medium shot. - SHOT 3 (13.29–15.5 s, hard cut): close-up of THE CAT's head and chest on a vivid purple backdrop with floating violet glass cubes; it looks at the camera, sings, raises a front paw at the left edge, turns its head to the right. - 15.5–15.8 s: dissolve into SHOT 4. - SHOT 4 (15.8–21.04 s): close-up of THE CAT's head and chest on a shiny gold backdrop with glowing golden glass blocks and sparkles; it turns its head, sings, raises a front paw in the lower foreground, then calmly looks into the camera and sings again. SINGING: THE CAT sings the song like a singing animal in a music video: whenever the source cat opens its mouth, the cat's lower jaw visibly drops open and snaps shut on every sung syllable, EXACTLY in time — about 13.9–15.3 s, 17.1–17.4 s, 17.9–18.1 s, 18.6–19.6 s and 20.4–21.0 s. Real cat mouth, small pink tongue inside, no panting, no human lips. When the source cat's mouth is closed, the mouth stays closed. THE CAT must NOT look like the cat in the source video: no grey-and-white bicolor British longhair coat, no solid grey cap and back with a white face blaze, no white chest bib and white belly, no round grey-and-white face with copper-orange eyes — unless the same thing is visible in its own reference photo. The cat stands and moves on its hind legs exactly like the source cat. Real cat legs ending in round fluffy cat paws with toe beans and fur — NO fingers, no thumb, no human hands; when it raises a paw, it lifts the whole round paw, toes together, no finger pointing up. Belly and groin covered with dense fur, anatomically neutral like a plush toy, no nipples, no genitals — also when it dances with its legs spread. Realistic real cat, natural fur only, no clothes, no collar, no bow, no accessories. BACKGROUNDS (stay exactly as in the source, nothing redesigned, never the room from the photo): starry space with the blue planet horizon and the hyperspace warp, the orange bubbles and glass cubes, the purple backdrop with violet cubes, the gold backdrop with golden blocks. NO OTHER ANIMALS OR PEOPLE: this video contains only this one cat. No text, no subtitles, no Chinese characters, no watermark. Photorealistic music video.
축구 24초 · 780크레딧

보낸 지시문 원문 (3262자)
Source video: @Video1. Replace ONLY the main player (the one wearing number 10 in the source) with the person from @Image1. Copy that person's face, hairstyle, hair color, facial hair, skin tone, age, body type, height, body proportions, headwear and clothing exactly as in their photo, in every frame. If the person in the photo is short, old, bearded or wears a hat, keep exactly that: same height and proportions, same hat on the head the whole time, also while sprinting, falling and rolling on the grass. Do not invent or change their look. Everybody else stays as in the source. CLOTHES AND HAIR COME ONLY FROM THE PHOTO, NEVER FROM THE SOURCE PLAYER: the person wears exactly the outfit, socks and shoes from @Image1 from the first to the last frame, even when seen from behind, in motion blur, lying on the ground and in the close-up at the end. Keep the source video's motion, timing, camera moves, zooms, framing and lighting exactly. Normal real-time speed, exactly matching the source video timing, no slow motion, no speed-up, no time compression. ONE continuous shot, no cuts. WHAT HAPPENS (one continuous handheld broadcast shot on a sunny football pitch): - 0.0–1.5 s: PERSON A stands on the grass with the back to the camera, then starts to move. - 1.5–8.5 s: PERSON A sprints forward across the pitch past opponents, then dribbles one football with the feet; the camera pans and follows. - 8.5–11.0 s: an opponent in a white-and-sky-blue striped shirt runs into PERSON A from the left, they collide and BOTH fall and roll on the grass. - 11.5–16.0 s: PERSON A sits on the grass with legs bent, arms out to the sides, looking around and gesturing with both hands, complaining. - 16.0–20.5 s: a tall teammate (navy blue shirt with number 9, white shorts, red socks) comes in from the left, leans over and pulls PERSON A up by the arm; for a moment the teammate's back covers PERSON A. - 20.5–23.6 s: close-up, PERSON A stands up bent forward, hands near the knees, grimacing and then laughing, looking around. WHO IS WHO: - PERSON A = @Image1: the main player in the centre of the frame who runs, gets fouled, falls, sits and is lifted up. PERSON A must NOT look like the source player: no navy blue football shirt, no number 10 or any number or name on the back, no white football shorts, no red knee socks, no football boots, no white tape or wristbands on the wrists, no short black buzz-cut hair — unless the same thing is visible in their own photo. OTHER PEOPLE stay exactly as in the source and never take the look or clothes of PERSON A: opponents in white-and-sky-blue vertical striped shirts with black shorts; the tall teammate with number 9 in a navy blue shirt, white shorts and red socks, short dark hair and dark beard; the referee in a bright neon-green shirt at the very start. PERSON A never wears their kit. LOCATION AND OBJECTS stay exactly as in the source: real stadium grass pitch with mowing stripes and white lines, bright sunshine with hard shadows. Exactly ONE football (white with red pattern), it stays at the feet of PERSON A while dribbling, never multiplies. Normal human anatomy, five fingers on each hand, two legs, no extra limbs during the fall. No text on screen, no subtitles, no watermark, no logos.
엘리베이터 24초 · 790크레딧

보낸 지시문 원문 (3667자)
Source video: @Video1. IDENTITY / OUTFIT LOCK @Image1 = THE MAN. @Image2 = THE WOMAN. THE MAN is exactly the character from @Image1 in every frame: same face, hair, skin, body, height, proportions, full outfit and accessories, including his headphones from @Image1. His complete look from @Image1 stays identical from the first frame to the last. THE WOMAN is exactly the character from @Image2 in every frame: same face, same hairstyle and hair color, same body shape, height and proportions, same outfit as in @Image2. Her face, hair and body stay identical from the first frame to the last. Both characters keep the render style of their own reference images. ROLES AND BLOCKING (two people only, the whole video) 0.0–19.9 s, medium two-shot inside the elevator: THE MAN stands on the left side of the frame, THE WOMAN stands on the right side of the frame, both framed from the chest up, side by side. - 0.0–2.0 s: THE MAN wears his @Image1 headphones on his head and looks ahead; THE WOMAN glances at him and smiles. - 2.0–3.5 s: THE MAN raises his hand and pulls his @Image1 headphones off his head down around his neck. - 3.5–19.9 s: THE MAN's headphones rest around his neck. The two turn to each other and talk; THE WOMAN speaks to him in profile (about 3–11 s), then turns toward the camera, smiles, looks up and sings (about 12–18 s); THE MAN listens in profile and smiles. 19.9 s: hard cut to a wide frontal shot of the closed elevator doors. - 19.9–21.0 s: the doors slide open; THE MAN stands on the left inside the elevator, THE WOMAN stands next to him on the right, both seen full body. - 21.0–22.5 s: THE WOMAN walks forward out of the elevator and exits the frame on the right. - 22.5–24.1 s: THE MAN stands alone in the elevator, full body, headphones around his neck, looking ahead, and speaks a short line to himself: his lips clearly move in sync with his voice in the original audio, even though his face is small in this wide shot. LIP SYNC (TOP PRIORITY) Lip sync is the most important part of this video. Both characters keep the original lip sync: THE MAN's and THE WOMAN's mouths open, close and shape every word exactly in sync with the speech and singing of the original audio, frame by frame, with the same timing as the original performers. When the original mouth is closed, the character's mouth is closed. Facial expressions, smiles, eye direction and head turns follow the original. FINAL LINE: at the end of the video (22.5–24.1 s), alone in the wide shot, THE MAN must visibly say his line: his mouth opens and closes on every syllable, exactly in sync with his voice in the original audio. His mouth is clearly visible and clearly moving while he speaks. LOCATION Same elevator as the input: warm dark-brown panelled walls with brushed-steel vertical strips, soft warm overhead light; in the wide shot, brushed-steel sliding doors and a metal handrail on the back wall. Same lighting and color as the input. HARD CHECKS - Exactly two people in every frame where people are visible: THE MAN (@Image1) and THE WOMAN (@Image2); in the last 1.6 s only THE MAN. - THE MAN looks exactly like @Image1 in every frame, including both close shot and wide shot; his @Image1 headphones are on his head 0–2 s and around his neck from 3.5 s to the end. - THE WOMAN looks exactly like @Image2 in every frame, including her walk-out in the wide shot: same face, hairstyle and body. - Lip movements of both characters match the original audio, including THE MAN's final spoken line alone in the wide shot (22.5–24.1 s). - One hard cut at 19.9 s, no other cuts, no transitions, no morphing. Keep motion, camera, timing, background unchanged.
A4. 5초 영상 + 5초만 적은 지시문 — 공정 확정판
위에서부터 원본 5초 · 넣은 그림 · 힉스필드 · 우리(A2) · 우리(이번). 165크레딧씩.
고양이 (동물) — 됨 · 없던 장면 0

보낸 지시문 원문 (1276자)
Source video: @Video1. Replace the cat in the source video with the cat from the reference image. The reference image shows one cat: copy that cat's breed, head shape, face, ear shape, fur length, fur color, markings, face mask, tail and eye color exactly as in its photo, in every shot. Keep its breed, fur color and markings exactly as in the photo. Do not invent or change its look. The pose and the room in the photo do not matter: take the pose and motion only from the video, and the backgrounds only from the video. Keep the source video's motion, timing, camera moves, zooms, framing, lens, cuts, dissolves and lighting exactly. Normal real-time speed, exactly matching the source video timing, no slow motion, no speed-up, no time compression. THE VIDEO IS 5 SECONDS, ONE SHOT, no cuts: - 0.0–5.0 s: THE CAT stands upright on its hind legs in deep blue starry space above a glowing blue planet horizon, front paws held in front of its chest, looking at the camera; the camera pulls back and the cat starts to dance; from about 4.8 s the stars begin to stretch into a white hyperspace warp. Nothing else happens. Same background for all 5 seconds. THE CAT must NOT look like the cat in the source video: no grey-and-white bicolor coat, no grey head, no orange eyes.
차 (차 → 다른 차) — 됨 · 컷·차 안 시점 그대로

보낸 지시문 원문 (877자)
Source video: @Video1. Car reference: @Image1 (front three-quarter view) and @Image2 (rear view) show one dark navy-blue sedan with black wheels, red brake calipers, red accent stripe and a red full-width LED tail light bar. Replace ONLY the silver car with dark-blue racing stripes in @Video1 with the car from @Image1 and @Image2. Copy its body shape, dark navy-blue paint, black wheels, red calipers, front and rear design exactly, in every shot. No silver paint, no racing stripes. Keep everything else in @Video1 exactly as it is: the same 5 seconds, the same shots and hard cuts, camera moves, framing, the view from inside the following car, the road, other cars, smoke, people and lighting, and the black bars and TikTok watermark. The replaced car drives, drifts and moves exactly like the original car, at the same moments and positions. Normal speed, no new shots.
엘리베이터 두 사람 — 됨 · 헤드폰 벗는 시점까지 같음

보낸 지시문 원문 (1349자)
Source video: @Video1. IDENTITY / OUTFIT LOCK @Image1 = THE MAN. @Image2 = THE WOMAN. THE MAN is exactly the character from @Image1 in every frame: same face, hair, skin, body, height, proportions, full outfit and accessories, including his headphones from @Image1. His complete look from @Image1 stays identical from the first frame to the last. THE WOMAN is exactly the character from @Image2 in every frame: same face, same hairstyle and hair color, same body shape, height and proportions, same outfit as in @Image2. Her face, hair and body stay identical from the first frame to the last. Both characters keep the render style of their own reference images. ROLES AND BLOCKING (5 seconds, two people only, one medium two-shot inside the elevator, no cuts) THE MAN stands on the left side of the frame, THE WOMAN on the right, both framed from the chest up, side by side. - 0.0–2.0 s: THE MAN wears his @Image1 headphones on his head and looks ahead; THE WOMAN glances at him and smiles. - 2.0–3.5 s: THE MAN raises his hand and pulls his @Image1 headphones off his head down around his neck. - 3.5–5.0 s: the headphones rest around his neck; the two turn to each other; THE WOMAN speaks to him in profile. The doors stay closed. Nobody leaves. Keep the source framing, camera and cuts exactly; mouths move in time with the original performers.
드레스 (옷만) — 됨 · 전신 구도까지 힉스필드와 같음(A2 에서 가까이 잡힌 건 지시문을 비워서)

보낸 지시문 원문 (775자)
Source video: @Video1. Outfit reference: @Image1 shows a white satin off-the-shoulder bridal gown with puffed short sleeves, a fitted ruched bodice, a long flowing skirt and a long sheer white veil. Change ONLY the clothes of the dancing woman in @Video1: she now wears exactly the bridal gown and veil from @Image1. Keep the SAME woman from @Video1: her face, hair, skin and body stay exactly as in @Video1. Do NOT use the face or hair of the woman in @Image1, do not copy the bouquet. Keep everything else exactly as in @Video1: the same 5 seconds, her dance moves and timing, the full-body wide framing, the camera position and distance, the night street, houses, street lamps, lighting and the small watermark. Do not move the camera closer, no close-ups, no new shots.
축구 한 명 — 사람·동작은 됨 · 구도 빗나감(가까이 · 관중석 새로 그림) → A5

보낸 지시문 원문 (1676자)
Source video: @Video1. Replace ONLY the main player (the one wearing number 10 in the source) with the person from @Image1. Copy that person's face, hairstyle, hair color, facial hair, skin tone, age, body type, height, body proportions, headwear and clothing exactly as in their photo, in every frame. If the person in the photo is short, old, bearded or wears a hat, keep exactly that: same height and proportions, same hat on the head the whole time, also while sprinting, falling and rolling on the grass. Do not invent or change their look. Everybody else stays as in the source. CLOTHES AND HAIR COME ONLY FROM THE PHOTO, NEVER FROM THE SOURCE PLAYER: the person wears exactly the outfit, socks and shoes from @Image1 from the first to the last frame, even when seen from behind, in motion blur, lying on the ground and in the close-up at the end. Keep the source video's motion, timing, camera moves, zooms, framing and lighting exactly. Normal real-time speed, exactly matching the source video timing, no slow motion, no speed-up, no time compression. ONE continuous shot, no cuts. WHAT HAPPENS (5 seconds, one continuous handheld broadcast shot on a sunny football pitch): - 0.0–1.5 s: PERSON A stands on the grass with the back to the camera, then starts to move. - 1.5–5.0 s: PERSON A sprints forward across the pitch, the camera pans and follows. Nobody falls, no collision. PERSON A = @Image1: the main player in the centre of the frame. PERSON A must NOT look like the source player: no navy blue football shirt, no number 10, no white football shorts, no red knee socks, no football boots, no short black buzz-cut hair. Other players stay as in the source.
A5. 축구 — 카메라·동선을 0.25초 단위로 다시 쓴 판 (구도 해결)
원본을 0.25초마다 보고 «높은 중계 카메라 · 내려다봄 · 관중석·하늘 안 보임 · 전신이 화면 높이 2/3» 와 «1.6초 심판 지나감 · 2.5~3.4초 상대 둘이 바로 옆을 지나감 · 4.5초 센터서클 선»을 적었다. 결과는 힉스필드와 거의 같다. 비교 영상: 원본 · 힉스필드 · 우리(짧은 지시문) · 우리(자세한 지시문). 165크레딧.


보낸 지시문 원문 (3133자)
Source video: @Video1. Replace ONLY the main player (the one wearing number 10 in the source) with the person from @Image1. Copy that person's face, hairstyle, hair color, facial hair, skin tone, age, body type, height, body proportions, headwear and clothing exactly as in their photo, in every frame. If the person in the photo is short, old, bearded or wears a hat, keep exactly that: same height and proportions, same hat on the head the whole time, also while sprinting, falling and rolling on the grass. Do not invent or change their look. Everybody else stays as in the source. CLOTHES AND HAIR COME ONLY FROM THE PHOTO, NEVER FROM THE SOURCE PLAYER: the person wears exactly the outfit, socks and shoes from @Image1 from the first to the last frame, even when seen from behind, in motion blur, lying on the ground and in the close-up at the end. Keep the source video's motion, timing, camera moves, zooms, framing and lighting exactly. Normal real-time speed, exactly matching the source video timing, no slow motion, no speed-up, no time compression. ONE continuous shot, no cuts. CAMERA AND FRAMING (very important, same as the source for all 5 seconds): A TV broadcast camera high up in the stands, looking DOWN at the pitch with a long telephoto lens. Only green grass with mowing stripes fills the background: NO stands, NO crowd, NO advertising boards, NO sky are visible at any moment. PERSON A is seen full body from head to shoes, filling about two thirds of the frame height, roughly in the centre. Do NOT move the camera closer, no close-ups, no ground-level view. The camera pans smoothly to keep PERSON A in the centre while running. WHAT HAPPENS (5 seconds, one continuous shot, no cuts): - 0.0–1.0 s: PERSON A stands still on the grass with the back to the camera, arms hanging; only the legs of a referee and other players are visible at the top edge of the frame. - 1.0–1.6 s: PERSON A bends the knees and turns to the right. - 1.6–2.5 s: PERSON A starts to sprint to the right; a referee in a bright yellow-green shirt runs past at the top right edge. - 2.5–3.4 s: PERSON A sprints to the right, seen from the side; two opponents in white-and-sky-blue striped shirts and black shorts come in from the left and run past right next to PERSON A, one crossing in front of PERSON A in the foreground. - 3.4–5.0 s: PERSON A keeps sprinting to the right, seen from the side; at about 4.5 s a white line of the centre circle appears at the bottom left and an opponent in a striped shirt runs at the right edge. Nobody falls, no collision, no ball in the frame. WHO IS WHO: - PERSON A = @Image1: the main player in the centre who stands, turns and sprints. PERSON A must NOT look like the source player: no navy blue football shirt, no number 10 or name on the back, no white football shorts, no red knee socks, no football boots, no short black buzz-cut hair. OTHER PEOPLE stay exactly as in the source and never take the look or clothes of PERSON A: opponents in white-and-sky-blue striped shirts with black shorts, the referee in a bright yellow-green shirt. No text on screen, no subtitles, no watermark.
7. 제품 바꾸기 예시
B1. 광고 편집본·영화 장면의 물건 바꾸기
향수 광고(흑백 9컷, 힉스필드 카타나 예시 영상 3.65~8.65초) → 빨간 «FIZZ» 캔 — 됨 · 165크레딧
향수병 클로즈업 → 물방울 맺힌 캔 윗면, 광고 문구도 «FIZZ THE MOST WANTED»로 바뀜. 첫 두 컷 길이가 조금 밀리고 가운데 작은 자막이 빠짐. 위 원본 · 아래 우리.


보낸 지시문 원문 (1627자)
Source video: @Video1. Product reference: @Image1 shows one cherry-red aluminum soda can with a white wave stripe and the word 'FIZZ'. Replace the black perfume bottle in @Video1 with the soda can from @Image1 (THE PRODUCT). Copy its can shape, red metallic paint, white stripe and 'FIZZ' lettering exactly; in the black-and-white shots it keeps the same can shape and lettering. No perfume bottle, no perfume cap, no glass. Keep everything else in @Video1 exactly as it is: the same 5 seconds, the same shots and hard cuts at the same moments, camera moves, framing, black-and-white color grading, lighting, the people, rooms and the small italic white caption in the middle of the first shots. Normal speed, no new shots. SHOTS (5 seconds, black and white): - 0.0–0.5 s: extreme close-up of the top of THE PRODUCT, slowly rotating, filling the frame. - 0.5–1.25 s: THE PRODUCT stands alone on a glossy black surface in darkness, centered, soft rim light, slow push-in. - 1.25–1.75 s: same shot, THE PRODUCT on the right; white title text on the left: 'FIZZ' in large thin letters, below it 'THE MOST WANTED' and a small line 'THE NEW ICE-COLD SODA'. - 1.75–2.0 s: a man in dark clothes stands by a bright window in profile. - 2.0–2.5 s: close-up of the man's face, his fingers near his chin. - 2.5–3.0 s: a woman in a long white gown seen from behind in a doorway. - 3.0–3.5 s: the man sits at a table facing the camera, flowers in the foreground. - 3.5–4.25 s: wide shot of an empty elegant dining hall with chandeliers. - 4.25–5.0 s: extreme close-up of the man's eyes. THE PRODUCT appears only in the first three shots.
같은 향수 광고 → 치즈버거 (극과극) — 결과 «안전 기준» 거절 · 0크레딧

보낸 지시문 원문 (1662자)
Source video: @Video1. Product reference: @Image1 shows one big cheeseburger with sesame bun, beef patty, melted cheese, lettuce and tomato. Replace the black perfume bottle in @Video1 with the cheeseburger from @Image1 (THE PRODUCT), shown as a luxury object in the same elegant style. Copy its bun, patty, cheese, lettuce and tomato exactly. In the first close-up the camera looks down at the sesame bun top. No perfume bottle, no perfume cap, no glass. Keep everything else in @Video1 exactly as it is: the same 5 seconds, the same shots and hard cuts at the same moments, camera moves, framing, black-and-white color grading, lighting, the people, rooms and the small italic white caption in the middle of the first shots. Normal speed, no new shots. SHOTS (5 seconds, black and white): - 0.0–0.5 s: extreme close-up of the top of THE PRODUCT, slowly rotating, filling the frame. - 0.5–1.25 s: THE PRODUCT stands alone on a glossy black surface in darkness, centered, soft rim light, slow push-in. - 1.25–1.75 s: same shot, THE PRODUCT on the right; white title text on the left: 'BIG BITE' in large thin letters, below it 'THE MOST WANTED' and a small line 'THE NEW DOUBLE CHEESE'. - 1.75–2.0 s: a man in dark clothes stands by a bright window in profile. - 2.0–2.5 s: close-up of the man's face, his fingers near his chin. - 2.5–3.0 s: a woman in a long white gown seen from behind in a doorway. - 3.0–3.5 s: the man sits at a table facing the camera, flowers in the foreground. - 3.5–4.25 s: wide shot of an empty elegant dining hall with chandeliers. - 4.25–5.0 s: extreme close-up of the man's eyes. THE PRODUCT appears only in the first three shots.
드리프트 차(A2 차 예시 원본 5초) → 쇼핑 카트 (극과극) — 됨 · 165크레딧
카트가 차 자리에서 연기를 뿜으며 미끄러짐. 1~2초 차 안에서 본 컷이 길 장면으로 바뀜. 위 원본 · 아래 우리(화면 가운데 띠).


보낸 지시문 원문 (882자)
Source video: @Video1. Object reference: @Image1 shows one empty metal supermarket shopping cart with a red plastic handle and four small black wheels. Replace ONLY the silver car with dark-blue racing stripes in @Video1 with a car-sized version of the shopping cart from @Image1. It rolls and drifts on its small wheels exactly where the car drives, at the same moments, positions, speed and angles, kicking up the same tire smoke. Copy the cart's metal wire basket, red handle and black wheels exactly. No silver car, no racing stripes, nobody sits in the cart. Keep everything else in @Video1 exactly as it is: the same 5 seconds, the same shots and hard cuts, camera moves and framing, the view from inside the following car with the driver's hands and the dashboard, the road, other cars, people, lighting, the black bars and the TikTok watermark. Normal speed, no new shots.
B2. 실제 TV 광고(유튜브) — 모두 거절, 0크레딧
토레타(1.2~6.2초) → 코카콜라 · 스프라이트 제로 위드 티(0~5초) → 립스틱 · 나랑드사이다(0~5초) → 레드불. 소리 뺌.
- 토레타·스프라이트: 입력 단계 «an input was flagged as possibly containing copyrighted content». 토레타는 두 번 모두 같음.
- 나랑드사이다 → 레드불: 1회차 입력 통과 → 결과 «violates the safety policy». 2회차 레드불 + 남자도 다른 사람으로 → 같은 결과 거절. 3회차 상표 없는 FIZZ 캔 → 이번엔 입력 단계 저작권 거절(같은 영상인데 검사 결과가 바뀜).
- 결론: 유명 광고 원본은 씨댄스 검사에 걸린다. 검사를 피하려 영상을 뒤집거나 자르는 우회는 하지 않았다. 실제 상표 자체는 문제가 아님(B3).

토레타→코카콜라 지시문 (1457자)
Source video: @Video1. Product reference: @Image1 shows one Coca-Cola 500 ml plastic bottle with dark cola inside, the red label with the white Coca-Cola logo and a red cap. Replace the clear water bottle with the green cap and pink-green label in @Video1 with the Coca-Cola bottle from @Image1, everywhere it appears. Copy its shape, dark cola color, red label, white Coca-Cola logo and red cap exactly. No clear water, no green cap, no Korean label on the bottle. Keep everything else in @Video1 exactly as it is: the same 5 seconds, the same shots and hard cuts at the same moments, the same people with the same faces, hair, clothes, poses, expressions and motion, camera moves, framing, rooms, backgrounds, lighting and colors. Keep all on-screen Korean text and graphics exactly where and when they appear. Normal speed, no new shots, no slow motion. SHOTS (5 seconds): - 0.0–0.25 s: close-up of a young woman with platinum blond hair lying down. - 0.25–1.5 s: the woman in a green top and white cardigan in a bright kitchen; THE BOTTLE stands large in the left foreground; she leans in and at about 1.0 s picks THE BOTTLE up in her hand. - 1.5–2.75 s: wide shot of a living room; the woman stretches on a yoga mat on the floor; THE BOTTLE stands small on the floor next to the mat. - 2.75–3.75 s: soft close-up of the woman's tilted face. - 3.75–5.0 s: top-down close-up of the woman lying on a blue mat; a hand holds THE BOTTLE against her cheek.
스프라이트→립스틱 지시문 (1401자)
Source video: @Video1. Product reference: @Image1 shows one luxury lipstick: glossy black-and-gold tube with a bright red bullet. Replace the green-and-orange soda bottle in @Video1 with the lipstick from @Image1, everywhere it appears, at a size that fits in the hand. Copy its black-and-gold tube and red bullet exactly. No soda bottle, no orange drink, no green bottle label. Keep everything else in @Video1 exactly as it is: the same 5 seconds, the same shots and hard cuts at the same moments, the same people with the same faces, hair, clothes, poses, expressions and motion, camera moves, framing, rooms, backgrounds, lighting and colors. Keep all on-screen Korean text and graphics exactly where and when they appear. Normal speed, no new shots, no slow motion. SHOTS (5 seconds): - 0.0–1.75 s: top-down shot of crushed ice; THE LIPSTICK lies on the ice in the centre, the camera slowly turns and slides; green Korean title text appears at the top left. - 1.75–2.75 s: a woman with long dark wavy hair at a sunny pool club smiles and holds THE LIPSTICK up near her shoulder on the left; small subtitle at the bottom. - 2.75–3.5 s: grey wall with diagonal white stripes of light; the woman stands, then three copies of her stand side by side, one holding THE LIPSTICK. - 3.5–5.0 s: medium close-up of the woman against the grey striped wall, holding THE LIPSTICK beside her face and talking.
나랑드→레드불 지시문 (1254자)
Source video: @Video1. Product reference: @Image1 shows one Red Bull Energy Drink slim can: blue and silver checkered pattern, two red bulls and a yellow sun, red 'Red Bull' logo. Replace the light-blue soda can in @Video1 with the Red Bull can from @Image1, everywhere it appears. Copy its slim shape, blue-silver pattern, red bulls, yellow sun and red logo exactly. No light-blue can, no 'Narangd Cider' text on the can. Keep everything else in @Video1 exactly as it is: the same 5 seconds, the same shots and hard cuts at the same moments, the same people with the same faces, hair, clothes, poses, expressions and motion, camera moves, framing, rooms, backgrounds, lighting and colors. Keep all on-screen Korean text and graphics exactly where and when they appear. Normal speed, no new shots, no slow motion. SHOTS (5 seconds): - 0.0–0.75 s: a hand raises THE CAN against a bright blue sky with white clouds. - 0.75–2.0 s: extreme close-up: fingers pull the ring tab of the cold, wet THE CAN open. - 2.0–2.25 s: a yellow-orange food truck, blue Korean graphic text appears. - 2.25–5.0 s: a man in black sunglasses, light-blue shirt and blue vest seen in profile in front of colorful food trucks, looking up; blue and pink Korean text on the left.
나랑드→레드불 + 사람 바꾸기 지시문(사람 그림 = A2 엘리베이터 남자 그림) (1545자)
Source video: @Video1. Product reference: @Image1 shows one Red Bull Energy Drink slim can: blue and silver checkered pattern, two red bulls and a yellow sun, red 'Red Bull' logo. Replace the light-blue soda can in @Video1 with the Red Bull can from @Image1, everywhere it appears. Copy its slim shape, blue-silver pattern, red bulls, yellow sun and red logo exactly. No light-blue can, no 'Narangd Cider' text on the can. PERSON: also replace the man in black sunglasses with the man from @Image2. Copy his face, curly brown hair and skin exactly from @Image2; he wears the same black sunglasses, light-blue shirt and blue vest as in @Video1 and makes the same head turns and expressions. He must NOT look like the man in the source video. Keep everything else in @Video1 exactly as it is: the same 5 seconds, the same shots and hard cuts at the same moments, the same poses, expressions and motion, camera moves, framing, rooms, backgrounds, lighting and colors. Keep all on-screen Korean text and graphics exactly where and when they appear. Normal speed, no new shots, no slow motion. SHOTS (5 seconds): - 0.0–0.75 s: a hand raises THE CAN against a bright blue sky with white clouds. - 0.75–2.0 s: extreme close-up: fingers pull the ring tab of the cold, wet THE CAN open. - 2.0–2.25 s: a yellow-orange food truck, blue Korean graphic text appears. - 2.25–5.0 s: THE MAN from @Image2 in black sunglasses, light-blue shirt and blue vest seen in profile in front of colorful food trucks, looking up; blue and pink Korean text on the left.
나랑드→FIZZ 캔 지시문 (1210자)
Source video: @Video1. Product reference: @Image1 shows one cherry-red aluminum soda can with a white wave stripe and the made-up word 'FIZZ'. Replace the light-blue soda can in @Video1 with the red FIZZ can from @Image1, everywhere it appears. Copy its shape, red metallic paint, white stripe and 'FIZZ' lettering exactly. No light-blue can, no 'Narangd Cider' text on the can. Keep everything else in @Video1 exactly as it is: the same 5 seconds, the same shots and hard cuts at the same moments, the same people with the same faces, hair, clothes, poses, expressions and motion, camera moves, framing, rooms, backgrounds, lighting and colors. Keep all on-screen Korean text and graphics exactly where and when they appear. Normal speed, no new shots, no slow motion. SHOTS (5 seconds): - 0.0–0.75 s: a hand raises THE CAN against a bright blue sky with white clouds. - 0.75–2.0 s: extreme close-up: fingers pull the ring tab of the cold, wet THE CAN open. - 2.0–2.25 s: a yellow-orange food truck, blue Korean graphic text appears. - 2.25–5.0 s: a man in black sunglasses, light-blue shirt and blue vest seen in profile in front of colorful food trucks, looking up; blue and pink Korean text on the left.
B3. 우리가 만든 광고 장면 → 실제 제품 (3/3 성공) ★
잘 나온 광고의 제품 장면을 본떠 원본을 직접 만들었다(상표 없는 제품). GPT 이미지 2 로 첫 장면(1536×864, 35크레딧) → 미니맥스 H3 Max Turbo 로 5초 영상(약 25크레딧) → 앞 5초를 원본으로 씨댄스 제품 바꾸기(165크레딧). 비교 영상: 왼쪽 원본 · 오른쪽 바꾼 것. 그림: 위 원본 · 아래 바꾼 것(0.5초 간격).
원본 장면 만들기 — 그림 지시문(IMG) · 영상 지시문(VID) 원문 (2125자)
import sys
sys.path.insert(0, "/Volumes/Samsung T5/Projects/HaHa Animation/app")
import creagen_gen as G, falq
W = "/Volumes/Samsung T5/Projects/HaHa Animation/app/uploads/genjutsu_1116/own/"
k = sys.argv[1]
IMG = {
"kitchen": "Cinematic TV beverage commercial still, 16:9: a young Korean woman with long platinum blond braided hair, green crop top and white cardigan, smiling in a bright sunny pastel kitchen; a clear plastic water bottle with a green cap and a plain mint label stands large and sharp in the left foreground on the counter; soft morning light, shallow depth of field, premium ad look, no text, no logo.",
"ice": "Cinematic TV soda commercial still, 16:9, top-down view: one green plastic soda bottle with orange tea-colored drink and a plain green label lies diagonally on a bed of sparkling crushed ice, water droplets, bright cool light, premium ad look, no text, no logo.",
"sky": "Cinematic TV soda commercial still, 16:9: a man's hand raises one slim light-blue aluminum soda can with a plain white label high against a bright blue sky with fluffy white clouds, low angle, sunny, cold droplets on the can, premium ad look, no text, no logo.",
}
VID = {
"kitchen": "The woman leans toward the camera smiling, reaches out and picks up the water bottle from the counter with her right hand and lifts it beside her face, cheerful. Smooth slow push-in. Premium beverage commercial.",
"ice": "The camera slowly rotates and glides over the bottle lying on the crushed ice, ice sparkles, cold mist drifts. Premium soda commercial.",
"sky": "The hand slowly turns the can toward the camera against the blue sky, then the camera cuts to an extreme close-up of fingers pulling the ring tab open with a fizzy spray. Premium soda commercial.",
}
G.gen("t2i", "openai/gpt-image-2", W + k + ".png", {"prompt": IMG[k], "quality": "high", "image_size": {"width": 1536, "height": 864}})
print("IMG", k, flush=True)
G.gen("i2v", "fal-ai/minimax/h3-max-turbo/image-to-video", W + k + "_src.mp4", {"prompt": VID[k], "image_url": falq.upload(W + k + ".png"), "duration": 5, "prompt_optimizer": False})
print("VID", k)
부엌 물병 → 코카콜라 — 됨 · 165크레딧
사람·웃음·손 동작·구도 그대로, 병만 코카콜라. 얼굴 옆으로 들어 보일 때 라벨도 보임.
원본 첫 그림 · 넣은 제품 그림:

보낸 지시문 원문 (1140자)
Source video: @Video1. Product reference: @Image1 shows one Coca-Cola 500 ml plastic bottle with dark cola inside, the red label with the white Coca-Cola logo and a red cap. Replace the clear water bottle with the green cap and plain mint label in @Video1 with the Coca-Cola bottle from @Image1, everywhere it appears, at the same size. Copy its shape, dark cola color, red label, white Coca-Cola logo and red cap exactly. No clear water, no green cap, no mint label. Keep everything else in @Video1 exactly as it is: the same 5 seconds, the same woman with the same face, blond braid, green top and white cardigan, her smile, laugh and every motion, the bright kitchen, camera position, framing, lighting and colors. Normal speed, one continuous shot, no cuts. WHAT HAPPENS (5 seconds): - 0.0–1.0 s: the woman leans on the counter with her chin on her hand, smiling; THE BOTTLE stands large on the counter in the left foreground. - 1.0–2.0 s: she laughs, reaches out with her right hand and grabs THE BOTTLE. - 2.0–5.0 s: she lifts THE BOTTLE and holds it up beside her face toward the camera, label facing the camera, smiling happily.
하늘 캔 → 레드불 — 됨 · 165크레딧
로고·황소까지, 2.75초 컷 전환·손·구름 그대로. 고리 따는 클로즈업에서도 캔 옆면이 레드불 파랑.
원본 첫 그림 · 넣은 제품 그림:

보낸 지시문 원문 (1030자)
Source video: @Video1. Product reference: @Image1 shows one Red Bull Energy Drink slim can: blue and silver checkered pattern, two red bulls and a yellow sun, red 'Red Bull' logo, silver top. Replace the plain light-blue and white can in @Video1 with the Red Bull can from @Image1, everywhere it appears. Copy its slim shape, blue-silver pattern, red bulls, yellow sun and red logo exactly. No plain white label, no light-blue can. Keep everything else in @Video1 exactly as it is: the same 5 seconds, the same two shots and the hard cut at the same moment, the hand, fingers and their motion, the blue sky and clouds, camera moves, framing, lighting and colors. Normal speed, no new shots. SHOTS (5 seconds): - 0.0–2.75 s: low angle, a man's hand holds THE CAN up high against a bright blue sky with white clouds, slowly turning it. - 2.75–5.0 s: hard cut to an extreme close-up of the silver top of THE CAN; fingers pull the ring tab open, cold foam bubbles up. The blue-silver side of THE CAN is visible at the bottom edge.
얼음 위 페트병 → 립스틱 — 됨 · 165크레딧
같은 자리·방향·비슷한 크기, 카메라 회전·찬 김 그대로. 그림대로 뚜껑이 옆에 따로 나옴.
원본 첫 그림 · 넣은 제품 그림:

보낸 지시문 원문 (831자)
Source video: @Video1. Product reference: @Image1 shows one luxury lipstick: glossy black-and-gold tube with a bright red bullet, and its black-and-gold cap. Replace the green plastic soda bottle with orange drink in @Video1 with the lipstick from @Image1, shown big like a hero product: the lipstick lies diagonally on the crushed ice exactly where the bottle lies, in the same direction, about as long as the bottle, red bullet at the top end where the bottle cap was. Copy its black-and-gold tube and red bullet exactly. No bottle, no orange drink, no green label. Keep everything else in @Video1 exactly as it is: the same 5 seconds, the sparkling crushed ice, cold mist and droplets, the slow top-down camera rotation and glide, framing, bright cool lighting and colors. Normal speed, one continuous shot, no cuts, no text.
제품 그림 지시문 (코카콜라 · 립스틱 · 레드불)
원문
import sys
sys.path.insert(0, "/Volumes/Samsung T5/Projects/HaHa Animation/app")
import creagen_gen as G
W = sys.argv[1]
P = {"coke": "Studio product photo of one classic Coca-Cola 500 ml plastic PET bottle: dark cola inside, iconic red label with the white Coca-Cola script logo, red cap, cold condensation droplets, standing upright, centered, plain light grey background, sharp, no other objects.",
"lipstick": "Studio product photo of one luxury lipstick: glossy black and gold tube, cap off and placed beside it, bright red bullet twisted up, standing upright, centered, plain light grey background, sharp, no other objects, no logo.",
"redbull": "Studio product photo of one Red Bull Energy Drink 250 ml slim aluminum can: blue and silver checkered pattern, two red bulls charging at a yellow sun, 'Red Bull' red logo text, cold condensation droplets, standing upright, centered, plain light grey background, sharp, no other objects."}
G.gen("t2i", "openai/gpt-image-2", W + "/" + sys.argv[2] + ".png", {"prompt": P[sys.argv[2]], "quality": "high", "image_size": {"width": 1024, "height": 1024}})
print("OK", sys.argv[2])
8. 남은 일 · 주의
- 오려 붙이기(마스크 합성) — 힉스필드 «물건 바꾸기»처럼 바꾼 대상 밖 화소를 원본 그대로 두려면: 원본에서 대상 마스크(SAM 계열) → 씨댄스 결과에서 그 자리(원본·결과 마스크 합집합)만 원본 위에 합성. 씨댄스가 구도를 지켜야 맞물림 → 지시문의 카메라 블록이 먼저.
- 사람이 여럿 나오는 넓은 장면은 구도가 흔들리기 쉽다 → 카메라·동선을 초 단위로.
- 제품을 원본과 다른 크기로 넣을 땐 «어디에, 무엇만큼» 명시(립스틱을 병 길이만큼).
- 유명 광고 원본은 쓰지 않는다(거절 · 권리). 권리 있는 영상이나 직접 만든 장면만.
- 1080p 는 시험하지 않았다(480p 만). 소리 생성도 껐다.
- 거절은 0크레딧, 성공 5초판 165크레딧. 같은 입력 재제출은 재과금 없음(멱등 키).