GPT Image 2 提示词指南：结构、案例与样式控制

GPT Image 2 是 OmniArt 上的高控制力图片模型，适合同时涉及构图、可见文字、参考图和明确细节要求的 brief。它不保证一次生成就得到准确文字，结果仍需人工校对。

本指南提供可复用的六段式结构与 12 份可直接使用的 prompt。它们是经过筛选的起点，不代表受控 benchmark；为便于粘贴，prompt 在所有语言版本中保持英文。

OmniArt 上的 GPT Image 2

当前支持 1K、2K、4K，Low/Medium/High detail，最多 9 张输入图片和可复现 seed。所有套餐都能使用模型；更高 detail 与 2K 以上需要 Starter 或更高套餐。积分同时取决于分辨率与 detail，先用 1K Low 验证布局。

六段式提示词结构

按“交付物与风格 + 主体 + 环境 + 光线与材质 + 构图层级 + 精确约束”组织提示词。先写准确可见文字，为每张输入图指定身份、结构、姿势、配色或环境等唯一职责；固定 seed 后一次只修改一个问题。

六个有效习惯

1. 先写交付物类型；2. 用完整句说明关系；3. 给准确可见文字加引号；4. 每张参考图只负责一件事；5. 小改动时锁定 seed；6. 一次只修一个失败。

Prompt library：12 份可直接使用的 brief

以下 1–12 依次用于电商产品图、配色编辑、双语海报、角色转面图、房间改造、美食摄影、时尚主视觉、信息图、应用商店图、YouTube 缩略图、六格分镜和身份保持肖像。为便于直接粘贴，prompt 在所有语言版本中保持英文。

1

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.

2

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.

3

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.

4

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.

5

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.

6

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.

7

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.

8

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.

9

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.

10

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.

11

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.

12

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.

检查时优先确认身份与结构、准确文字、画面层级、参考素材边界和可用留白；每次只调整一个变量。

样式控制：什么有效，什么无效

GPT Image 2 对自然语言式的样式指令的接受度，远比关键词堆叠更好。三种能稳定生效的写法。

两种行不通的写法。

• 堆叠多个风格形容词（"dreamy ethereal cinematic photoreal hyperrealistic"）。模型会把它们平均成一锅糊。
• 要求精确的品牌 logo。logo 复刻并不可靠，logo 请在后期合成。

使用参考图片

最多 9 张输入图只有在职责明确时才有价值。可让 image 1 负责产品结构、image 2 负责构图、image 3 负责材质或配色。编辑时先列出必须保持不变的内容，再描述变化。

真实限制

可见文字仍需逐字校对；精确 logo 与受监管包装文字应在后期合成；多视图可能在身份、配件或结构上漂移；更高分辨率和 detail 不能修复含糊 brief；seed 有助于对比，但不保证提示词或输入变化后的像素一致。

开始使用 OmniArt

打开 GPT Image 2，从 1K Low 开始，选择下方一个 prompt。锁定 seed，一次修改一个变量，待构图、保持项和文字稳定后再提高 detail 或分辨率。更多起点可浏览 prompt gallery。