Purpose
When the user needs to use GPT-Image-2 to generatenon-text-only postershigh-quality images, use this skill.
Applicable scenarios: realistic photography, product mockups, App UI screenshots, cinematic quality, Pixar 3D characters, infographics, Logo, comic strips, 360° panoramas, isometric miniature scenes, action figures, retro trading cards, creative concept art, etc.
Not applicable scenarios: text posters centered on typographic aesthetics → use image2-typography-poster skill。
I. GPT-Image-2 Core Capabilities Quick Reference
| Capability | Performance | Prompt suggestion |
|---|---|---|
| Text rendering | Latin/Chinese/Japanese/Korean/Arabic ≥95% | Key text 1-5 words, wrapped in English quotation marks |
| Multi-element composition | Stably carries 150+ elements in a single image | Group with numbering or lists |
| Face consistency | persistent embedding maintained across images | Fix description of age/race/features/clothing |
| Physical materials | Metal reflections, wet surfaces, glass refraction correct | Explicitly mention material names and light sources |
| Edit mode | Original image + edit endpoint for precise local adjustments | "preserve everything else" lock |
| World knowledge | Built-in reasoning, understands real events/scenes | Can imply context |
| 4K output | Native 4K + custom dimensions | Automatically scales when exceeding pixel budget |
II. 5 Iron Rules of Prompt Writing
⚠️ These 5 are OpenAI official + community consensus,must be followed every time a prompt is written。
Iron Rule 1: Subject first
Place the core subject at the beginning of the prompt. The model gives the highest weight to the first 30% of the text. → The subject will not be drowned out by environmental descriptions.
Iron Rule 2: Structured scene
Order:Scene → Subject + Action → Visual Style/Medium → Camera Parameters → Lighting → Composition → Constraints → Complex compositions will not lose elements.
Iron Rule 3: Use English quotation marks for text
All text that should appear in the image should use "..." wrapped.
→ Text rendering success rate goes from 70% → 95%+.
Iron Rule 4: Specify camera and lighting
Specify concrete parameters: focal length (24mm/35mm/50mm/85mm/100mm macro), aperture (f/1.4/f/2/f/4), angle (eye-level/low-angle/bird's-eye), color temperature (3200K warm/5600K daylight). → gpt-image-2 integrates a physics engine, and camera parameters have real effects.
Iron Rule 5: When editing, separate "what changes and what does not"
When modifying an image, clearly divide into:Change: ... / Preserve: ...
→ Repeat the invariants list every time you edit to prevent drift.
III. Key differences between gpt-image-2 and older versions
| Dimension | 1.5 Era Approach | 2.0 Correct Approach |
|---|---|---|
| Keyword Stacking | "8K, ultra detailed, masterpiece" must | ❌ Completely ineffective, wastes semantic space |
| Negative Prompts | "no text, no watermark" | ✅ Switch to positive constraint sentences |
| Text Rendering | Limited to 1-2 words, error-prone | Supports 3-5 words, multi-line short sentences |
| Camera Description | Optional decoration | Strongly recommended, with physical effects |
| Editing Mode | Regenerate | Prioritize edit endpoint |
💡 Migration tip: about 70% of old prompts only need redundant adjectives removed to get better results.
IV. Prompt Structure Formula
Universal Structure (required for every generation)
[场景描述，1-2句]
[主体描述 + 动作/状态，2-3句]
[视觉风格 / 介质，1句]
[镜头参数：焦距 + 光圈 + 角度，1句]
[光线描述：方向 + 色温 + 质感，1-2句]
[构图：框架 + 前景/中景/背景，1句]
[约束：排除项 / 不变量，1-2句]
Style: [风格标签]
Style Tag Quick Reference
| Tag | Style | Best For |
|---|---|---|
editorial-magazine |
Magazine Layout | Posters, UI |
studio-product |
Studio Product Shot | Product Packaging |
cinematic-anamorphic |
Anamorphic Widescreen | Cinematic Texture |
pixar-3d |
Pixar 3D | Characters, Mascots |
kodak-portra-400 |
Kodak Film | Realistic Portraits |
raw-documentary |
Raw Documentary | Street Photography, Lifestyle |
V. Scene Recognition and Template Selection
After receiving the user's request, first determine which of the following scene categories it belongs to, then apply the corresponding template.
Scene Classification Table
| Scene Category | Keyword Signals | Corresponding Template |
|---|---|---|
| 🎯 Realistic Photography | Portrait, street photography, film, photographic feel, real | → Template A |
| 📦 Product Mockup | Product photo, e-commerce, packaging, advertising image | → Template B |
| 🎬 Cinematic Texture | Film, still, widescreen, atmosphere | → Template C |
| 🧸 Pixar 3D | 3D, cartoon characters, mascots, Pixar | → Template D |
| 📱 App UI | App interface, phone screenshot, UI design | → Template E |
| 📊 Infographic | Infographic, diagram, data visualization, infographic | → Template F |
| 🎨 Logo and Brand System | Logo, mark, brand VI, brand manual, brand touchpoints | → Template G |
| 🏙️ Isometric Miniature | Isometric, 3D miniature, isometric, SaaS hero image | → Template H |
| 🃏 Creative Concept | Trading card, figure packaging, comic strip, 360 panorama, exploded view, ink-wash poster, science poster, brand visual board, action sheet, collectible toy | → Template I |
| 🖼️ Editing Workflow | Modify existing image, style transfer, local adjustments | → Template J |
| 🎭 Concept Illustration/Scene Card | Holiday card, story scene, emotional illustration, concept art | → Template K |
Template effectiveness quick reference (33 sets of blind evaluation hard data)
Core principle: Knowing when not to use a template is more important than the template itself. Check this table before deciding whether to apply a template.
| Scenario | Template effect | Conclusion |
|---|---|---|
| Pixar 3D | 0:3 major loss | ❌ Don't use a template, keep it concise |
| Conceptual illustration | 0:3 major loss | ❌ Don't use a template, maintain narrative quality |
| Realistic photography | 3:3 tie | ⚠️ Template provides no boost, concise mode is enough |
| Creative concept | 2:2 tie | ⚠️ Depends on the situation, don't apply to simple scenes |
| Isometric miniature | 1:1 tie | ⚠️ Template provides no boost |
| App UI | 2:0 major win | ✅ Must use Template E |
| Logo/Brand | 2:0 major win | ✅ Must use Template G |
| Editorial workflow | 3:0 major win | ✅ Must use Template J |
| Cinematic quality | 2:1 win | ✅ Template C recommended |
| Infographic | 1:1 tie | ⚠️ Template F helps but is not required |
| Product Mockup | 1:2 loss | ⚠️ Original prompt may be better, see Template B |
VI. Scenario Template Library
Template A: Realistic Photography (Photorealism)
Core principle: Describe it like a real photo being taken. Use photographic language and demand realistic texture (pores, wrinkles, fabric wear, flaws). Avoid words that imply studio retouching.
⚠️ Key distinction: Realistic photography has two types—scenario portrait(with story, action, environment) andstudio portrait(static, controllable lighting, plain background). Choosing the wrong template will lose narrative quality.
A1: Scenario Portrait / Story Portrait
Applicable: portraits with a sense of story—the subject is doing something, in a real environment, with interactive elements Features: requires narrative, action, environmental details, emotional guidance
Create a photorealistic candid photograph of a [AGE]-year-old
[ETHNICITY] [GENDER] with [HAIR DESCRIPTION] and [DISTINCTIVE FEATURE].
[人物正在做什么动作] while [伴随状态/互动元素].
[LOCATION] — [环境细节1], [环境细节2], [环境细节3].
Shot like a 35mm film photograph, [SHOT TYPE] at [ANGLE],
using a [FOCAL LENGTH] lens.
[LIGHTING: direction, quality, color temperature].
The image should feel [情绪词: honest/unposed/intimate/contemplative],
with real skin texture, worn materials, and everyday detail.
[约束: No glamorization, no heavy retouching, no studio polish.]
Emotion word quick reference：
| Emotion | Applicable scenario |
|---|---|
| honest, unposed | Documentary, life |
| intimate, warm | Intimate, family |
| contemplative, powerful | Contemplative, lonely |
| raw, gritty | Street, underclass |
| dreamy, ethereal | Dreamy, romantic |
A2: Studio Portrait
Applicable: standard portrait photos—controllable lighting, simple background, emphasized facial features
Photorealistic medium close-up portrait of a [AGE]-year-old
[ETHNICITY] [GENDER] with [HAIR DESCRIPTION] and [DISTINCTIVE FEATURE].
Wearing [CLOTHING DESCRIPTION], seated in [LOCATION].
Shot on a 35mm full-frame camera with a 50mm f/1.4 lens,
shallow depth of field, golden hour window light from camera left,
3200K warm color temperature.
Natural skin texture with visible pores, sharp focus on eyes,
slight film grain, no smoothing or beauty filter.
Vertical 4:5 framing.
Cross-image consistency: keep [ETHNICITY] [HAIR] [DISTINCTIVE FEATURE] consistent; embedding persistence will maintain facial consistency.
A3: iPhone street photography / life snapshot
Applicable: iPhone snapshots, life records, social media authenticity Features:simplicity is king, do not over-structure
Amateur iPhone photo of [SCENE + 主体 + 动作].
Shot from [ANGLE], natural [TIME OF DAY] light, no flash.
Environment: [2-3 个真实环境细节].
The image should look like a very authentic life slice.
⚠️ If the user input is already concise and strong enough (e.g. "Amateur iPhone photo at Apple Park, Tim Cook presenting on stage. Shot from the crowd at a distance"),Don't inflate it. Only fill in missing elements.
A4: 35mm film style
35mm film photography, [AESTHETIC ADJECTIVE] aesthetic,
soft ambient [LIGHTING TYPE] lighting mixed with gentle natural window light,
subtle film grain, gentle color shift, high atmosphere editorial style,
intimate medium shot, [SUBJECT DESCRIPTION].
[PHYSICAL DETAILS — 皮肤质感、发型、服装]
[LIGHTING: direction, color temperature, quality]
[POSE & LOCATION]
A5: RAW documentary / candid moment capture
Applicable: subway stations, convenience stores, street—pursuing "unadorned authenticity" Features:Poetic and concise, preserving the original version's visual feel and rhythm
Create a completely RAW quality, unprocessed, unedited image
with full iPhone camera quality.
[SCENE — 用短句描述场景，保留诗意]
[人物 + 动作 — 简洁，不清单化]
[1-2 个关键环境细节，不要堆砌]
⚠️ The good thing about this type of prompt is the "a momentary blur"-style poetry; don't change it to "dynamic motion blur with slight optical distortion".
Negative constraint phrasing(Canghe's experience): In realistic documentary scenes, explicit negative exclusions are more precise than positive constraints. Example:
负面约束：不要插画、动漫、CGI、棚拍光、过度干净、过度构图、假液体、漂浮物、品牌文字、水印、海报设计感。Used together with positive constraints, the effect is best.
Template B: Product Mockup (Product Photography)
Core principle: Describe it like a product ad image that is already on sale. Focus on materials, lighting, and label completeness.
⚠️ Blind evaluation tip: In blind evaluation, Product Mockup lost to concise prompts 1:2. If the user's description is already specific (with product name, material, scene), prioritize using concise mode for direct optimization instead of applying a template.
B1: Standard product studio shot
A close-up product photograph of a [PRODUCT TYPE] standing upright
on a [SURFACE] with a clean [BACKGROUND] backdrop.
The packaging is [MATERIAL] with [TEXTURE], featuring:
- A bold logo "[BRAND]" in [LOGO STYLE]
- A descriptive line "[DESCRIPTION]" below the logo
- A small badge in the upper-right reading "[BADGE TEXT]"
Lighting: large softbox at 45° from camera left,
small fill light from camera right, subtle reflection on the surface.
Shot at f/4, ISO 100, 1/125s, on a 100mm macro lens,
3:4 vertical crop, ultra-sharp focus on the label.
Packaging type quick reference：
| Product | Material | Surface |
|---|---|---|
| Coffee beans | Kraft paper bag + metallic foil seal | Wooden table |
| Skincare products | Frosted glass bottle + embossed cap | Marble |
| Food can | Matte tin can + paper label | Light gray concrete |
| Digital accessories | Premium tactile black box | Dark leather |
B2: Transparent background product extraction
A [PRODUCT] photographed on a transparent/white background.
Clean edges, no color fringing, no light spill.
Studio product photography, soft even lighting,
subtle shadow directly beneath the object.
Sharp focus on the entire product, f/8, 100mm macro lens.
B3: Product application scene
A [PRODUCT] in use in a [CONTEXT/ENVIRONMENT].
[USER/ACTION DESCRIPTION]
Natural [TIME OF DAY] lighting, lifestyle photography style.
Shot at f/2.8, 35mm lens, eye-level angle.
Focus on the product with the environment softly blurred.
Template C: Cinematic
Core principle: Describe it like a frame from a movie screenshot. Specify film stock, color grading, composition, and director reference.
A cinematic still from an imaginary [GENRE] film,
shot on Kodak Vision3 500T 35mm film stock.
The frame shows [SUBJECT + ACTION] in a [LOCATION]
during [TIME OF DAY].
Color palette: teal shadows and orange highlights,
slight halation around bright areas, organic film grain,
anamorphic 2.39:1 widescreen aspect ratio.
Camera: 40mm lens at f/2, slight motion blur on the foreground,
deep focus on the subject's face.
Mood: [MOOD ADJECTIVES], inspired by the visual language of
[DIRECTOR REFERENCE].
Style quick reference：
| Sub-style | Color grading | Lighting | Typical scene |
|---|---|---|---|
| Film Noir | High-contrast black and white | Venetian blind shadows | Detective, suspense |
| Coming-of-Age | Warm tones | Natural light + 16mm grain | Youth, coming of age |
| Cyberpunk | Neon blue/purple | Rainy night wet ground reflections | Sci-fi, cyber |
| Wabi-sabi | Low saturation | Soft window light + 16:9 | Quiet, everyday |
| Denis Villeneuve | Orange-red gradient + silhouette | Strong backlight + reflective floor | Epic, loneliness |
| Wes Anderson | Symmetry + pastel colors | Even flat light | Fun, symmetry |
Template D: Pixar 3D character
Core: Describe "who the character is and what they are doing", plus key visual anchors (expression, material), and do not list rendering parameters. v1.2 was over-simplified and lost to the opponent in PK—need to find a balance between "too detailed" and "too minimal".
D1: Standard mode (recommended)
A 3D Pixar-style [SUBJECT], [KEY ACTION/POSE].
[表情+体态描述, 如 large expressive eyes, rosy cheeks, gentle smile].
[1-2 KEY VISUAL DETAILS, e.g. outfit texture, prop, fur/hair detail].
Clean [BACKGROUND TYPE], soft warm lighting, shallow depth of field.
Principle: 4 lines. Line 1 character + action, line 2 expression anchor, line 3 visual details, line 4 background + lighting. Why it's better than minimalist: The PK competition proved that adding expression anchors and visual details can improve the score by 0.5-1.0 points.
D2: Minimalist Mode
A cute 3D animated [SUBJECT] doing [ACTION].
[Pixar/Disney/Illumination] style, [warm/bright/soft] lighting.
When to use minimalist: When the user only says "draw a cute XX" or "Pixar-style XX," with no complex requirements.
Counterexample (don't write like this)：
❌ A 3D Pixar-style character of a cat, 3/4 front view, soft cinematic key light from above,
❌ warm rim light from behind. Smooth subsurface scattering on skin,
❌ fluffy hair with stray strands, subtle fabric folds on clothing.
❌ Background: clean pastel gradient, shallow depth of field with creamy bokeh.
❌ Render quality: feature-film polish, soft global illumination, no harsh shadows.
→ v1.1 blind evaluation: 1.12 points worse than the concise prompt. The model was over-constrained and lost its creativity.
Template E: App UI Mockup
A high-fidelity mobile app screenshot, iPhone 15 Pro frame,
vertical 9:19.5 aspect ratio.
The screen shows a [APP CATEGORY] app with the following layout:
- Top: status bar (9:41, 100% battery, full signal)
- Header: app name "[APP NAME]" in bold, profile icon on the right
- Main: a [HERO COMPONENT] taking 60% of the screen
- Below: [LIST/CARDS/CONTENT BLOCKS]
- Bottom: tab bar with 4 icons (home / explore / notifications / profile)
Design language: [COLOR PALETTE], rounded corners (16px),
subtle drop shadows, system font (SF Pro), [light/dark] mode.
Render the screen pixel-perfect, all text fully legible.
Template F: Infographic
Create a detailed infographic about [TOPIC].
Target audience: [AUDIENCE].
Layout: [vertical/vertical scroll/single panel].
Sections (top to bottom):
1. Title: "[TITLE]" — large bold sans-serif, [COLOR]
2. Section 1: [TOPIC] — [ICON] + [STAT/DATA POINT]
3. Section 2: [TOPIC] — [VISUAL ELEMENT] + brief explanation
4. Section 3: [TOPIC] — [CHART/DIAGRAM TYPE]
5. Footer: "[SOURCE/CALL TO ACTION]"
Color palette: [PRIMARY], [SECONDARY], [ACCENT].
Typography: clean sans-serif for body, bold geometric for headers.
Visual style: flat vector illustrations with subtle shadows,
no 3D effects, generous whitespace between sections.
All text must be legible at 100% zoom.
quality="high" recommended for dense text layouts.
Template G: Logo Design and Brand System
Core: Logo generation requireshard constraints(color values, symmetry, no gradients) + brand personality description. PK competition lesson: lacking specific color values and technical constraints will lose to an opponent's precise description. v1.6 upgrade: expanded from a single Logo to a four-in-one system of Logo + brand identity package + brand touchpoint board + brand-enveloped advertising.
G1: Single Logo Design
Design a [STYLE: minimalist / geometric / organic / vintage] logo for "[BRAND NAME]",
a [BUSINESS TYPE / INDUSTRY].
Concept: [CORE VISUAL IDEA, e.g. "a leaf shape integrated with the letter E"].
Brand personality: [ADJECTIVE 1], [ADJECTIVE 2], [ADJECTIVE 3].
Visual specs:
- Color: [PRIMARY COLOR] ([HEX CODE]) on pure white background
- Shape: [SYMMETRICAL / ASYMMETRICAL], [GEOMETRIC / ORGANIC] lines
- Weight: [SINGLE STROKE / VARIED THICKNESS], [THIN / MEDIUM / BOLD]
- Layout: centered logo mark with brand name below in [SERIF / SANS-SERIF / CUSTOM] font
Constraints: no gradients, no shadows, no 3D effects, no watermarks,
scalable from favicon to billboard, original and non-infringing.
Logo multi-touchpoint validation(recommended to add to the prompt):
Also show the logo applied across touchpoints:
- Business card (90×55mm)
- App icon (rounded square)
- Website header
- Signage / billboard
This helps verify readability at different sizes.
Source: Canghe awesome-gpt-image-2. A Logo can't be judged from a single image alone—whether it's still readable when scaled down to an app icon, and whether its horizontal and vertical proportions fit, only a multi-touchpoint display can expose the problems.
Why Constraints are needed: Both the official OpenAI Cookbook and the fal.ai guide emphasize that the key to Logo generation is using constraints to prevent the model from "freely improvising in directions you don't want." Color values, symmetry, and no gradients are not over-structuring—they are essential hard parameters for a Logo.
PK competition data: The v1.2 Logo template, lacking color values and technical constraints, lost to an opponent that specified #2D5016 + no gradients.
G2: Complete Brand Identity Package
Source: Canghe awesome-gpt-image-2. Suitable for complete VI delivery for a brand from 0 to 1.
为[业务名]交付一套完整品牌身份系统。
输入信息：
业务描述：[一句话说明]
行业：[行业]，目标受众：[描述]
品牌个性：[5个关键词]
希望触发的感受：[信任/兴奋/奢华/亲近/力量]
请输出：
1. Logo概念：3-5个完全不同的方向，每个说明视觉理念、形状语言、象征意义、字体方向
2. 配色系统：主色+辅助色+强调色+中性色，附HEX代码和使用规则
3. 字体系统：标题+正文+强调字体，字号层级和免费替代方案
4. 应用触点：名片、App图标、网站首页、社媒模板、广告牌上的应用效果
5. 品牌规则：3条永远不要打破的核心规则
6. 禁用规则："不要怎么用"的明确清单
设计语言：[现代极简/科技品牌/奢华编辑]，主色[颜色+HEX]，大量留白。
输出：结构化品牌手册，任何设计师或AI工具都能在10分钟内理解并复用。
G3: Brand-Enveloped Product Advertising
Source: Canghe awesome-gpt-image-2. When swapping in different products under the same brand, the visual world stays consistent.
PHASE 1 / ANCHOR：用 2 行描述[品牌身份]，包括调色板、材质、光影和情绪。
PHASE 2 / INJECT：把[产品]放入这个品牌世界中，产品要服从品牌气质和环境语言。
PHASE 3 / FORMAT：指定[输出格式]，例如 hero 图、方形广告、竖版 story 或电商头图。
PHASE 4 / SIGNATURE：加入[品牌元素]，例如颗粒、阴影、叠加纹理、包装符号或图形边框。
目标：同一品牌下替换不同产品时，视觉世界保持一致，广告图仍有明确主角和商业质感。
Template H: Isometric Miniature
A 45° top-down isometric miniature 3D scene of a [SCENE THEME]
diorama on a wooden display base.
Soft refined PBR textures, realistic materials,
clean unified composition, minimalistic aesthetics.
Tiny props integrated into the architecture: [3 SPECIFIC ELEMENTS].
Studio softbox lighting, subtle ambient occlusion,
pastel color palette dominated by [COLOR1] and [COLOR2].
Square 1:1 frame, centered subject, plenty of negative space.
Template I: Creative Concepts
I1: Retro Trading Card
A premium holographic trading card, vertical 3:4 layout.
Center: a [SUBJECT] in dynamic pose, vibrant cinematic lighting.
Border: ornate gold filigree with rune-like icons in four corners.
Top banner reads "[RARITY]" in bold serif caps.
Bottom panel: name plate "[CHARACTER NAME]", three small stat icons
(power / speed / magic) with numeric values.
Holographic foil effect, slight grain, studio backdrop.
I2: Action Figure Blister Pack
A stylized action figure of [SUBJECT] sealed inside a premium
plastic blister pack, photographed straight-on.
The cardboard backing is glossy with a bold header reading
"[BRAND / NAME]" in oversized sans-serif caps and a smaller
tagline "[TAGLINE]".
The figure is posed upright with [ACCESSORY 1] and [ACCESSORY 2]
slotted into molded compartments next to it.
Studio product photography, soft top lighting,
clean off-white background, subtle reflection on the floor.
I3: 360° Panorama
A 360° equirectangular panoramic photograph of [LOCATION],
aspect ratio 2:1.
The horizon is perfectly level across the middle of the frame.
Foreground (bottom 1/3): [具体前景元素].
Mid-ground (middle 1/3): [建筑/人物/场景].
Background (top 1/3): [天空/远景].
Lighting: natural [TIME OF DAY] sun, soft atmospheric haze.
No fish-eye distortion at the poles, ready for VR projection.
I4: Comic Strip
Create a short vertical comic-style reel with [N] equal-sized panels.
Panel 1: [场景 + 动作]
Panel 2: [场景 + 动作]
Panel 3: [场景 + 动作]
Panel 4: [场景 + 动作]
Style: clean line art, flat colors, consistent character design
across all panels. Each panel should read as a clear narrative beat.
I5: Concept Product R&D Breakdown Board
Source: Canghe awesome-gpt-image-2. This is a kind ofmulti-stage narrative composition—not a single finished render, but a complete R&D story from inspiration → analysis → iteration → validation → specifications.
为[产品/家具/装置]生成一张完整的概念产品研发拆解板，而不是单张成品渲染图。
核心概念：
把[灵感来源，如揉皱纸团/贝壳/折纸/机械结构]转译成[产品类型]。
设计哲学：[一句话说明功能与情绪，例如"把受控混乱转化为高舒适度座椅"]。
画面结构：
中心：高质量 hero render，展示最终产品的主要形态、材质和比例。
左侧：观察与形态分析，包含灵感图、轮廓提取、结构线、折痕/纹理/受力方向标注。
中部：形态迭代过程，展示从原始形态到产品外壳的 3-5 个演化步骤。
下方：人体工学或使用场景验证，包含尺寸、角度、使用姿态和关键功能说明。
右侧：结构集成与材料方案，展示内部骨架、外壳、软垫/面料/连接件等分层拆解。
底部：最终材质、表面纹理、颜色方案和关键规格表。
视觉风格：
工业设计提案板，干净白底或浅灰背景，技术图纸 + 产品摄影混合风格，
细线标注，清晰标题，真实阴影，材质细节可见。
约束：
不要只画一个漂亮产品；必须展示分析、迭代、人体工学、结构、材料和规格。
不要让文字挤满画面；每个阶段只保留短标题和关键标签。
产品外形应保留[灵感来源]的识别特征，但必须看起来可制造、可使用。
I6: Ink-Wash Double Exposure Character Poster
Source: Canghe awesome-gpt-image-2. Suitable for high-end character posters for personal IPs, brand founders, and athletes.
生成一张[人物/角色/品牌主理人/运动员]的水墨双重曝光人物海报。
画幅：9:16 竖版，高级电影海报构图。
主体结构：
- 上半区：放大的人物头部、面部轮廓或半身剪影，形成最强识别锚点。
- 中下区：同一人物的全身或半身主体，姿态为[站姿/动作姿态/凝视镜头]。
- 剪影内部：融合[关键场景]、[象征物]、[叙事片段]、[环境纹理]，形成双重曝光叙事。
视觉连接：用云雾、水墨扩散、飞白边缘、负空间和柔和明暗过渡，把上方剪影、内部拼贴和下方主体连成一条从上到下的视觉动线。
风格：东方水墨美学 + 写实电影感，克制、高级、留白充足，层次丰富但不杂乱。
文字：可加入[标题/姓名/短句]，必须少量、可读、像海报题签而不是信息图说明。
约束：不要硬拼贴，不要把背景塞满，不要廉价武侠特效，不要复制真实海报版式，不要让剪影和主体互相抢焦点。
输出：海报级完成图，主体清晰，水墨边缘自然，叙事元素与人物身份强相关。
I7: Nature Popular Science Poster (Apple Style)
Source: Canghe awesome-gpt-image-2. Suitable for popular science content, natural history museums, and educational visuals.
生成一张 9:16 竖版高级科普海报，Apple Keynote 产品发布风格。
画面结构：
顶部左侧：中文大标题 [{物种名}]，副标题 [{一句定位}]，细短横线，英文名 [{英文物种名}]，分布信息。
中部：超高清、真实、强烈立体感的 {物种名}，占画面 50%-70%。白色背景，可保留少量承托物（树枝/岩石/雪地），真实阴影。
底部：四个极简信息栏目，细线 icon + 彩色小标题 + 1-3行短说明，栏目间极细浅灰竖线分隔。
最底部：一句灰色小字总结。
设计原则：
- 主体极度放大，最强视觉中心
- 纯白/极浅灰白渐变背景，大量留白
- 不使用圆角卡片、厚边框、大面积装饰图形
- 文字像高端发布会视觉，标题巨大，副标题克制
禁止项：不要淡黄色旧纸背景、复杂信息图网格、圆角卡片、儿童科普风、卡通风。
I8: Brand Touchpoint System Visual Board
Source: Canghe awesome-gpt-image-2. Not a single poster, but a complete set of brand application displays—suitable for brand proposals and VI reports.
为[品牌名]生成一张高端品牌触点系统视觉板。
品牌定位：[行业/生活方式/产品品类]
核心气质：[关键词1]、[关键词2]、[关键词3]
主视觉场景：[核心产品/服务/体验]，放在[材质表面/空间场景]中。
触点系统必须包含：
- 主产品 hero shot
- 包装盒 / 手提袋 / 杯子 / 标签 / 贴纸 / 封签等品牌物料
- 菜单卡 / 价目表 / 小型排版样张
- 生活方式场景或用户使用片段
- 配色、字体、图形语言在不同触点上的统一应用
设计语言：[现代极简/日式留白/奢华编辑/科技品牌]，主色[颜色]，辅助色[颜色]，大量留白，细腻材质，真实阴影。
构图：像顶级设计机构提案页，主视觉最突出，辅助物料层级清楚，整体有品牌系统感。
约束：不要只生成一个 logo；不要把所有物料挤成杂乱拼贴；不要使用随机乱码文字；不要让各触点风格割裂。
I9: Character Action Breakdown Reference Sheet
Source: Canghe awesome-gpt-image-2. Suitable for animation/dance/game action reference, 4×4 grid structured action sequence.
生成一张[角色/人物]动作分解参考表。
风格：[黑白线稿/3D 灰阶/漫画分镜/教学图]，背景纯净。
版式：4×4 网格，共 16 个等尺寸面板，细线分隔，每格左上角编号 1-16。
角色一致性：所有面板使用同一角色，保持脸型、服装、比例和发型一致。
每格结构：
- 顶部：动作标题
- 中央：完整身体动作姿态
- 底部：3-4 行动作说明
- 叠加：方向箭头、旋转箭头或运动轨迹线
动作序列：[从基础站姿到结束动作的完整步骤]
约束：不要复杂背景，不要新增角色，不要彩色干扰，不要改变角色身份。
输出：清晰可读、可用于动画/舞蹈/游戏动作参考的角色动作表。
I10: Reference image to 3D collectible toy
Source: Canghe awesome-gpt-image-2. Suitable for converting portrait photos into high-end trendy toy/collectible figures.
将输入照片转换为高端 3D 收藏玩具形象。
身份保持：保留原始人物/角色的脸部身份、主要发型、表情气质和服装识别点。
造型比例：大头设计，五官轻微夸张，身体比例玩具化，但整体仍保持高级设计感。
材质：哑光 vinyl / resin / collectible figure finish，皮肤和服饰材质要有细节。
灯光与背景：柔和棚拍光，干净背景，[黑色/白色/品牌色]，主体居中，轮廓清晰。
质感：超清锐度，真实材质反射，premium designer toy aesthetic。
约束：不要改变身份，不要廉价塑料感，不要多角色，不要复杂背景，不要文字水印。
输出：一张完整的高端收藏玩具渲染图。
Template J: Edit Workflow (Edit)
Core Principle: Clearly divide into Change (what changes) and Preserve (what stays the same). Repeat the invariant list with every edit to prevent drift.
J1: Style Transfer
Use the same style from the input image and generate [NEW SUBJECT].
Preserve: palette, texture, brushwork, film grain, composition framing.
Change: only the subject/content.
J2: Local Modification
Change: replace [SPECIFIC ELEMENT] with [NEW ELEMENT].
Preserve: keep the subject's face, pose, clothing, and lighting
exactly the same as the input.
J3: Lighting/Weather Transformation
Change: transform the lighting to [NEW LIGHTING CONDITION].
Preserve: identity, geometry, camera angle, object positions,
all physical elements unchanged.
J4: Multi-Image Compositing
Place the [ELEMENT] from the second image into the setting of image 1,
[POSITION DESCRIPTION].
Use the same style of lighting, composition and background.
Do not change anything else in image 1.
J5: Virtual Try-On
Edit the image to dress the person using the provided clothing images.
Do not change face, facial features, skin tone, body shape, pose,
or identity in any way.
Preserve exact likeness, expression, hairstyle, and proportions.
Replace only the clothing, fitting the garments naturally to
the existing pose and body geometry with realistic fabric behavior.
Match lighting, shadows, and color temperature to the original photo
so the outfit integrates photorealistically.
Do not change the background, camera angle, framing, or image quality.
J6: Interior Design Replacement
In this room photo, replace ONLY [SPECIFIC ELEMENT] with [NEW ELEMENT].
Preserve camera angle, room lighting, floor shadows, and surrounding objects.
Keep all other aspects of the image unchanged.
Photorealistic contact shadows and [MATERIAL] texture.
J7: Product Extraction (Transparent Background)
Extract the [PRODUCT] from this image and place it on a clean,
transparent background.
Preserve the exact shape, color, texture, and details of the product.
Clean edges, no color fringing, no background remnants.
Studio-quality lighting with subtle shadow beneath.
J8: Sketch → Render
Turn this drawing into a photorealistic image.
Preserve the exact layout, proportions, and perspective.
Choose realistic materials and lighting consistent with the sketch intent.
Do not add new elements or text.
J9: Image Translation (Localization)
Translate the text in this image to [TARGET LANGUAGE].
Do not change any other aspect of the image.
Preserve typography style, placement, spacing, and hierarchy.
Translate verbatim and accurately, with no extra words.
J10: Clothing/Product Replacement
Edit the image to replace the [ORIGINAL PRODUCT] with [NEW PRODUCT].
Do not change the person's face, facial features, skin tone,
body shape, pose, or identity.
Preserve exact likeness, expression, hairstyle, and proportions.
Fit the new product naturally to the existing pose with realistic behavior.
Match lighting, shadows, and color temperature to the original.
Template K: Concept Illustration / Scene Card
⚠️ Important: Concept illustration scenesThe core is creative storytelling, not piling up parameters. Blind evaluation results: Overly structured illustration prompts score 0.6-1.0 points worse than concise originals. Maintain visual feel and narrative quality; do not list composition/color/texture parameters.
K1: Scene Card (with text)
[插画类型] illustration: [核心场景描述，1-2句话].
[Mood/氛围，一个短句].
Style: [视觉风格关键词，2-3个].
Include ONLY this text: "[要出现的文字]"
Principle: Let the model decide composition, color scheme, and lighting on its own. You only describe "what to draw" and "what feeling".
K2: Story Scene (no text)
A [风格] illustration of [SCENE].
[主体 + 动作].
Mood: [情绪].
Extreme simplicity. 2-3 sentences are enough. The model will automatically fill in all visual details.
Negative Example (do not write like this)：
❌ Create a holiday card illustration.
❌ Scene: [详细场景描述，包括主体、动作、环境、细节].
❌ The scene suggests [情感/叙事暗示].
❌ Mood: [情绪形容词1], [情绪形容词2], [情绪形容词3].
❌ Style: [视觉风格描述], [光线], [质感], [构图], [品质].
❌ Constraints: Original artwork only. No trademarks. No watermarks. No logos.
→ Blind evaluation results: This writing style scores 0.6-1.0 points worse than concise prompts. Templated descriptions kill the liveliness that illustrations need.
Seven, 4 Major Advanced Techniques
Technique 1: Use Constraints to Control the Number of Elements
Multi-element scenes may be over-rendered; add at the end:
Constraints: exactly [N] elements, no extra props,
no additional text beyond what's specified above.
→ Positive constraints are much more stable than negative prompts on gpt-image-2.
Technique 2: Use Seed to Reproduce Composition
img = client.images.generate(
model="gpt-image-2",
prompt=PROMPT,
size="1024x1536",
quality="high",
extra_body={"seed": 20260421},
)
→ Fixed seed + fixed prompt = maintain consistent composition, lighting, and character features across generations.
Technique 3: Production-Grade Resolution Workflow
| Stage | Goal | Resolution | n | Budget |
|---|---|---|---|---|
| Concept Exploration | Find direction | 1024×1024 | 4 | 10% |
| Composition Iteration | Lock subject layout | 1024×1536 | 2 | 25% |
| Style Convergence | Determine lighting and color | 1024×1536 | 1 | 20% |
| Text Refinement | edit fine-tuning | 1024×1536 | 1 | 15% |
| Final Output | 4K upscale | 2048×3072 | 1 | 30% |
→ Cost reduced by 60%, quality pass rate from 40% → 85%+.
Technique 4: Iterate Rather Than Overload
Start with a clean base prompt, then follow up with small single changes:
"make lighting warmer""remove the extra tree""shift the subject to the left third""same style as before, but change the background to..."
Eight, 6 Common Pitfalls (scan before every generation)
| Pitfall | Wrong Approach | Correct Approach |
|---|---|---|
| 1. Cramming into one long sentence | A paragraph turns into a blur | Segment: scene→subject→details→lighting→constraints |
| 2. Conflicting styles | "photorealistic" + "cartoon" | Choose only one main style |
| 3. Negative prompts | "no watermark, no text" | Switch to positive constraints |
| 4. Ignore camera parameters | Don't write focal length and aperture | Explicitly specify, has physical effects |
| 5. Text without quotation marks | Write the text directly | "..." wrap |
| 6. Edits don't clarify what changes and what stays the same | Only say "change the background" | Change: ... / Preserve: ... |
IX. Size and Aspect Ratio Quick Reference
| Ratio | Pixels | Applicable scenarios |
|---|---|---|
| 1:1 | 1024×1024 | Logo, avatar, social media |
| 4:5 | 1024×1280 | Instagram vertical image, portrait |
| 3:4 | 1024×1365 | Trading cards, vertical posters |
| 16:9 | 1536×864 | WeChat Official Account cover, horizontal main image |
| 9:16 | 864×1536 | Mobile full screen, Stories |
| 2:3 | 1024×1536 | Magazine pages, vertical posters |
| 2:1 | 1536×768 | 360° panorama, Banner |
| 2.39:1 | 1536×642 | Cinematic widescreen |
X. Execution Process and Guidance (Phase 0-5)
Phase 0: Complexity assessment (critical!)
Before applying a template, first assess the complexity of the user input. Also refer to §V Template Effectiveness Quick Reference to confirm whether this scenario needs a template.
| User input characteristics | Mode | Handling method |
|---|---|---|
| Already a structured prompt / only needs minor adjustments | 🟢 Concise mode | Only optimize and complete, don't force a template |
| One-sentence requirement, no details | 🔵 Expansion mode | Apply template, fill placeholders |
| Multiple scene elements + specific requirements | 🟡 Custom mode | Reference template, mix and assemble |
| Edit an existing image | 🟠 Edit mode | Use template J |
| Creativity-driven requirement(3D cartoon/illustration/art) | 🔴 Skip template | Directly optimize the user description, don't apply any template |
Judgment criteria：
- If the user input already contains ≥3 of the following elements (scene, subject, action, camera, lighting, style) → use concise mode
- If the scene belongs to Pixar 3D / concept illustration / artistic creation → skip the template, only do minimal optimization (add quality/size)
- The blind evaluation table in §V will tell you whether the scenario template is effective → check first, then apply
Concise mode principles(when user input is already very good):
- Don't inflate 2 sentences into 5
- Don't force structure onto an already concise and powerful prompt
- Only complete when key elements are missing (missing camera parameters? missing lighting?)
- Preserve the original's poetic quality, rhythm, and visual feel
- Don't change "a momentary blur" to "dynamic motion blur"
Phase 1: Scenario + theme confirmation (ask all at once, 5-6 questions)
After receiving the user's requirement,at oncethrow out the following questions (skip those the user has already made clear):
帮你做一张好图！几个问题确认方向：
1️⃣ 场景类型：
🎯写实照片 | 📦产品Mockup | 🧸3D卡通 | 🎬电影质感 | 📱App UI
🎨Logo/品牌系统 | 📊信息图 | 🎃创意概念 | 🖼️编辑图片 | 🎭插画 | 🏙️等轴微缩
💡 创意概念包含：交易卡/人偶/漫画条/拆解板/水墨海报/科普海报/品牌视觉板/动作表/收藏玩具
💡 Logo/品牌系统包含：单Logo / 完整品牌身份包 / 品牌包络广告
2️⃣ 核心主体：画面的主角是谁/什么？
3️⃣ 用途：公众号封面 / 产品展示 / 品牌VI / 社交媒体 / PPT / 其他
4️⃣ 情绪氛围：温暖治愈 / 科技未来 / 奢华高端 / 可爱萌 / 神秘幽暗 / 清新自然
5️⃣ 有参考图或参考风格吗？可以发给我
不确定的跳过就行，我来推荐 👍
Smart skip rules：
- Parameters the user has already made clear → don't ask again
- User provides a reference image → skip style/color tone questions, extract from the image
- User says "you decide" → use the best default values for that scenario
Phase 2: Scenario deep dive guidance (ask 6-8 parameters per scenario)
Based on the scenario determined in Phase 1,in one goask for the core parameters of that scenario:
🧸 Pixar 3D
⚠️ Blind evaluation conclusion: Template 0:3 major loss. Do not use Template D; only collect key information to build a concise description.
3个核心参数决定画面质量：
- 角色是谁？在做什么？（一只小狐狸抱着猫咪 / 机器人举着气球）
- 表情和性格？（大眼睛好奇歪头 / 开心大笑 / 害羞低头）⚡ 最影响质量
- 参考哪个动画风格？（默认Pixar | Disney/Illumination/DreamWorks/吉卜力）
📱 App UI
✅ Blind evaluation conclusion: Template 2:0 major win. Template E is recommended (see §6).
- App类型？（冥想/外卖/健身/金融/社交/音乐）
- App名称？（出现在Header中）
- 深色还是浅色模式？
- 主色调？（蓝色系/绿色系/紫色系；可给HEX色值）
- 首页主要展示什么？（呼吸圆圈+课程卡片 / 信息流 / 地图 / 仪表盘）
- 底部Tab几个？分别是什么？（默认4个: home/explore/stats/profile）
- iOS还是Android？（默认iPhone 15 Pro | Android Material）
- 想参考哪个App的风格？
🎨 Logo and Brand System
✅ Blind evaluation conclusion: Template 2:0 major win. Template G is recommended (see §6).
First determine which tier you need：
- Just one Logo → G1 Single Logo Design
- Complete VI for a brand from 0 to 1 → G2 Complete Brand Identity Package
- Consistency across multiple product ads under the same brand → G3 Brand-Enveloped Product Ads
G1 Single Logo — Guiding Questions：
- 品牌名？
- 行业/领域？
- 品牌个性？选3个形容词（如clean, trustworthy, earth-conscious）
- 核心视觉概念？（如叶子+E字母融合 / 山峰轮廓 / 无限符号）
- Logo风格？极简几何 / 有机圆润 / 复古经典 / 手绘质感 / 字母标
- 主色调？（给HEX色值最好，如 #2D5016）
- 需要带文字吗？（纯图标 / 图标+品牌名 / 只要字体设计）
- 对称性？（默认对称 | 不对称）
G2 Complete Brand Identity Package — Guiding Questions：
- 业务名和一句话描述？
- 行业和目标受众？
- 品牌个性？5个关键词
- 希望触发的感受？（信任/兴奋/奢华/亲近/力量）
- 喜欢的视觉身份参考？3个
- 讨厌的视觉身份参考？3个
- 设计语言偏好？（现代极简/科技品牌/奢华编辑）
G3 Brand-Enveloped Product Ads — Guiding Questions：
- 品牌身份描述？（调色板、材质、光影、情绪）
- 要放进去的产品？
- 输出格式？（hero图/方形广告/竖版story/电商头图）
- 品牌签名元素？（颗粒/阴影/叠加纹理/包装符号/图形边框）
🎯 Realistic Photography
⚠️ Blind evaluation conclusion: Template 3:3 tie. Templates provide no boost; concise mode is sufficient. The following parameters are used to fill in missing elements.
- 主体是谁？（年龄/性别/种族/发型/穿着/特征）
- 在做什么？（坐在咖啡馆看书 / 走东京街头 / 抱着孩子微笑）
- 场景环境？（咖啡馆 / 地铁站 / 海边日落 / 雨中街道）
- 整体氛围？温暖治愈 / 冷酷都市 / 胶片复古 / 奢华时尚 / 纪实新闻
- 镜头感？（默认35mm candid | 50mm人像 / 85mm浅景深 / iPhone业余感）
- 光线？（默认natural | golden hour / 霓虹灯 / 影棚 / 逆光剪影）
- 拍摄角度？（默认eye-level | 微俯 / 仰拍 / 俯拍）
- 构图？（默认居中 | 三分法 / 对称 / 引导线 / 留白）
🎬 Cinematic Quality
✅ Blind evaluation conclusion: Template 2:1 win. Template C is recommended (see §6).
- 场景描述？（雨夜城市天台 / 废弃工厂 / 太空舱）
- 画面主体？（穿风衣的背影 / 情侣拥吻 / 复古跑车）
- 电影风格？Noir / Villeneuve科幻 / Wes Anderson对称 / Warm Romance / Thriller
- 色调？（默认teal&orange | 低饱和 / 黑白 / 暖琥珀）
- 光线类型？（侧光rim light / 顶光god ray / 霓虹背光 / 烛光）
- 宽高比？（默认2.39:1电影宽银幕 | 16:9 / 4:3 / 1:1）
- 画面中需要文字吗？
📦 Product Mockup
⚠️ Blind evaluation conclusion: Template 1:2 loss. Templates provide no boost; concise mode is preferred. Only refer to Template B when the user's requirements are very vague.
- 产品是什么？（咖啡豆/护肤品/食品罐/数码配件/其他）
- 拍摄场景？（纯色棚拍 / 透明背景提取 / 生活场景应用）
- 核心卖点？材质/颜色/标签上的文字
📊 Infographic
⚠️ Blind evaluation conclusion: Template 1:1 tie. Templates are helpful but not mandatory.
- 主题？
- 面向谁？
- 要展示哪些数据/信息点？（列3-5个）
- 布局？（默认竖版滚动 | 单版面）
- 视觉风格？（默认flat vector | 手绘 / 3D / 等轴）
- 配色方案？
- 图表/可视化类型？（柱状图 / 饼图 / 流程图 / 时间线）
🖼️ Editing Workflow
✅ Blind evaluation conclusion: Template 3:0 major win. Template J must be used (see §6).
J1-J10 Sub-template Selection Decision Tree：
用户想做什么？
├── 换风格/把A变成B风格 → J1 风格迁移
├── 只改局部元素 → J2 局部修改
├── 换光线/天气/时段 → J3 光线变换
├── 两张图合成 → J4 多图合成
├── 换衣服/试穿 → J5 虚拟试穿
├── 换房间/家具/装饰 → J6 室内设计替换
├── 抠图/透明背景 → J7 产品提取
├── 草图变渲染 → J8 草图→渲染
├── 换文字语言 → J9 图片翻译
├── 换产品/服装 → J10 服装/产品替换
└── 不确定 → 先问用户“你想改哪方面？”
- 编辑类型？风格迁移 / 局部修改 / 添加元素 / 删除元素
- 变成什么风格/改成什么？（发参考图或描述）
- 必须保留什么？（人脸 / 构图 / 色调 / 文字）
- 变化程度？（微调 / 适中 / 彻底改变）
- 发原图给我看看
🎭 Concept Illustration
⚠️ Blind evaluation conclusion: Template 0:3 major loss. Do not use Template K; only collect the visual feel in the user's mind.
3个问题抓住灵魂：
- 核心场景？描述你脑海中的画面
- 情绪/氛围？
- 视觉风格？（水彩 / 油画 / flat design / 手绘线稿 / 绘本风）
🏙️ Isometric Miniature
⚠️ Blind evaluation conclusion: Template 1:1 tie. Templates provide no boost.
- 微缩场景主题？（云计算数据中心 / 东京路口 / 咖啡馆 / SaaS界面）
- 想包含的小道具/细节？（列3-5个）
- 主色调？
- 风格？（默认写实PBR | low-poly / 像素 / 萌系）
- 基座样式？（默认木质展台 | 白色展台 / 无基座漂浮）
- 有参考图吗？
🎃 Creative Concept
⚠️ Blind evaluation conclusion: Template 2:2 tie. Use Template I depending on the situation.
First determine the sub-type (now choose one of 10, using the decision tree)：
你要做什么类型的创意图？
├── 复古交易卡 / 收藏卡 → I1
├── 动作人偶包装 → I2
├── 360°全景 → I3
├── 漫画条/四格 → I4
├── 产品研发拆解板（工业设计） → I5
├── 水墨双重曝光人物海报 → I6
├── 自然科普海报（Apple风） → I7
├── 品牌触点系统视觉板 → I8
├── 角色动作分解参考表 → I9
├── 照片转3D收藏玩具 → I10
└── 都不是 → 描述你想做什么，我来匹配
After selecting the sub-type, collect the key parameters according to the placeholders of the corresponding template:
- I1-I4: Refer to the placeholders in the templates in §6 (subject/brand name/action/scenario)
- I5: Inspiration source + product type + design philosophy
- I6: Character identity + core scenario + symbolic object + pose
- I7: Species name + positioning statement + four popular science sections
- I8: Brand positioning + core temperament + touchpoint list
- I9: Character description + complete action sequence
- I10: Send reference images + confirm identity-preservation elements
Phase 3: Output + Technical Settings
最后确认几个技术参数（都有默认值，可以直接跳过）：
📐 尺寸/比例？
1:1正方形(默认) | 16:9横版 | 9:16竖版 | 4:5 Instagram | 2.39:1电影
⚙️ 质量？
medium快速预览(省50%成本) | high正式输出(默认)
🔢 生成几张？
默认1张 | 可选2-4张对比
🔄 A/B双方案？（可选）
在prompt中加“请输出一版主方案+一版备选方案”，一次生成两个方向直接挑选。
适合不确定具体风格时快速探索。
Phase 4: Output prompt + user confirmation
This step is critical — first generate the prompt for the user to see; do not call the API directly.
根据你的需求，这是为你定制的prompt：
---
[完整prompt文本]
---
参数：size=1024x1024, quality=high, model=gpt-image-2
你想怎么处理？
1. ✅ 看着不错，帮我直接调用生成吧
2. ✏️ 我想调整一下（告诉我改哪里）
3. 🔄 换个方向重新来
User chooses 1 → help him call the API to generate the image User chooses 2 → adjust the prompt and display it again User chooses 3 → return to Phase 1
Phase 5: Post-generation Feedback Loop
生成完毕！[展示图片]
满意吗？
- 😍 完美！
- 🔄 微调一下（告诉我改什么，我用edit模式调整）
- 🔁 重新生成（换个随机种子）
- 📐 换个尺寸/比例再出一张
Quality check (internal self-check before displaying the prompt in Phase 4)
- Is the subject at the beginning of the prompt?
- Is the structure clearly segmented?
- Is text wrapped in quotation marks?
- Are the camera parameters clear?
- Is the lighting described specifically?
- No piling on "8K ultra detailed"?
- No conflicting styles?
- Are constraints phrased positively? (For realistic documentary scenes, a negative constraint list can be added to assist)
- Is it not over-engineering a simple requirement?
- If it is a brand-type requirement, was the correct template tier chosen? (G1 single Logo / G2 identity package / G3 envelope ad)
- If it is a creative concept, was the correct sub-template chosen through the decision tree?
Default value strategy
Every parameter has a safe default value—even if the user skips all of them, a qualified image can still be generated:
- Tone → warm tone
- Lighting → natural soft light
- Background → clean gradient / neutral
- Composition → centered with padding
- Quality → high
- Size → 1:1
The default values come from the highest-scoring parameter combinations in 33 blind evaluation groups + 6 PK matches.
Eleven, mixed scenario handling
When the user's requirement is unclear or spans multiple scenarios:
- Prioritize identifying the main scenario: realistic vs design vs creative
- Main template + sub-elements: for example, "realistic portrait + cinematic texture" → use template A as the main body and incorporate template C's color grading and film description
- Do not force multiple templates: choose one as primary and take only key elements from the other
- If completely uncertain: use the general structure formula + the closest Style tag
- Simplicity first: if the user input is already good, do not inflate it just to fit a template
Twelve, boundaries with image2-typography-poster
| Judgment condition | Which skill to use |
|---|---|
| The core of the image is the typography design of "a certain character/word" | → image2-typography-poster |
| The core of the image is a scene, person, product, or concept | → yinyo-image2-prompt |
| Official account cover + text is the protagonist | → image2-typography-poster |
| Official account cover + scene/concept is the protagonist | → yinyo-image2-prompt |
| Uncertain | → First ask the user, "Is the core of this image text or imagery?" |
Reference sources
- OpenAI official Cookbook — GPT Image Generation Models Prompting Guide
- apiyi.com — GPT-Image-2 prompt collection: 10 most popular templates (April 2026)
- GitHub ZeroLu/awesome-gpt-image — community-curated prompt collection
- gptimageai.org — GPT-Image 1.5 Prompt Guide
- PixVerse — GPT Image 2 Review: Prompt Guide and Use Cases
- befreed.ai — GPT Image 2: Complete Guide 2026
- Civitai's Guide to GPT Image 1
- OpenAI Developer Community — DALL-E 3 & gpt-image-1 tips thread