GPT Image 2 优质提示词汇总

2026/05/01

GPT Image 2 是 OpenAI 于 2026 年 4 月 21 日发布的最新图像生成模型。它的核心升级包括:原生 2K 分辨率、多语言文字渲染准确率超过 95%、以及在绘制前先进行构图规划的推理管线。单帧可稳定承载 5–8 个命名元素,支持从 3:1 到 1:3 的多种宽高比,排版渲染精度超越了此前所有图像模型。

本文整理了社区中最优质的 GPT Image 2 英文提示词,按类别编排,可直接复制、修改和使用。

提示词六要素

GPT Image 2 对有结构的提示词响应更好。按照清晰的指令序列编写的提示词,效果远优于自由形式的描述。每一条高质量提示词都包含以下六个要素,按此顺序排列:

  1. 场景 / 背景 — 画面发生在哪里?
  2. 主体 — 包含比例和视线方向,确定谁是画面焦点。
  3. 材质与纹理 — 布料、金属、玻璃、皮肤,用来描述物体是什么材质的。
  4. 构图 — 取景方式、视角、焦距、元素位置。
  5. 光影与氛围 — 光线的方向、质感、色温、情绪基调。
  6. 约束条件 — 用引号标出需渲染的文字,哪些必须保持不变?哪些绝对不能出现?

两条额外规则:需要用引号将画面中的文字包裹起来;想要照片级真实感时,明确使用 "photorealistic" 一词。类似 "8K, ultra-detailed, masterpiece" 这种通用风格标签是早期扩散模型的遗留习惯——GPT Image 2 基本上会忽略它们。所以把这部分提示词预算留给光影、构图和约束条件的更好的选择。


各场景提示词示例


1. 文字排版与字体设计

这是 GPT Image 2 的王牌能力。此前所有图像模型在文字渲染上都存在缺陷——字形残缺、字距错乱、非拉丁文字变成乱码。GPT Image 2 渲染排版的效果堪比设计师在 InDesign 中排出来的。

极简排版海报

Create a premium minimal poster for a design conference. Use the exact headline "Future Type Lab" in large clean sans-serif typography, the subtitle "Letters, Layouts, and AI Systems" below it, crisp alignment, generous spacing, black text on warm white paper, a subtle embossed texture, and a single abstract geometric shape behind the title without covering any words.

咖啡包装标签

Design a realistic matte coffee bag standing on a stone counter with a readable front label. The label should say "North Peak Coffee" as the brand, "Single Origin Ethiopia" as the product line, and "Notes: citrus, honey, jasmine" in smaller text. Use premium packaging photography, soft morning light, accurate label hierarchy, and keep all text straight and legible.

餐厅菜单看板

Create a cozy cafe menu board photographed behind a counter. The board should have three clear sections titled "Espresso", "Tea", and "Bakery", with neat short item names under each section. Use hand-lettered but readable white text on a dark green board, warm pendant lighting, shallow depth of field, and keep every section title spelled correctly.

杂志封面多级标题

Create a realistic editorial magazine cover about future workspace design. The masthead should read "WORKFLOW" at the top, the main cover line should read "The New Creative Desk", and three smaller blurbs should read "Lighting", "Focus", and "Tools". Use a clean portrait of a modern desk setup, readable magazine typography, accurate hierarchy, and no misspelled filler text.

电影感 3D 文字

"EXPLORE" in giant 3D-text placed across a sea bridge, with striking cinematic lighting, photorealistic ocean and rock textures, golden hour backlight, and dramatic shadows.

多语言文字海报(日文 + 英文)

A 1960s Japanese National Railways travel poster style. Bold kanji heading "京都の秋" centered, English subtitle "Autumn in Kyoto" below, a vermilion temple pagoda surrounded by crimson maple leaves, woodblock print texture, warm vintage paper tone, clean typography hierarchy with both scripts perfectly legible.

文字提示词技巧:

  • 用直引号包裹需要精确渲染的文字——模型将引号内的字符串视为精确渲染目标。
  • 对于品牌名称或不常见的拼写,逐字母拼出。
  • 单行文字控制在 6 个单词以内效果最佳。明确指定字体颜色和样式。
  • 末尾添加约束:"verbatim, perfectly legible, no extra text, no duplicate text"。
  • 极小文字(如表盘、细则)单条控制在 15 个字符以内。

2. 电影级场景

GPT Image 2 理解镜头语言、光比和场景设计。描述光的物理特性,而非仅仅描述视觉效果。

霓虹雨夜街头

Create a cinematic night scene of a lone courier crossing a rain-soaked street in a dense futuristic city. Use wet asphalt reflections, neon signage glow, a 35mm film look, volumetric mist, teal and amber color contrast, low camera angle, dramatic backlighting, and a sense of quiet motion without making the scene look chaotic.

黎明时分的沙漠观测站

Generate a wide cinematic shot of a research observatory in a silent desert at dawn. Show a small team preparing equipment near a white dome structure, warm sunrise light hitting red sand, long shadows, realistic atmospheric haze, a restrained color palette, and a composition that feels like a science-fiction film still.

静谧的车窗

Generate a cinematic portrait-oriented scene of a traveler sitting beside a train window at sunrise. Reflections of passing fields appear in the glass, soft golden light touches the seat fabric, the mood is thoughtful and calm, the framing is intimate, and the image should look like a still from an independent film.

科幻控制室

Create a cinematic control room scene with a practical sci-fi console, translucent display panels, and an engineer studying a live system map. Use believable production design, soft blue screen light on the face, dark background depth, precise reflections, and a grounded near-future aesthetic.

蓝调时刻的山地救援

Create a dramatic mountain rescue scene at blue hour. Show a helicopter spotlight sweeping across a snowy ridge, two rescuers in bright technical jackets, wind-blown snow, realistic scale, high contrast lighting, a telephoto lens feel, and a serious documentary film style rather than fantasy action.

3. 产品摄影

适用于主视觉图、电商创意、包装概念和高端广告视觉。

精华液产品主视觉

Create a premium skincare product photo of a frosted glass serum bottle on a pale stone platform. Use soft diffused studio lighting, subtle water droplets, botanical shadows on the background, clean negative space, realistic reflections, and an upscale beauty brand look suitable for a landing page hero image.

跑鞋广告

Generate a dynamic product photography shot of a lightweight running shoe suspended above a textured track surface. Add small dust particles, crisp side lighting, sharp material detail, a deep blue background, realistic shadow contact, and a polished sports campaign style without adding logos or brand names.

耳机细节特写

Create a close-up commercial product photo of wireless over-ear headphones resting on a brushed metal surface. Emphasize fabric texture, hinge detail, soft rim lighting, premium black and graphite tones, realistic reflections, and a composition that leaves clean space for web copy on the right.

巧克力包装组合

Create a premium product arrangement of three artisan chocolate bars in textured paper sleeves on a marble surface. Use readable but fictional package names, warm side lighting, small cacao nibs as props, careful alignment, realistic packaging folds, and a refined editorial food photography style.

电商白底图

A stainless steel water bottle, isolated on pure white background, no shadows extending beyond product base, studio lighting, 3/4 angle, sharp product edges, realistic condensation droplets. E-commerce product photography style, no logos, no watermarks.

4. UI/UX 与 App 原型

GPT Image 2 能生成截图级的 App 原型,标签文字清晰可读,UI 层次结构正确,间距真实自然。

理财 App 仪表盘

Create a realistic mobile finance app dashboard screenshot. Include a balance card, spending chart, three transaction rows, and a bottom navigation bar. Use readable labels such as "Overview", "Budget", and "Savings", a calm white and blue interface, clean spacing, rounded cards, and a polished SaaS product design style.

旅行规划器引导页

Design a mobile onboarding screen for a travel planner app. Show the headline "Plan smarter trips", a simple destination card, a map preview, a primary button labeled "Start planning", soft sky colors, readable UI text, consistent margins, and screenshot-like realism without browser chrome.

数据分析 Web App

Create a desktop web app analytics dashboard for a subscription product. Include a left sidebar, top filter bar, revenue line chart, conversion funnel card, retention table, and readable labels. Use a restrained professional color palette, dense but scannable layout, and high-fidelity product screenshot styling.

AI 写作工具落地页

Create a clean web landing page mockup for an AI writing tool. Include a hero headline, short subheading, two call-to-action buttons, a prompt editor preview, and three feature cards. Use readable text, modern SaaS spacing, subtle shadows, and a credible product marketing layout.

带 AI 助手的效率 App

A modern smartphone with dynamic island, light mode, soft blue accents, soft drop shadows. Header: small circular photo, greeting "Good morning, Alex ☀️", subtitle "Let's make today productive.", notification bell with red dot. Hero card: "TODAY'S FOCUS — Launch marketing campaign", circular progress ring at 75%. Up Next section with one item. Reminders section with two items. Bottom: pill-shaped AI input field "Ask your AI assistant..." and 4-tab navigation bar (Home, Tasks, Assistant, Profile).

UI 提示词技巧:

  • 明确指定 UI 层级结构:"Top navigation with 4 icons, below that a search bar reading 'Search recipes...'"。
  • 模糊的提示词会产生毫无意义的通用结果。明确写出元素位置:"logo top-right, headline centered, CTA bottom-left."。
  • 迭代优化优于一步到位:第一轮定布局,第二轮调颜色,第三轮改文案。

5. 人像与摄影

GPT Image 2 能处理身份敏感的图像编辑,生成的人像避免了此前模型常见的塑料感、过度磨皮效果。

创业者肖像(编辑风格)

Create a realistic editorial portrait of a startup founder in a quiet studio office. Use natural window light, a navy blazer, relaxed confident expression, clean background shelves, shallow depth of field, sharp eye detail, and a professional magazine interview style.

时尚棚拍人像

Create a high-fashion studio portrait of a model wearing a structured cream jacket and silver accessories. Use a smooth gray backdrop, directional softbox lighting, precise shadows, elegant posture, detailed fabric texture, and a refined editorial campaign mood.

户外的黄金时刻

Generate a natural outdoor portrait of a creative professional standing on a quiet city rooftop at golden hour. Use warm rim light, soft wind in the hair, realistic clothing folds, an urban skyline blurred in the background, and a composed lifestyle photography style.

电影感角色研究

A highly detailed, cinematic portrait of a young woman with dark hair in an updo, wearing a futuristic black trench coat with silver metallic accents, tactical belts, and a holstered sci-fi pistol. She has a serious, focused expression and is interacting with a glowing blue holographic display. The scene is inside a spaceship command center with cool blue lighting. Through a large window, a starry space environment with a fleet of spaceships is visible. Photorealistic textures, dramatic lighting, high-tech sci-fi aesthetic.

音乐人专辑写真

Create a moody portrait of an independent musician sitting beside a vintage keyboard in a small recording room. Use low warm light, soft film grain, rich shadows, relaxed posture, detailed hands and instrument keys, and an intimate album press photo aesthetic.

真实感秘诀: 避免使用 "perfect skin"、"flawless"、"professional retouching" 这类词汇——它们会产生千篇一律的 AI 人像效果。用真实摄影的语言替代:"visible pores"、"fine lines"、"asymmetry"、"available light"、"no heavy retouching"。


6. 信息图与教育视觉

3D 进化信息图

A 3D stone staircase ascending from left to right, each step carved with a different technological era label: "Stone Tools" → "Bronze" → "Iron" → "Steam" → "Electricity" → "Computing" → "AI". Ancient artifacts on lower steps, microchips and holograms on upper steps. Isometric view, warm museum lighting, realistic stone texture, clean readable labels.

手绘风格城市美食地图

An illustrated tourist map of Osaka's food districts. Hand-drawn watercolor style, neighborhood boundaries in thin ink lines, iconic dishes illustrated at their famous locations (takoyaki in Dotonbori, kushikatsu in Shinsekai), compass rose in the corner, readable bilingual labels in Japanese and English, vintage map aesthetic with warm cream paper.

前后对比图

A split-screen comparison: left side shows a dated 1990s living room with floral wallpaper and heavy curtains; right side shows the same room renovated with minimalist Scandinavian design, light oak floors, and floor-to-ceiling windows. Keep room dimensions, camera angle, and lighting direction identical on both sides. Architectural before/after photography style.

物理公式参考表

A clean physics reference sheet on dark chalkboard background. Title "Snell's Law & Refraction" in bold serif at top. Three diagrams: incident ray refracting through water surface, prism dispersion with rainbow spectrum, and total internal reflection in optical fiber. Each diagram has proper Greek notation (θᵢ = θᵣ, n = c/v), labeled angles, and a one-line caption. Footer reads "Geometric Optics — Principles". Hand-drawn chalk diagram aesthetic with crisp white lines.

信息图注意事项: 永远不要信任图中的数字。GPT Image 2 会编造统计数据和数据点。用它来生成视觉结构和布局,然后在 Figma 或 Canva 中添加最终的数据标签。数据点控制在 5–7 个以内——超过这个数量标签就会开始重叠。


7. 角色设计与一致性

角色参考表(五视图)

A professional character reference sheet showing the same female warrior in five views: front, three-quarter, side profile, back, and action pose. Red-haired woman, age 25, green jacket, round glasses, short practical haircut. Consistent facial features, outfit details, and proportions across all views. Clean studio lighting, neutral gray background, character design sheet format with annotation lines for height.

多格漫画一致性

A 4-panel vertical comic strip featuring a red-haired woman in a green jacket. Panel 1: She discovers a mysterious glowing package on her doorstep (wide shot, afternoon light). Panel 2: Close-up of her hesitant expression as she reaches for it. Panel 3: She opens it — a burst of golden light illuminates her face. Panel 4: She holds up a small floating crystal, expression shifting from shock to wonder. Keep character identical across all panels: same face, same hairstyle, same green jacket, same round glasses. Comic book illustration style, clean ink lines, flat colors.

一致性法则: 跨提示词使用完全相同的角色描述,逐字复制粘贴。在多轮迭代中,每次都重新声明身份信息:"same face, same outfit, same lighting — only change the background."。模型不会记住两轮之前你关心的内容,你必须反复告诉它。


8. 游戏与娱乐截图

神话战斗场景

A hyper-realistic epic fantasy digital painting of Sun Wukong, the Monkey King, in mid-combat, wearing ornate golden and red armor with long red headpiece feathers. He fiercely strikes down armored celestial soldiers using a glowing golden staff. The dynamic scene features flying sparks, shattered debris, and flowing fabric. The background is a heavenly realm with ornate pillars and clouds, showing distant warriors on grand staircases. Cinematic lighting, intense action, and dramatic particle effects.

动漫武术对决

A dynamic anime-style illustration of two martial artists clashing mid-air in a bamboo forest at sunset. Motion blur on flowing fabric and hair, impact shockwave radiating outward, warm orange and cool blue color contrast, painted background with depth, 2D animation keyframe aesthetic.

等距视角 RPG 小镇

An isometric view of a medieval fantasy town square at dusk. Cobblestone plaza with a central fountain, timber-framed buildings with glowing windows, market stalls with fabric awnings, tiny NPC figures going about their day. Warm lantern light, soft shadows, game asset concept art style, 3/4 isometric perspective, rich environmental detail.

VR 头显爆炸图

An exploded view technical poster of a VR headset. All components — lenses, display panels, sensors, straps, circuit boards, outer shell — floating in precise alignment with thin connecting lines showing assembly order. Each part labeled with small clean sans-serif text. Dark technical background, studio lighting, product design presentation style. No brand logos.

9. 图像编辑与风格迁移

Logo 变体网格

Transform this logo into a grid of minimalist logo variations using the main subject as the core icon. Create 16–20 unique vector-style logo marks. Each variation should reinterpret the same subject in different ways. Arrange evenly on a light background. Keep designs clean, modern, with balanced spacing. Maintain consistency while exploring creative variations.

角色成长序列

Show the evolution of a character from child to elder in 6 stages, left to right. Same person, consistent facial structure and identifying features across all ages. Each stage in a different season of life, matching environment and lighting. Photo-realistic, consistent lighting direction, seamless aging progression.

物件蜕变序列

A horizontal sequence of 5 images showing a raw wooden log transforming into a finished acoustic guitar. Stage 1: Raw log. Stage 2: Rough-cut body shape. Stage 3: Sanded and assembled body. Stage 4: Stained and lacquered. Stage 5: Finished guitar with strings. Consistent lighting, same angle, same distance, workshop background.

编辑工作流法则: 编辑时,同时声明要改什么和保留什么。使用这样的措辞:"keep [X, Y, Z] identical + change only [A] + repeat the invariant list."。这能大幅减少未指定元素上的漂移。


10. 营销与品牌创意

电商主视觉海报(护肤品)

A high-end e-commerce hero poster for a ceramide repair serum. Style: Clean, light luxury, strong product focus. Center: frosted glass bottle with droplet reflections. Background: off-white to warm gray gradient with subtle molecular structure decorations. Include clearly readable copy: product name, tagline "Barrier repair, soothe redness, radiant skin", formula generation, key ingredients (ceramide, panthenol B5, centella asiatica), suitable skin types (sensitive, sleep-deprived, seasonal-instability skin), limited-time price, and gift bundle details. Fine print: "Individual results may vary, use consistently." Overall must be high-end, not tacky.

Instagram 四宫格广告

Generate 4 coherent Instagram panels for a sustainable sneaker brand launch. Panel 1: Hero product shot on natural linen. Panel 2: Close-up of recycled material texture. Panel 3: Lifestyle shot — runner on coastal trail, golden hour. Panel 4: Launch details card with date, price, and CTA. Consistent warm neutral palette, same brand aesthetic across all panels, clean negative space for text overlays. No brand logos — leave space for compositing.

深夜电音派对传单

Generate a modern event flyer for a late-night electronic music show. The main title must read "MIDNIGHT SIGNAL" in bold condensed lettering, with "Friday 11 PM" and "Warehouse Room 4" as smaller supporting text. Use electric blue and white type on a dark charcoal background, high contrast, clear margins, and a realistic printed flyer texture.

品牌建议: 用于客户交付的品牌设计时,先生成不带 Logo 的版式(在正确位置留白),然后在 Figma 或 Photoshop 中合成实际的 Logo SVG。这对生产级品牌设计是不可妥协的流程——AI 生成的品牌 Logo 可能存在知识产权问题。


常见错误

1. 堆砌形容词而非陈述视觉事实。

比如 "Beautiful, stunning, gorgeous sunset",没有给模型任何有效信息。

正确的方式应该是像 "Golden hour, 85mm lens, warm rim light, cirrus clouds at 30% coverage" 这样,给了模型可以规划的素材。

2. 迭代中丢失不变约束。

在多轮对话中优化时,每次都要重新声明哪些部分必须保持不变:"same face, same outfit, same lighting — only change the background."。否则模型会认为所有元素都可以随意改动。

3. 遗漏构图指令。

同一主体的"特写"和"广角"是完全不同的画面。每次都明确指定角度、距离和取景方式。

4. 忘记负面约束。

提示词末尾没有 "no watermark, no extra people, no logo drift",模型就会在你不需要的方向上即兴发挥。

不得不说的是,有些平台哪怕加了 "no watermark" 这种提示词,其实还是会有水印。

5. 使用 Midjourney 式的关键词堆砌。

逗号分隔的关键词链式提示词虽然也勉强能用,但无法发挥 GPT Image 2 的优势。模型对会话式描述的理解远优于关键词列表,所以用自然语言些提示词会更好。

6. 要求精确计数。

比如 "Generate exactly 47 trees" 这种大概率会失败。使用定性描述可能会更好,比如 "a dense forest"、"a small group of",如果数量确实重要,在后期处理中手动调整。换句话说对于需要精确计数的生成场景,可能并不能一步到位,还需要专业工具做调整。


专业工作流

  1. 从中等质量、1024×1024 方形画布开始。

先生成两张校准提示词效果,确认无误后再切换到高质量和非方形比例输出最终成品。

  1. 迭代而非一步到位。

第一轮:确定构图和主体。

第二轮:调整光影或氛围。

第三轮:细化某个细节或添加文字。

  1. 优先使用编辑而非重新生成。

需要修改时,用自然语言指令编辑现有图像,而不是从头生成。重新生成是生产工作中品牌风格漂移的最大原因。

  1. 合成输出成品。

先以不包含 Logo 或数据的形式生成版式布局,然后在 Figma 或 Photoshop 中合成实际素材。这是几乎所有客户交付级品牌或数据视觉的标准工作流。

  1. 短提示词 vs 长提示词。

短提示词(30 词以下)适用于有明确概念并希望模型发挥创意的情况。

长结构提示词(提示词六要素)适用于需要精确构图、指定光影或精确排版的情况。

最佳区间:30–60 词。少于 15 词结果不可预测,超过 80 词则开始被忽略。

  1. GPT Image 2 在同一对话中会记住上一张生成结果。

你可以用简短的后续消息进行优化:"warm up the sky"、"move the cup to the left third"、"make her expression more relaxed"。

模型会在原有基础上做针对性调整,而非从头开始。


参考资料

Simple AI Box

Simple AI Box