GPT IMAGE 2 COMPLETE PROMPT LIBRARY
使用路由：openai/gpt-image-2（第三方平台名稱可能不同）
OpenAI 原廠模型代碼：gpt-image-2
固定快照：gpt-image-2-2026-04-21
版本日期：2026-08-30

模型定位
========
適合 production-quality 生圖與編修、長篇自然語言、寫實攝影、文字較多的廣告、資訊圖、多格構圖、透明商品素材、合成與身分敏感編修。可透過多輪對話一次只改一項。

API／介面設定
=============
quality：low、medium、high。
- low：大量發想、預覽與低延遲工作。
- medium：大多數商業迭代的平衡起點。
- high：小字、密集資訊圖、近距離人像、重要最終稿可比較使用。

size：gpt-image-2 支援符合條件的彈性 WIDTHxHEIGHT。
- 每邊必須是 16 的倍數。
- 長短邊比例不可超過 3:1。
- 總像素介於 655,360 與 8,294,400。
- 最長邊必須小於 3840px；超過 2560x1440 的輸出視為較具實驗性。
- 常用穩定尺寸：1024x1024、1024x1536、1536x1024、2560x1440。

透明背景：preview。
- background="transparent"
- output_format="png" 或 "webp"
- Prompt 明寫 isolated subject on a fully transparent background。

重要：size、quality、background、output_format 都是設定，不是畫面 Prompt 文字。
官方模型摘要指出 gpt-image-2 已預設高輸入保真，不應依賴 input_fidelity 參數改善結果。

Prompt 核心架構
===============
INTENT / USE：這張圖用在哪裡、要完成什麼商業任務。
SCENE / BACKGROUND：地點、時間、情境、環境真實資料。
SUBJECT：人物、商品、空間、車輛與不可變特徵。
COMPOSITION：比例、鏡位、位置、文案安全區。
LIGHT / MATERIAL：光向、色溫、材質、反射、接觸陰影。
TEXT：逐字文案、位置、字體感、層級、只出現一次。
PRESERVE：編修時保持不變的內容。
CHANGE：本輪唯一要改的內容。
CONSTRAINTS / QA：手部、透視、文字、品牌與真實性驗收。

通用 Master Template
====================
Create a [PHOTOREALISTIC PHOTOGRAPH / ILLUSTRATION / INFOGRAPHIC] for [INTENDED USE].

SCENE
[LOCATION, TIME, WEATHER, BACKGROUND AND CONTEXT]

SUBJECT
[EXACT SUBJECT, MATERIAL, AGE, ACTION, PRODUCT GEOMETRY OR REAL PROPERTY FACTS]

COMPOSITION
[ASPECT RATIO, SHOT SIZE, CAMERA ANGLE, PERSPECTIVE, SUBJECT PLACEMENT, COPY SAFE AREA]

LIGHT AND MATERIAL
[KEY LIGHT, FILL, PRACTICALS, COLOR TEMPERATURE, TEXTURE, REFLECTIONS, CONTACT SHADOWS]

EXACT TEXT
Render exactly once: "[EXACT_TEXT]"
[POSITION, TYPE STYLE, SIZE, COLOR, ALIGNMENT]

CONSTRAINTS
[PRESERVE LIST]. Keep typography legible, anatomy and perspective physically consistent, and all claims grounded in supplied facts. Use no extra text, logos or watermarks.

01｜企業人物寫實照
=================
Create a photorealistic professional portrait for a technology company About page.

SUBJECT
An authorized adult East Asian founder in their early 40s, preserving real age, facial proportions, skin tone and hairstyle from the supplied reference. They stand naturally beside a real office window, shoulders relaxed, one hand lightly resting on the table, calm direct eye contact and a restrained confident smile.

COMPOSITION
Medium portrait, eye-level 70mm perspective, person on the left third, right side kept visually quiet for the page headline, target ratio 4:5.

LIGHT AND FINISH
Soft overcast window light from camera-left with neutral fill. Authentic pores, fine age lines, under-eye texture, eyebrow hairs, fabric fibers and physically correct catchlights. Honest editorial retouching, balanced skin color, both eyes sharp, believable city reflections. Avoid glamorization, heavy smoothing, fake bokeh, invented logos, text and watermarks.

建議：1024x1280 或符合 4:5 且為 16 倍數的尺寸；quality=medium/high。

02｜電商白底商品主圖
===================
Create a photorealistic e-commerce catalog image of the exact supplied [PRODUCT].

PRESERVE
Keep the product silhouette, dimensions, cap, label layout, color, material, SKU markings and all visible package text unchanged.

SCENE AND COMPOSITION
Single product centered on a seamless neutral-white studio background, front three-quarter view at eye level, generous equal margins, target ratio 1:1.

LIGHT AND MATERIAL
Large softbox from upper-left, white frontal fill, narrow edge light separating the silhouette. Natural material texture, controlled highlights, one soft physically plausible contact shadow directly below. Clean production retouching while keeping edges and label sharp.

Do not redesign the package, add props, invent copy or duplicate the product. Output one clean catalog image ready for manual SKU verification.

03｜透明背景商品去背
===================
Extract the exact product from the supplied image and isolate it on a fully transparent background.

Preserve the product geometry, label, cap, material, color, proportions and every readable character exactly. Deliver one centered product with generous padding, clean alpha edges, crisp silhouette and natural internal reflections. Remove the original environment and its cast shadow. Keep only a very subtle object-internal shading that describes the form. Do not add a solid backdrop, checkerboard, scenery, new text, watermark, halo or fringing.

設定：background="transparent"；output_format="png" 或 "webp"；quality=medium。

04｜含精準文案的品牌廣告
=======================
Create a premium 4:5 social advertisement for [BRAND] featuring the exact supplied [PRODUCT].

LAYOUT
Product centered-right on a pale stone plinth. Upper-left contains the headline. Lower-left contains one short supporting line. Keep generous margins and a clear mobile reading hierarchy.

TEXT — EXACT, VERBATIM
Headline: "[EXACT_HEADLINE]"
Supporting line: "[EXACT_SUBHEAD]"
Render each line exactly once, with no extra characters. Use bold modern sans-serif for the headline and regular sans-serif for the supporting line, left aligned, high contrast.

VISUAL
Photorealistic studio product lighting, brand colors [HEX 1] and [HEX 2], accurate label, real contact shadow and controlled reflections. No unrelated logo, decorative pseudo-text or watermark. Leave final legal copy and official vector Logo for post-production.

05｜包裝翻譯／文字局部修改
========================
Edit only the specified text region in the supplied package image.

CHANGE
Replace the text "[ORIGINAL_TEXT]" with the exact Traditional Chinese text "[NEW_TEXT]". Match the original baseline, alignment, visual weight, color, print texture, curvature and perspective.

PRESERVE
Keep the package shape, dimensions, brand mark, all other text, barcode, color, lighting, surface texture, background, framing and shadow exactly unchanged. The replacement must appear once in the same region. Make no other visual changes.

After generation, provide an image suitable for human character-by-character verification; the result is not final legal packaging artwork.

06｜飯店室內訂房封面
===================
Create a photorealistic hospitality photograph of the verified [ROOM TYPE] at [HOTEL].

Use only these confirmed features: [LIST ACTUAL BED TYPE, WINDOW, FURNITURE, BATHROOM, VIEW AND AMENITIES]. Show a welcoming eye-level view from the entrance using a controlled 24mm architectural perspective with vertical lines straight. Bed and circulation path fully readable; target 16:9 with upper-right title safe area.

Warm 2700K bedside practicals balanced with soft blue-hour window light. Preserve real wood, stone, linen and upholstery texture; reflections correspond to visible sources. Styling feels freshly prepared but occupied by real life. Do not invent a balcony, ocean view, room size, furniture, amenity or architectural opening. Human QA must compare the image with the actual room checklist.

07｜建築外觀概念／實景分流
=========================
Create a photorealistic [ACTUAL PHOTOGRAPH / CLEARLY LABELED CONCEPT VISUALIZATION] of [PROJECT].

FACTS TO PRESERVE
[SITE, MASSING, NUMBER OF FLOORS, FAÇADE MATERIALS, OPENINGS, LANDSCAPE, ACCESS AND CURRENT COMPLETION STATUS].

COMPOSITION
Three-quarter street-level architectural view, controlled 28mm perspective, verticals straight, realistic neighboring context, pedestrians and vehicles used only for scale. Late-afternoon side light with accurate window and façade reflections.

Ensure the structure is physically buildable and all joints, railings, drainage, stairs and entrances are coherent. Do not add unapproved floors, amenities, signage, landscape, ocean or skyline. Reserve lower-right area for the later label "AI-assisted concept visualization" when applicable.

08｜汽車商業視覺
===============
Create a photorealistic automotive campaign image of the exact supplied [YEAR MAKE MODEL AND TRIM].

Preserve body proportions, grille, headlights, wheel design, tire profile, mirrors, panel lines, badge placement and paint color [COLOR CODE]. Place the vehicle on a dry coastal overlook at blue hour, three-quarter front view, camera at headlight height, 50mm perspective. Wheels remain circular and aligned; tires visibly contact the road. A large soft source traces the body while sky reflections follow each panel’s curvature. Empty roadway, subtle environmental reflections, left-side copy space, 16:9. Do not redesign the vehicle, change trim, invent accessories, duplicate badges or add text.

09｜餐廳菜單主視覺
=================
Create an appetizing photorealistic menu photograph of the real [DISH] using the supplied plating reference.

Preserve actual portion size, ingredient count, garnish, plate, sauce placement and cooking appearance. Three-quarter overhead angle, 70mm food perspective, dish centered lower-left with a calm menu-copy area upper-right, 4:5. Soft window key from left and gentle warm practical fill. Steam rises only from hot components; moist highlights and crispy textures remain realistic. Tableware has believable contact shadows. Do not increase portion, add premium ingredients, alter doneness or show items not included in the order.

10｜專業資訊圖
=============
Create a clean 16:9 classroom infographic explaining [TOPIC] to [AUDIENCE].

INFORMATION
Use only the following verified facts: [FACT LIST]. Organize them into [NUMBER] steps from left to right with one simple icon per step and clear connecting arrows.

TEXT
Title exactly: "[TITLE]"
Step labels exactly: "[STEP 1]", "[STEP 2]", "[STEP 3]", "[STEP 4]"
Render each text item once, large enough to read in a presentation. Use a white background, dark navy text, one accent color [HEX], flat vector-style icons, clear hierarchy and generous spacing.

Avoid tiny paragraphs, decorative pseudo-data and unverified claims. All wording and numerical relationships require human subject-matter review.

11｜多圖商品＋場景合成
=====================
Create one believable commercial photograph by compositing the supplied references.

SOURCE ROLES
Image 1: preserve as the base room, camera position and architecture.
Image 2: use the exact product, preserving geometry, label and material.
Image 3: use only as the styling palette for textiles and small movable props.

Place the product on the center table at realistic scale. Match Image 1’s lens perspective, light direction, white balance, occlusion, reflection and contact shadow. Keep all fixed architecture and major furniture from Image 1 unchanged. Apply Image 3 only to replace the specified cushions and vase. Add no other object, text or logo. The final image should look like one photograph captured in the original room.

12｜一次只修一項的迭代 Prompt
============================
Edit the current image with one change only.

CHANGE
[DESCRIBE ONE PRECISE CHANGE: move product 8% left / change background to #HEX / replace one line of text / soften the key light].

PRESERVE EXACTLY
[SUBJECT IDENTITY, PRODUCT GEOMETRY, CAMERA, CROP, POSE, WARDROBE, BACKGROUND, LIGHT DIRECTION, COLOR, TEXT, LOGO, SHADOWS — REMOVE THE ITEM THAT IS BEING CHANGED FROM THIS LIST].

Keep every unlisted visual detail pixel-consistent in intent. Do not reinterpret the design, add new elements or change framing. Return a single corrected version for side-by-side QA against the previous image.

失敗修復
========
文字不準：減少總字數；逐行加引號；寫 appears exactly once；先完成版面，再只改文字。
人物不自然：寫 real moment、authentic skin、worn materials、unposed；指定手有支撐與清楚動作。
商品變形：把 PRESERVE 放到前半段，逐項列 SKU、形狀、材質、標籤；以原圖編修而非重新生成。
改一處卻全圖漂移：使用第 12 組格式，每輪只有一個 CHANGE。
透明背景有白邊：明寫 clean alpha edges、no fringing / halo，並用 PNG 或 WebP。
尺寸錯誤：確認兩邊皆為 16 的倍數、比例與總像素在官方範圍內。

商業 QA
=======
[ ] 模型值使用 gpt-image-2；平台路由名稱沒有誤放進 OpenAI model 欄位
[ ] size、quality 與透明背景設定放在 API／介面設定
[ ] 人物參考為已授權成年人，沒有虛構背書
[ ] 文字、Logo、價格、數據逐字核對
[ ] 商品、車型、房型、建築與設施對照真實資料
[ ] 反射、手指、接觸點、透視與景深合理
[ ] AI 輔助／合成／概念示意依規定揭露

OpenAI 官方參考
================
https://developers.openai.com/cookbook/examples/multimodal/image-gen-models-prompting-guide
https://developers.openai.com/api/docs/guides/image-generation
https://developers.openai.com/api/docs/models/gpt-image-2

