Skip to content

experiment(ai): composite 한 장으로 에디토리얼 asset suite 전체 생성 가능한가 (Nano Banana vs GPT Image 2) #290

@cocoyoon

Description

@cocoyoon

Summary

Fashion Decode composite 이미지 한 장 을 입력으로, 매거진 에디토리얼 한 편에 필요한 사진 asset 스위트 전체 를 이미지 생성 모델로 뽑아낼 수 있는지 실험. 성공하면 #261 역검색 파이프라인 자체의 존재 이유를 재검토해야 할 만큼 architecture 에 영향이 큼.

Related:

프로덕션 구현 아님 — 실험·평가. 결과는 이 이슈 코멘트로 시각 자료 + 판정 리포트, 채택 결정은 별도 이슈로 분리.


핵심 가설

composite (좌: 아티스트 / 우: 제품 그리드 + 라벨) 한 장이면 Nano Banana 또는 GPT Image 2 가 매거진 에디토리얼에 필요한 identity-consistent, style-coherent asset suite 를 생성할 수 있다.

이게 성립하면 → 역검색 불필요. composite → asset suite → 매거진. 끝.
이게 부분 성립하면 → 어떤 asset 은 자동화 가능, 어떤 건 역검색/사람 손 필요. 세분화된 파이프라인 설계 근거.
이게 실패하면#261 역검색 그대로 유지, 원본이 identity 보존의 기반 reference 로 반드시 필요.


Asset Suite 정의 (pin 한 장당 생성 목표)

# Asset 사양 용도 난이도
1 Hero portrait 1:1 or 4:5, 중앙 인물, 어깨 위 프레이밍 매거진 커버, 썸네일 그리드 낮음
2 Environmental shot 3:4 or 2:3, 하프/풀바디, 배경 컨텍스트 에디토리얼 스프레드 메인 비주얼 중간 (composite 에 없으면 extrapolation)
3 Item cutouts (N개) 각 아이템 isolated, clean/neutral bg, 제품샷 톤 레이아웃에 개별 배치 높음 — 그리드에서 개별 추출 + 재생성
4 Detail close-up (옵션) 특정 아이템 로고/텍스처 클로즈업 에디토리얼 악센트 중간
5 Title card (옵션) 타이포그래피 treatment, 배경 무드 커버 / 섹션 디바이더 텍스트 렌더링 품질 의존

최소 목표: #1 hero + #3 item cutouts (composite 의 items 개수만큼). 이 두 개만 되어도 매거진 그리드에 쓸 수 있음.

Stretch 목표: #1 + #2 + #3 + #4 가 전부 한 아티스트로 일관된 identity 를 유지하면서 하나의 매거진 에디토리얼로 조합 가능한 visual tone 을 공유.


테스트 대상 모델

항목 Gemini 2.5 Flash Image "Nano Banana" GPT Image 2
공식 marketing "multi-image fusion", "character consistency", "드래그앤드롭 제품 삽입" "8 coherent images from single prompt with character continuity"
우리 유스케이스에 직접 해당? ✅ — 제품 isolation, 인물 identity 보존 둘 다 세일즈 포인트 ✅ — 캐릭터 일관성 batch 생성 공식 지원
가격/이미지 $0.039 고정 low $0.006 / medium $0.053 / high $0.211
텍스트 렌더링 보통 우수
해상도 512 / 1K / 2K / 4K 1024 / 1792 / 2K(2560x1440)

Experiment Design

입력 (이미 확보됨)

케이스 매트릭스

10 pin x 두 API x asset suite (최소 2 asset: hero + items) = pin 당 2 API x (1 hero + N item cutouts)

예: Jennie hoodie 핀 (5 items) 의 경우

  • Nano Banana: 1 hero + 5 item cutouts = 6 호출
  • GPT Image 2: 동일 = 6 호출
  • 합계 12 호출

10 pin 전체로는 약 100~120 호출 예상.

프롬프트 전략

Hero portrait

"Editorial portrait of the person in the left half of this image. Magazine cover style, clean studio lighting, neutral background, centered composition, serif title treatment reading [ARTIST_NAME] at the bottom. Identity preserved from source image. 1:1 aspect ratio."

Environmental

"Editorial full-body shot of the person in the source image, same outfit, shot in a location that matches the mood (airport / street / studio inferred from source). Wide aspect ratio 3:4, hands visible, natural pose."

Item cutout (per item)

"Product cutout of the [SUBCATEGORY] in the source image's right panel, brand [BRAND]: clean white background, studio lighting, centered, 1:1 aspect ratio. Remove all other items and labels. Photorealistic catalog style."

(parse_result items JSON 에서 brand/subcategory 채움)


평가 축

설명 판정 기준
Identity consistency hero + environmental 에서 동일 아티스트 인식 가능? 팬 시점 blind test: 3/5 명이 동일인 판정해야 pass
Item isolation quality composite 우측 그리드에서 개별 아이템 cutout 이 깨끗한가 배경 오염 없음, 라벨 텍스트 잔재 없음, subcategory 정확
Label comprehension composite 내 "THE ROW — $1,590" 같은 라벨을 읽고 해당 제품만 isolated 하는가 정확한 item 이 생성된 비율
Visual tone coherence asset suite 가 하나의 에디토리얼로 조합 가능한 톤인가 라이팅·컬러·framing 일관성 주관 평가
Fidelity vs drift 원본 아이템 디자인이 얼마나 유지되는가 (브랜드 로고, 실루엣) 육안 비교
비용/시간 asset suite 한 벌 생성 총 비용 + 대기시간 Nano Banana vs GPT Image 2 per-pin 합계

비용 시뮬레이션

asset suite 한 벌 = 1 hero + 평균 3 items = 4장

시나리오 Nano Banana GPT Image 2 (medium) GPT Image 2 (high)
10 pin x 4 asset $1.56 $2.12 $8.44
+ 3 variation/asset (identity consistency 관찰) $4.68 $6.36 $25.32

iteration 포함 총 < $40 — 실험 가능.


결정 필요 사항

  • asset 범위: 최소(hero + items) / 확장(environmental / detail / title 포함)
  • variation 수: 각 asset 당 1장 / 3장 (identity drift 관찰 위해 3장 권장)
  • identity 실험 방법: 생성 후 얼굴 비교 수동 / CLIP similarity 자동
  • item cutout 기대 스타일: white bg catalog / editorial still life (shadow) / flat lay

산출물

  1. `/tmp/mag_experiment_v2/` — asset suite 파일들
    • `<pin_id>/nano_banana/hero.jpg`, `item_0.jpg`, ...
    • `<pin_id>/gpt_image_2/hero.jpg`, ...
  2. 이 이슈에 코멘트로 비교 리포트 업로드

이 실험의 아키텍처 임팩트

결과에 따라 #261 역검색 파이프라인의 운명 이 결정됨:

실험 결과 architecture 결정
Asset suite 품질 우수 + identity 일관 #261 제거, composite → 생성 으로 단순화
Hero 는 OK / item cutout 은 불만 부분 도입: hero 만 생성, 아이템은 현행 유지
Identity drift 심함 원본 없으면 quality 안 나온다 확인 → #261 강화
법적/윤리 우려 (celeb likeness) 기능 자체 보류, decoded 는 원본 사용 유지

Follow-up

  • 실험 성공 → `feat(ai): composite-based editorial asset generation pipeline` 이슈
  • 매거진 자동 레이아웃 단계 → `feat(web): decoded magazine auto-layout` 이슈

범위 밖

  • 프로덕션 API 통합
  • admin UI 편집 기능
  • 실 배포 / 셀럽 초상권 검토 (법무 리뷰는 별도)

Metadata

Metadata

Assignees

Labels

aiAI/자동화

Type

No type
No fields configured for issues without a type.

Projects

Status
Done

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions