GPT Image 2 提示詞指南：結構、範例與風格控制

GPT Image 2 是 OmniArt 上高控制度的圖片模型，適合那種同時要處理構圖、可見文字、參考圖和明確細節要求的案子。真正有用的做法不是把形容詞一直堆進提示詞裡，而是照著一個可以逐項驗收的順序，把主體、層級、場景、光線、構圖和交付限制講清楚。

這份指南給你一套可重複使用的六段式結構，加上 12 組可以直接貼上的提示詞。它們是精選過的起點，不是什麼受控測試的結論。貼進模型的時候提示詞保持英文，只依照你的案子改掉具體細節就好。

OmniArt 上的 GPT Image 2

點數同時看解析度和畫質，所以請挑能回答眼前這個問題的最便宜設定。低畫質的 1K 草稿就足以測版面；高畫質的 4K 生成應該等到構圖和文字都穩定之後再跑。

六段式提示詞結構

照這個順序寫：

[交付物與風格] + [主體] + [環境] + [光線與材質] + [構圖與層級] + [明確限制]

一個精簡的例子：

Editorial product photograph of a matte lilac wireless speaker on a warm ivory plinth. Soft window light from camera left, subtle coral bounce, crisp fabric texture. Three-quarter view, centered vertically with generous negative space on the right. 4:5 aspect ratio. No hands, no extra products, no visible logo, no border.

這個結構讓失敗很好診斷。產品不對，就修主體或參考圖的角色。版面不對，就修構圖和層級。字體排版不對，先把要顯示的文案縮短，再去加風格描述。

六個值回票價的習慣

1. 先講交付物是什麼。 海報、產品去背圖、角色設定表、分鏡、App Store 宣傳圖，各自隱含不同的版面規則。
2. 用完整句子描述關係。 說清楚什麼東西在什麼的上面、後面或旁邊。
3. 把要顯示的文案原文引出來。 保持簡短、保留大小寫，並且說明它該出現在哪裡。
4. 每一張輸入圖都指派一個角色。 識別、產品幾何、姿勢、色盤、環境，是各自獨立的工作。
5. 鎖住 Seed 再做受控的文案調整。 只有在構圖本身需要換方向時才換 Seed。
6. 一次只修一個問題。 整篇重寫會毀掉這次和上次之間的可比性。

提示詞庫：12 組可直接貼上的簡報

下面的提示詞在這篇文章的每一個語言版本裡都保持英文，這樣你可以原封不動貼上。

1. 電商產品去背圖

Premium ecommerce packshot of a matte lilac wireless speaker, exact front three-quarter view, centered on a warm ivory seamless background. Soft large-box lighting from upper left, controlled coral rim light, crisp woven fabric and button details, realistic contact shadow. Leave clean space above and below for marketplace cropping. No hands, no props, no logo, no text, no border. Aspect ratio 4:5.

檢查重點： 輪廓、材質紋理、按鍵位置、接地陰影，還有邊緣乾不乾淨。

2. 產品配色改版

Use image 1 as the exact product-geometry reference. Preserve shape, proportions, seams, buttons, camera angle, lighting, background, and shadow. Change only the fabric color from charcoal to soft lilac and the metal controls to brushed warm silver. Do not add or remove any component. Aspect ratio 1:1.

檢查重點： 是不是只有指定的材質變了，其他結構細節全部沒動。

3. 雙語活動海報

Contemporary bilingual event poster on a warm off-white textured background. Large headline at the top reads exactly “FUTURE MADE HUMAN”. Directly below, smaller Chinese text reads exactly “未来，由人创造”. Add the date “18 OCT 2026” and location “SHANGHAI” in a clean information block at the bottom left. One abstract lilac ribbon crosses the center with a single coral accent. Strong grid, generous negative space, sharp legible typography, no extra words, no logo. Aspect ratio 2:3.

檢查重點： 文案是否一字不差、兩種語言有沒有分開、層級、留白，以及有沒有多冒出來的字。

4. 角色三視圖設定表

Professional character turnaround sheet for one original fantasy courier with short silver hair, violet eyes, a dark travel cloak, and a lilac satchel. Show front, left side, back, and three-quarter views of the same character with identical clothing and proportions. Add four facial expressions: neutral, smiling, concerned, determined. Clean warm-white background, organized concept-art grid, small unlabeled color swatches, no prose text. Aspect ratio 16:9.

檢查重點： 每一個視角的識別和服裝是否一致，而不是只看主視角那一張。

5. 依參考圖重新設計房間

Use image 1 as the exact room-layout reference. Preserve walls, windows, doors, camera position, floor plan, and natural light direction. Redesign the room as a calm contemporary reading studio with warm ivory walls, pale oak shelving, one soft lilac lounge chair, linen curtains, and a single coral ceramic lamp. Remove clutter but do not change the architecture. Photoreal interior photography. Aspect ratio follows image 1.

檢查重點： 建築結構有沒有守住、比例可不可信，以及改造有沒有偷偷移動結構元素。

6. 食物專題攝影

Editorial food photograph of handmade citrus tart on a cream ceramic plate, one slice removed to reveal clean layers. Morning window light from camera right, soft shadows, subtle crumbs, linen napkin, and one sliced blood orange as a supporting prop. Three-quarter overhead composition with restrained styling and natural color. No utensils crossing the tart, no hands, no text. Aspect ratio 4:5.

檢查重點： 食物質感、切面可不可信、配件有沒有節制，還有顏色準不準。

7. 時尚廣告主視覺

High-fashion campaign image of one model wearing a sculptural plum coat in a minimal warm-violet studio. Full-body pose with the coat silhouette clearly separated from the background. Hard key light from high camera left, soft reflected fill, subtle floor reflection. Leave the upper-right quadrant empty for campaign copy. No logo, no jewelry, no extra people, no text. Aspect ratio 9:16.

檢查重點： 服裝的結構、手、全身比例，以及留白夠不夠放文案。

8. 資料資訊圖表

Clean editorial infographic titled exactly “HOW A FIVE-SECOND CLIP GETS MADE”. Show five numbered stages in a left-to-right flow: “1 BRIEF”, “2 REFERENCES”, “3 GENERATE”, “4 REVIEW”, “5 EDIT”. Use simple geometric icons, a warm ivory background, lilac primary accents, and one coral highlight on stage 4. Strong grid, large readable labels, no paragraph text, no extra stages. Aspect ratio 16:9.

檢查重點： 階段順序、標籤有沒有一字不差、字級讀不讀得清楚，以及流程方向明不明確。

9. App Store 宣傳圖

App-store feature graphic for a fictional creative app named “MOSAIC”. Center one realistic phone showing a clean image-generation interface. Large headline reads exactly “MAKE THE FIRST FRAME”. Smaller line reads exactly “Image, video, and sound in one brief.” Warm ivory background, soft lilac gradient shapes, one coral call-to-action pill with the text “START CREATING”. Premium product design, no ratings, no store badges, no extra copy. Aspect ratio 16:9.

檢查重點： 文案是否一字不差、手機的透視、介面合不合理，以及縮到縮圖尺寸時層級還在不在。

10. YouTube 縮圖概念稿

High-contrast YouTube thumbnail showing a split-screen transformation of the same small home office. Left side is cluttered and dim; right side is bright, organized, and warm. One surprised creator is centered across the split with consistent identity. Add only the large text “AI REDESIGNED IT” in the upper left, fully inside safe margins. Bold readable composition, no logo, no arrows, no extra words. Aspect ratio 16:9.

檢查重點： 分割線兩邊是不是同一個人、前後對比一眼看不看得懂，還有文字縮小之後讀不讀得到。

11. 六格分鏡

Six-frame storyboard for a 10-second product film about a compact wireless speaker. Frame 1: quiet desk at dawn. Frame 2: speaker powers on. Frame 3: close-up of fabric texture. Frame 4: camera circles as music begins. Frame 5: friends enter the room. Frame 6: clean product end frame. Keep the same speaker geometry and room in every panel. Cinematic grayscale storyboard drawing with small frame numbers only, no captions. Aspect ratio 16:9.

檢查重點： 動作有沒有照時間順序、產品幾何有沒有重複一致、運鏡有沒有變化，以及結尾那格能不能用。

12. 保留識別的人像改圖

Use image 1 as the exact identity and facial-geometry reference. Preserve face shape, eyes, nose, mouth, skin tone, age, hairstyle, and expression. Change only the wardrobe to a tailored cream jacket and the background to a softly lit warm-violet studio. Keep the original camera angle and crop. Natural skin texture, realistic fabric, no beauty retouching, no jewelry, no text. Aspect ratio follows image 1.

檢查重點： 先看識別、再看風格，然後是服裝的邊界、髮絲邊緣和背景的分離。

不靠堆形容詞的風格控制

風格描述寫成可觀察的選擇時最有效：

不要把近義詞疊成一串，像是 “dreamy, ethereal, cinematic, photoreal, hyperreal.” 這些字並沒有說明哪一個視覺決定才是重點。

怎麼用參考圖

最多 9 張圖片輸入，只有在角色講明白的時候才有用。一組實用的搭配可能是：圖 1 給產品幾何、圖 2 給鏡頭構圖、圖 3 給材質或色盤。如果兩張參考圖對同一個屬性有衝突，就拿掉一張，或者講清楚哪一張說了算。

要改圖的話，先列出哪些東西不能動，再描述要改什麼。先用低成本設定跑，等到保留的部分穩了，再拉高解析度或畫質。

誠實的限制

• 可見文字還是要校對；一次就拼對所有字，模型不保證做得到。
• 精確的標誌和有法規要求的包裝文案，應該在生成之後用確定性的素材合成上去。
• 多視角之間，識別、小配件或幾何仍然可能走鐘。
• 更高的解析度和畫質會提高點數成本，但修不好一份含糊的簡報。
• 可重現的 Seed 有助於做受控比較；它不保證你改了提示詞或輸入之後還能得到像素完全相同的結果。

在 OmniArt 上開始

打開 OmniArt 上的 GPT Image 2，從 1K 加低畫質開始，從提示詞庫裡挑一組來用。鎖住 Seed，一次只改一個變數。等到構圖、保留和文案都穩定了，再往上調畫質或解析度。想要更多可重複使用的起點，就去逛提示詞圖庫。