Complete Guide to GPT Image 2.5 Prompting and Image Editing: A Comprehensive Analysis of Core Practical Cases

Quick Overview: Substantive Upgrades in GPT Image 2.5

1. Dual-Model Division of Labor Established: gpt-image-2.5-flare is a small model optimized for speed (image quality comparable to GPT Image 2), while gpt-image-2.5-sunburst is a base model optimized for quality (image quality higher than GPT Image 2).
2. Custom Resolutions: Supports custom dimensions meeting constraints: width and height must both be multiples of 16, maximum edge of 3840px, aspect ratio not exceeding 3:1, and total pixel count between 655,000 and 8.294 million pixels (655,360 to 8,294,400).
3. Native Transparent Background Channel: Supports directly requesting transparent backgrounds (background="transparent" outputting PNG/WebP with an alpha channel), though alpha, glass, hair, shadows, and edges still require inspection, combined with post-processing or engineering compositing when necessary, avoiding reliance on fake checkerboards.
4. Structured Constraints in Edit Mode: Supports role assignment across multiple reference images (Image 1 as subject, Image 2 as clothing, Image 3 as background); prompts can separately express items to change and items to preserve, helping reduce visual drift, with results still requiring inspection each time.
5. Text and Infographic Workflows: Can be used for typography, multi-panel narrative comics, academic diagrams, pitch deck slides, and localized text translation, with generated results still requiring manual verification.

If you want to learn about the model itself first, you can read the site's GPT Image 2.5 Model Introduction; this article continues to focus on official prompting guidelines and production workflows.

OpenAI has released an official prompting guide for the latest GPT Image 2.5 family of models (including the speed-optimized small model Flare and the quality-optimized base model Sunburst).

This guide aims to make it clear that: AI image generation is shifting from single-turn text-to-image prompt tricks toward "localized editing, multi-image compositing, character consistency, and engineering delivery" workflows that clearly express constraints, improve controllability, and verify outputs step-by-step.

Whether you are a creator using ChatGPT to craft posters, comics, UI mockups, and e-commerce imagery, or a developer migrating an API from older versions (GPT Image 1 / 1.5 / 2) to 2.5, this deep dive breaks down all core mechanisms, baseline parameters, 8 foundational prompting principles, and 22 named workflows (including 9 generation, 8 editing, 2 multi-turn, and 3 additional workflows) from the official guide, providing ready-to-use engineering-grade prompt blueprints and practical toolkits.

Part 1: Model Selection and API Core Baseline Parameters

Before writing prompts, one must first clarify model selection and interface parameters. It should be noted that while general prompting methods can be referenced, Flare/Sunburst model selection as well as quality, size, and background belong to the API context and must not imply that these controls necessarily exist in the ChatGPT UI. The official guidance emphasizes: interface parameters should be explicitly specified in API requests, not mixed into prompt text.

1. Flare vs. Sunburst: How to Choose and Migrate?

Migration Decision Tree

• Scenario A (Current GPT Image 2 results already meet requirements): Integrate Flare first. While keeping existing prompts, reference images, and resolutions unchanged, verify whether end-to-end response time can be shortened while maintaining comparable image quality.
• Scenario B (Current workflow results have pain points / complex, demanding tasks): Integrate Sunburst first to establish an acceptable quality baseline; once Sunburst passes acceptance, test Flare in reverse using identical inputs. If Flare can also meet quality requirements, then evaluate the speed advantage of switching to Flare; it must be clarified that speed cannot be directly inferred as cheaper—current pricing, failure rates, retry counts, and the comprehensive cost per qualified image must be audited, and if Flare fails to meet standards, retain Sunburst.

Special Reminder: Quality tiers cannot be directly equated across models
Even if both models select the same quality="medium", it does not mean actual image detail and generation time will be identical. When making model comparisons, live testing must be conducted on the same evaluation dataset with fixed dimensions and format.

2. Core API Parameters and Resolution Mathematical Constraints

# API 典型请求结构
response = client.images.generate(
    model="gpt-image-2.5-sunburst",  # 或 gpt-image-2.5-flare
    prompt=prompt_text,
    quality="high",  # auto (默认), low, medium, high, xhigh, max
    size="1536x1024",  # 支持预设或自定义
    background="transparent",  # auto, opaque, transparent
    output_format="png",  # png, jpeg, webp
)

Key Parameter Limits and Practical Rules

1. size custom resolutions must satisfy 4 hard mathematical constraints:
   • Maximum single-edge length cannot exceed 3,840 pixels.
   • Both width and height must be integer multiples of 16 (e.g., 1024, 1536, 2048, etc.).
   • Aspect ratio (longer edge : shorter edge) must not exceed 3:1 (prohibiting extreme ultrawide strips).
   • Total pixel count must fall between 655,360 and 8,294,400 pixels.
   • Experimental Range Warning: When total pixels exceed 3,686,400 pixels (i.e., 2560x1440), the system is in an experimental support phase.
2. Quick Reference for Common Standard Resolutions:
   • Square (1:1): 1024x1024, 2048x2048 (2K)
   • Landscape (3:2 / 16:9): 1536x1024, 2048x1152 (2K), 3840x2160 (4K)
   • Portrait (2:3 / 9:16): 1024x1536, 2160x3840 (4K)
3. quality Tier Strategy:
   • Available tiers are low, medium, high, xhigh, max, and the default auto.
   • Tuning Rule: Always finalize model selection before tuning quality tiers. When encountering typography, small-font annotations, or dense mixed fonts, starting baseline tests at medium or high is recommended. Consider xhigh or max only when latency budgets permit and lower tiers genuinely fail to meet sharpness requirements.
4. background Transparent Background:
   • Transparent output can be directly requested in API calls (setting background="transparent"), and output_format must be png or webp.
   • output_compression is prohibited for PNG (this parameter applies only to JPEG and WebP).
   • Step-by-Step Acceptance and Post-Processing: Although transparent output can be requested directly, one must still inspect the alpha channel, translucent glass, hair details, shadow transitions, and edge contours, confirming the model did not generate a fake checkerboard; combine with post-processing or engineering compositing where necessary to meet delivery standards.

3. Deprecation Schedule for Older Models (Migration Warning)

If your production environment still calls older model versions, note the following shutdown timelines and migration essentials:

• Migration Validation Path: Workflows on both GPT Image 1 and 1.5 should first be validated on GPT Image 2, evaluating 2.5 only after standards are met.
• Official Size Parameter Values: For both GPT Image 1 and 1.5, size is 1024x1024, 1024x1536, 1536x1024, or auto.
• input_fidelity Parameter Details: Both support low or high for input_fidelity (where high on GPT Image 1 increases image input tokens); omit input_fidelity when migrating to GPT Image 2 because GPT Image 2 is always high-fidelity (do not write "not needed for any 2.x").
• Transparent Background Status: GPT Image 2's transparent background is in preview; note this feature status during migration evaluations.

Part 2: 8 Foundational Principles for Writing Efficient Prompts

The official guide summarizes 8 writing principles spanning basic generation to multi-turn editing. To facilitate practical rewrites and implementation benchmarking, each principle is structured below into a compact "Prompt Diagnostic Card." Refining prompts aims to express intent clearly and provide acceptance criteria, reducing the risk of unintended results; however, model execution still retains uncertainty, and generated results must still be verified item-by-item.

【模糊写法｜中文】
一个白底上的陶瓷咖啡杯产品图。

【Vague｜English】
A product photo of a ceramic coffee mug on a white background.

【可执行改写｜中文】
电商白底产品图：一只纯白色哑光陶瓷咖啡杯，置于纯白无缝背景中央，从 45 度俯角拍摄，柔和漫射光，无干扰性阴影。

【Actionable｜English】
E-commerce white-background product photography: a matte white ceramic coffee mug centered on a seamless pure white background, photographed from a 45-degree downward angle, with soft diffused lighting and no distracting shadows.

【模糊写法｜中文】
画一个在科技实验室做实验的女科学家，桌上有发光蓝色试管和全息投影，画面明亮整洁，不要杂物和文字。

【Vague｜English】
Draw a female scientist doing an experiment in a tech lab, with glowing blue test tubes and holographic projections on the desk, bright and neat, no clutter and no text.

【可执行改写｜中文】
[Scene] 明亮整洁的现代科技实验室，冷白室内环境光
[Subject] 一位身穿白色实验服的女科学家，正低头专注记录实验数据
[Details] 实验台上整齐摆放发光的蓝色试管与半透明全息分子模型
[Constraints] 画面无杂乱多余物品，背景与界面无额外文字

【Actionable｜English】
[Scene] Bright and clean modern tech laboratory, cool white ambient indoor lighting
[Subject] A female scientist in a white lab coat, looking down and focusing on recording experiment data
[Details] Glowing blue test tubes and translucent holographic molecular models neatly arranged on the workbench
[Constraints] No messy redundant items in the scene, no extra text on the background or interfaces

【模糊写法｜中文】
老木匠的高清真实肖像，质感超强，大师级光影，8k 电影感。

【Vague｜English】
A high-definition realistic portrait of an old carpenter, incredible texture, master lighting, 8k cinematic feel.

【可执行改写｜中文】
真实摄影：一位年长木匠的面部特写，皮肤带有深浅不一的干燥皱纹与微小毛孔，侧方柔和窗光勾勒轮廓，浅景深使工坊背景自然虚化（85mm 镜头外观线索）。

【Actionable｜English】
Real photograph: a close-up portrait of an elderly carpenter, showing dry wrinkles of varying depth and visible pores, soft side window light outlining facial contours, shallow depth of field naturally blurring the workshop background (85mm lens appearance cue).

【模糊写法｜中文】
一个在街上骑自行车的男子。

【Vague｜English】
A man riding a bicycle on the street.

【可执行改写｜中文】
全身构图：一位三十多岁的男子在城市林荫道上骑行，全身可见包括双脚，双手自然抓握平把公路车车把，双眼注视前方路面，双腿处于蹬踏脚踏板的动作中。

【Actionable｜English】
Full-body composition: a man in his 30s riding on a tree-lined city street, full body visible including both feet, hands naturally gripping the flat handlebars of a road bike, eyes focused on the road ahead, legs actively pedaling.

【模糊写法｜中文】
咖啡馆门头木招牌，上面写着店名 Morning Brew 和副标题。

【Vague｜English】
A wooden storefront sign for a coffee shop, with the shop name Morning Brew and a subtitle on it.

【可执行改写｜中文】
复古木质咖啡馆门头招牌，正中以清晰衬线体排版文字 "Morning Brew"，下方以较小无衬线体排版副标题 "Artisan Coffee"，两处文字各出现且仅出现一次，招牌及画面其他区域无额外杂乱文字。

【Actionable｜English】
Vintage wooden coffee shop storefront sign, centered typography in clean serif font reading "Morning Brew", with smaller sans-serif subtitle "Artisan Coffee" below it, each text rendered exactly once, with no extra clutter text anywhere on the sign or scene.

【模糊写法｜中文】
把图里人物的上衣换成红色羽绒服。

【Vague｜English】
Change the person's jacket in the image to a red down jacket.

【可执行改写｜中文】
基于原图编辑：仅修改人物上衣为一件红色连帽羽绒服（change only the upper garment to a red hooded down jacket）。严格保留人物面部五官、发型发色、站立姿态、背景街道环境及原有光照角度不变。

【Actionable｜English】
Based on original image edit: change only the subject's upper garment to a red hooded down jacket. Keep facial features, hair style and color, standing posture, background street environment, and original lighting angle strictly unchanged.

【模糊写法｜中文】
结合图 1 和图 2，做一张女人和狗在客厅里的图。

【Vague｜English】
Combine Image 1 and Image 2 to make a picture of a woman and a dog in a living room.

【可执行改写｜中文】
多图组合：图 1 作为主场景底图，保留客厅空间结构与暖色室内环境光；从图 2 中提取金毛寻回犬，放置在图 1 沙发旁的地毯上；狗的受光角度与地毯投影严格继承图 1 的光源方向。

【Actionable｜English】
Multi-image composition: Image 1 serves as the base background scene, preserving the living room spatial structure and warm indoor ambient light; extract the golden retriever from Image 2 and place it on the rug next to the sofa in Image 1; the dog's lighting angle and cast shadow on the rug must inherit the light source direction from Image 1.

【模糊写法｜中文】
把上一轮生成的图换个海滩背景，人物放大一点，光线调暗，再加一副眼镜。

【Vague｜English】
Change the background of the image generated last round to a beach, make the person slightly bigger, dim the lighting, and add glasses.

【可执行改写｜中文】
基于上一轮已验收确认的图像作为输入：本轮仅执行单项修改——为人物面部添加一副黑色细圆框眼镜（change only: add thin black round-frame glasses）。上一轮已验收确认的人物面部五官、发型、服装、站立构图与背景完全保持不变。

【Actionable｜English】
Based on the verified image from the previous round as input: execute only a single change this round—add a pair of thin black round-frame glasses to the subject's face. Keep the previously verified facial features, hairstyle, clothing, standing composition, and background strictly unchanged.

Part 3: Comprehensive Breakdown of Core Practical Cases

The official guide provides 22 named workflows (including 9 generation cases, 8 editing cases, 2 multi-turn cases, and 3 additional workflows). The original English prompts serve as practical starting points; in actual deployment, they must be adapted to specific inputs and business scenarios, and outputs still require manual inspection. Below is a breakdown of prompt design goals, key parameters, and mechanistic takeaways for each case.

Scenario 1: High-Fidelity Base Image Generation

Case 1: Realistic Portraits and Film Lighting (Control style and lighting)

• Generation Parameters: size="1024x1536", quality="medium"
• Bilingual Prompt:

【中文】
创建一张逼真的纪实抓拍照片，画面中一位年迈的水手站在一艘小型渔船上。
他有着饱经风霜的皮肤，可见清晰的皱纹、毛孔与日晒纹理，手臂上有几处褪色的传统水手刺青。
他正从容平静地理着渔网，他的狗就坐在甲板一旁。拍摄质感如同 35mm 胶片照片，平视中近景，使用 50mm 镜头。
柔和的海岸自然日光，浅景深，细腻的胶片颗粒感，自然的色彩平衡。
整张画面应给人真切自然、毫无摆拍之感，具备真实的皮肤质感、磨损的材质痕迹与日常细节。不要美颜，不要重度磨皮。

【English】
Create a photorealistic candid photograph of an elderly sailor standing on a small fishing boat.
He has weathered skin with visible wrinkles, pores, and sun texture, and a few faded traditional sailor tattoos on his arms.
He is calmly adjusting a net while his dog sits nearby on the deck. Shot like a 35mm film photograph, medium close-up at eye level, using a 50mm lens.
Soft coastal daylight, shallow depth of field, subtle film grain, natural color balance.
The image should feel honest and unposed, with real skin texture, worn materials, and everyday detail. No glamorization, no heavy retouching.

• Core Mechanism Breakdown:
  • Texture Negative Suppression: Explicitly appending No glamorization, no heavy retouching at the end directs the model toward natural textures, helping reduce tendencies toward over-beautification.
  • Physical Cues: Explicitly listing "weathered skin with visible wrinkles, pores, and sun texture, and a few faded traditional sailor tattoos" sets concrete prompt targets for micro-surface details.
  • Single-Sample Observation: In the official single-run comparative sample, Flare and Sunburst demonstrated realistic candid effects; in this specific sample, Sunburst appeared more refined in beard texture, net weave edges, and background depth layering. However, single-sample results are for reference only and cannot be extrapolated into definitive conclusions about overall model superiority.

Official examples | Inputs and outputs (single official samples for inspection, not stable model rankings):

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Result comparison · GPT Image 2.5 Flare single official output	Figure：Result comparison · GPT Image 2.5 Sunburst single official output

Case 2: Process Diagrams and Infographics (Explain a process visually)

• Generation Parameters: size="1024x1536", quality="medium"
• Bilingual Prompt:

【中文】
制作一张详细的信息图，展示类似 Jura 全自动咖啡机的运作原理与工作流程。
从豆仓、研磨，到称重计量、水箱、锅炉等。
我希望从技术和视觉层面清晰理解其工作流程。

【English】
Create a detailed Infographic of the functioning and flow of an automatic coffee machine like a Jura.
From bean basket, to grinding, to scale, water tank, boiler, etc.
I'd like to understand technically and visually the flow.

• Core Mechanism Breakdown:
  • System-Level Guidance: Explicitly specifies the machine type (Jura automatic coffee machine) and core internal component flow (bean basket -> grinding -> scale -> water tank -> boiler).
  • Inspection Focus: After generating such diagrams, pipeline flows and logical component connections must be manually validated, without relying entirely on the AI's spontaneous understanding of mechanical principles.

Official examples | Inputs and outputs (single official samples for inspection, not stable model rankings):

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Result comparison · GPT Image 2.5 Flare single official output	Figure：Result comparison · GPT Image 2.5 Sunburst single official output

Case 3: Accurate Ad Tagline Typography (Render exact text)

• Generation Parameters: size="1024x1536", quality="medium"
• Bilingual Prompt:

【中文】
为一个名为 Thread 的品牌制作一张具有潮流文化感、酷炫的广告 / 时尚大片。
这是一个面向年轻人的时髦街头潮流品牌。广告展示了一群好友聚在一起玩乐的场景，并带有标语 "Yours to Create."。
让它看起来像是一张面向年轻街头服饰受众的高水准商业宣传大片：时尚、前卫、充满活力且富有品味。
运用干净利落的构图、鲜明强烈的色彩走向、自然的姿态以及高端时尚摄影的视觉质感。
将标语清晰可辨地渲染一次，且仅渲染一次，并自然融入广告版式布局中。
不要多余文字，不要水印，不要无关 logo。

【English】
Give me a cool in culture ad / fashion shot for a brand called Thread.
It's a hip young street brand. The ad shows a group of friends hanging out together with the tagline "Yours to Create."
Make it feel like a polished campaign image for a youth streetwear audience: stylish, contemporary, energetic, and tasteful.
Use clean composition, strong color direction, natural poses, and premium fashion photography cues.
Render the tagline exactly once, clearly and legibly, integrated into the ad layout.
No extra text, no watermarks, no unrelated logos.

• Core Mechanism Breakdown:
  • Explicit Text Frequency Constraint: Render the tagline exactly once clearly specifies tagline occurrence frequency, helping reduce instances of repeated, scattered text fragments across poster corners; text spelling and layout must be audited verbatim prior to delivery.
  • Natural Layout Integration: integrated into the ad layout explicitly requires text to blend into the overall poster layout hierarchy rather than floating haphazardly over the scene.

Official examples | Inputs and outputs (single official samples for inspection, not stable model rankings):

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Result comparison · GPT Image 2.5 Flare single official output	Figure：Result comparison · GPT Image 2.5 Sunburst single official output

Case 4: Vector-Style Transparent Background Logo (Design a reusable logo)

• Generation Parameters: size="1024x1536", quality="medium", background="transparent", output_format="png", n=1
• Bilingual Prompt:

【中文】
为一家名为 Field & Flour 的本地烘焙工坊设计一个原创、无侵权风险的 Logo。
该 Logo 应传递出温暖、简洁与历久弥新的质感。采用干净利落的类矢量图形、鲜明的轮廓剪影以及平衡的留白负空间。
追求极简而非繁复细节，以便在各种缩放尺寸下都能清晰辨识。扁平化设计，极简线条笔触，非必要不使用渐变。
完全透明的背景。交付单个居中 Logo，四周留白充裕，带有干净利落的 Alpha 边缘，无实体背景、场景、假棋盘格或水印。

【English】
Create an original, non-infringing logo for a company called Field & Flour, a local bakery.
The logo should feel warm, simple, and timeless. Use clean, vector-like shapes, a strong silhouette, and balanced negative space.
Favor simplicity over detail so it reads clearly at small and large sizes. Flat design, minimal strokes, no gradients unless essential.
Fully transparent background. Deliver a single centered logo with generous padding, clean alpha edges, and no solid backdrop, scenery, checkerboard, or watermark.

• Core Mechanism Breakdown:
  • Background Exclusions: no solid backdrop, scenery, checkerboard, or watermark explicitly demands no solid backgrounds, scenery, fake checkerboards, or watermarks; inspection is still required after generation.
  • Padding Requirements: generous padding serves to reduce the risk of logo boundaries being clipped by the canvas frame, with results still requiring verification.

Official examples | Inputs and outputs (single official samples for inspection, not stable model rankings):

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Logo variation 1 · GPT Image 2.5 Flare single official output	Figure：Logo variation 1 · GPT Image 2.5 Sunburst single official output

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Logo variation 2 · GPT Image 2.5 Flare single official output	Figure：Logo variation 2 · GPT Image 2.5 Sunburst single official output

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Logo variation 3 · GPT Image 2.5 Flare single official output	Figure：Logo variation 3 · GPT Image 2.5 Sunburst single official output

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Logo variation 4 · GPT Image 2.5 Flare single official output	Figure：Logo variation 4 · GPT Image 2.5 Sunburst single official output

Case 5: Historical Scenes and Real-World Context (Use historical and real-world context)

• Generation Parameters: size="1024x1536", quality="medium"
• Bilingual Prompt:

【中文】
生成一张写实的户外人群聚集场景，时间地点为 1969 年 8 月 16 日的纽约州贝塞尔。
逼真写实，具有符合该历史时期的服装、舞台布局与现场环境。

【English】
Create a realistic outdoor crowd scene in Bethel, New York on August 16, 1969.
Photorealistic, period-accurate clothing, staging, and environment.

• Core Mechanism Breakdown:
  • Historical Context Cues: The exact date and location (August 16, 1969 in Bethel, New York) provide concrete historical context cues for the model, steering the image toward the specific historical era.
  • Era Accuracy and Historical Fact-Checking: The prompt uses period-accurate clothing, staging, and environment to explicitly request period fidelity; however, clothing, staging, and environmental details in the generated image must still be verified manually against historical sources—the prompt itself does not guarantee historical accuracy.

Official examples | Inputs and outputs (single official samples for inspection, not stable model rankings):

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Result comparison · GPT Image 2.5 Flare single official output	Figure：Result comparison · GPT Image 2.5 Sunburst single official output

Scenario 2: Narrative, Interface, and Academic Productivity Graphics

Case 6: Four-Panel Narrative Comic Strip (Turn a story into a comic strip)

• Generation Parameters: size="1024x1536", quality="medium"
• Bilingual Prompt:

【中文】
创作一组 4 格竖版漫画风格的连环画短片。
分镜 1：主人走出正门离开。宠物被框在主人身后的窗户中，贴在玻璃前显得娇小，睁大眼睛，爪子高高扒在窗上，屋里突然安静下来。
分镜 2：门咔哒一声关上。寂静被打破。宠物缓缓转身面向空荡荡的屋子，姿态转变，眼神因各种可能性而变得敏锐。
分镜 3：整间屋子大变样。宠物像主人一样大摇大摆瘫躺在沙发上，身旁散落着零食碎屑，阳光如聚光灯般洒满房间。
分镜 4：门开了。宠物端端正正地坐在玄关处，警惕而从容，仿佛什么都没发生过一样。

【English】
Create a short vertical comic-style reel with 4 panels.
Panel 1: The owner leaves through the front door. The pet is framed in the window behind them, small against the glass, eyes wide, paws pressed high, the house suddenly quiet.
Panel 2: The door clicks shut. Silence breaks. The pet slowly turns toward the empty house, posture shifting, eyes sharp with possibility.
Panel 3: The house transformed. The pet sprawls across the couch like it owns the place, crumbs nearby, sunlight cutting across the room like a spotlight.
Panel 4: The door opens. The pet is seated perfectly by the entrance, alert and composed, as if nothing happened.

• Core Mechanism Breakdown:
  • Visual Beats & Panel Language: Avoid vague plot summaries; composition, action, and lighting must be defined individually from Panel 1 through Panel 4.
  • The dynamic transitions—from the pet's reluctant parting to wreaking havoc in the living room to instantaneous disguise—rely on explicit physical action descriptions within each panel.

Official examples | Inputs and outputs (single official samples for inspection, not stable model rankings):

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Result comparison · GPT Image 2.5 Flare single official output	Figure：Result comparison · GPT Image 2.5 Sunburst single official output

Case 7: Realistic Mobile App UI Prototype (Create an interface preview)

• Generation Parameters: size="1024x1536", quality="medium"
• Bilingual Prompt:

【中文】
为本地农贸集市制作一个逼真的移动端 App UI 原型展示图。
展示今日集市界面，包含一个简洁的页头、带有小缩略图和品类标签的摊商简短列表、一个精巧的“Today’s specials”专区，以及地点与营业时间的基础信息。
整体设计需实用且易用。纯白背景、柔和自然的点缀色、清晰的字体排版与极简装饰。
外观应如同一款真实上线、设计精良且美观的本地小型集市 App。
将该 UI 原型置于一个 iPhone 机模框架中呈现。

【English】
Create a realistic mobile app UI mockup for a local farmers market.
Show today’s market with a simple header, a short list of vendors with small photos and categories, a small “Today’s specials” section, and basic information for location and hours.
Design it to be practical, and easy to use. White background, subtle natural accent colors, clear typography, and minimal decoration.
It should look like a real, well-designed, beautiful app for a small local market.
Place the UI mockup in an iPhone frame.

• Core Mechanism Breakdown:
  • Rejecting Conceptual Sketch Jargon: Avoid terms like "UI concept/wireframe" in favor of real, well-designed, beautiful app alongside specific functional component names (header, vendor list, specials, hours), while specifying Place the UI mockup in an iPhone frame.

Official examples | Inputs and outputs (single official samples for inspection, not stable model rankings):

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Result comparison · GPT Image 2.5 Flare single official output	Figure：Result comparison · GPT Image 2.5 Sunburst single official output

Case 8: High School Biology Cellular Respiration Diagram (Scientific and educational visuals)

• Generation Parameters: size="1536x1024", quality="high"
• Bilingual Prompt:

【中文】
为高中生制作一张标题为 "Cellular Respiration at a Glance" 的简明生物学图解。

展示葡萄糖如何在细胞内部转化为能量。包含糖酵解、三羧酸循环以及电子传递链。
使用箭头连接各个步骤，并标注主要分子：glucose、pyruvate、ATP、NADH、FADH2、CO2、O2 和 H2O。
使其看起来像一份干净清晰的课堂讲义或幻灯片，具备白色背景、简洁图标、明确标签和易读文字。

避免微小字体、多余装饰或任何让图解难以理解的元素。

【English】
Create a simple biology diagram titled "Cellular Respiration at a Glance" for high school students.

Show how glucose turns into energy inside a cell. Include glycolysis, the Krebs cycle, and the electron transport chain.
Use arrows to connect the steps, and label the main molecules: glucose, pyruvate, ATP, NADH, FADH2, CO2, O2, and H2O.
Make it look like a clean classroom handout or slide, with a white background, simple icons, clear labels, and easy-to-read text.

Avoid tiny text, extra decoration, or anything that makes the diagram hard to understand.

• Core Mechanism Breakdown:
  • Instructional Design Brief: Geared toward a high school educational scenario, explicitly specifying the three core stages (glycolysis, the Krebs cycle, electron transport chain) and key molecular abbreviations.
  • Parameter and Scientific Logic Verification: Setting quality="high" is merely a comparative quality setting that aids detail sharpness, but does not guarantee accurate text spelling, arrow directions, or scientific relationships; all biochemical pathways and chemical formulas must be manually verified by subject matter experts.

Official examples | Inputs and outputs (single official samples for inspection, not stable model rankings):

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Result comparison · GPT Image 2.5 Flare single official output	Figure：Result comparison · GPT Image 2.5 Sunburst single official output

Case 9: YC-Style Pitch Deck Slide (Build slides, diagrams, and charts)

• Generation Parameters: size="1536x864", quality="high"

Fictional Data Notice: The $42B / $8.7B / $340M figures and footnotes "AGI Research, 2024" and "Internal analysis" in the prompt are fictional design inputs used in the official guide to demonstrate chart layout. In actual business deployment, they must be replaced with genuinely verified figures and citations; a highly polished visual presentation never equates to data credibility.

• Bilingual Prompt:

【中文】
制作一张标题为 **"Market Opportunity"** 的商业计划书幻灯片，质感如同出自 YC 孵化初创公司的真实 A 轮融资路演幻灯片。

采用干净的纯白背景、现代无衬线字体（如 Inter），以及利落极简的版式。幻灯片应包含：

* 一个采用低饱和度蓝灰配色的 TAM/SAM/SOM 同心圆图
* 具体、可信的市场规模数据：
  * **TAM:** $42B
  * **SAM:** $8.7B
  * **SOM:** $340M
* 下方配有一张简洁的柱状图，展示 **2021 to 2026** 的市场增长，呈现平缓上升趋势
* 小字号注脚：**"AGI Research, 2024"** 与 **"Internal analysis"**
* 右下角放置一个公司 Logo 占位符

整体设计应呈现出真正成功融到资的路演 PPT 质感：文字高度可读、数据层级分明、间距考究、采用专业的初创公司视觉语言。

避免剪贴画、图库照片、渐变、阴影、装饰性元素，或任何显得平庸或过度设计的元素。

【English】
Create one pitch-deck slide titled **"Market Opportunity"** that feels like a real Series A fundraising slide from a YC-backed startup.

Use a clean white background, modern sans-serif typography like Inter, and a crisp, minimal layout. The slide should include:

* A TAM/SAM/SOM concentric-circle diagram in muted blues and grays
* Specific, believable market sizing numbers:
  * **TAM:** $42B
  * **SAM:** $8.7B
  * **SOM:** $340M
* A clean bar chart below showing market growth from **2021 to 2026**, with a subtle upward trend
* Small footnotes: **"AGI Research, 2024"** and **"Internal analysis"**
* A company logo placeholder in the bottom-right corner

The design should look like it belongs in a deck that actually raised money: highly readable text, clear data hierarchy, polished spacing, and professional startup-style visual language.

Avoid clip art, stock photography, gradients, shadows, decorative elements, or anything that feels generic or overdesigned.

• Core Mechanism Breakdown:
  • Layout Hierarchy and Structural Guidance: Designating clean typography like Inter, a TAM/SAM/SOM concentric diagram, and a temporal trend bar chart builds a crisp visual information hierarchy while explicitly filtering out gratuitous ornamentation and drop shadows.
  • Separating Visual Polish from Factual Truth: While the model excels at rendering highly polished presentation styling, chart proportions, numerical authenticity, and footnote sources remain purely decorative placeholders; they cannot serve as commercial evidence and must be rigorously verified and replaced in production.
  • Writing to a Specification: Avoids vague adjectives in favor of concrete metrics ($42B, $8.7B, $340M), timeframes (2021-2026), chart geometries (concentric circles, bar charts), and verbatim footnote text.
  • Layout Purity: Explicitly strips out clip art, gradients, drop shadows, and generic illustration styles.

Official examples | Inputs and outputs (single official samples for inspection, not stable model rankings):

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Result comparison · GPT Image 2.5 Flare single official output	Figure：Result comparison · GPT Image 2.5 Sunburst single official output

Scenario 3: Precision Image Editing and Multi-Image Composition

In image editing workflows, the invocation method switches to client.images.edit.

Case 10: Translate While Preserving Layout (Translate while preserving layout)

• Input Image: Coffee machine diagram generated in Case 2
• Editing Parameters: size="1024x1536", quality="high"
• Bilingual Prompt:

【中文】
将信息图中的文字翻译成西班牙语。不要改动图像的任何其他方面。

【English】
Translate the text in the infographic to Spanish. Do not change any other aspect of the image.

• Core Mechanism Breakdown:
  • In-Place Text Replacement Intent and Layout Verification: The prompt's objective is to replace text with the target language while preserving original layout and graphic elements. However, the model does not guarantee complete text-image separation or invariant backgrounds; generated outputs require strict inspection for untranslated text, leftover source language, or unintended alterations to mechanical lines, icons, arrows, or overall layout geometry.

Official examples | Inputs and outputs (single official samples for inspection, not stable model rankings):

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Result comparison · GPT Image 2.5 Flare single official output	Figure：Result comparison · GPT Image 2.5 Sunburst single official output

Case 11: Extract and Transfer Visual Style to a New Subject (Transfer a visual style)

• Input Image: Pixel art game reference image (pixels.webp)
• Editing Parameters: size="1024x1536", quality="medium"
• Bilingual Prompt:

【中文】
沿用输入图片的视觉风格，在白色背景上生成一位骑摩托车的男子。

【English】
Use the same style from the input image and generate a man riding a motorcycle on a white background.

• Core Mechanism Breakdown:
  • Style Transfer Objectives and Element Verification: The prompt aims to decouple the visual style from the original subject, requesting extraction of the "color palette, pixel grid grain, and retro medium texture" while swapping the subject and background to a new description (motorcyclist on a white background). Sustaining the identical style is a requested goal rather than an absolute guarantee; color fidelity, pixel grit, and silhouette integrity of the new subject must still be audited post-generation.

Official examples | Inputs and outputs (single official samples for inspection, not stable model rankings):

Figure：Input / reference · pixel-art game screen

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Result comparison · GPT Image 2.5 Flare single official output	Figure：Result comparison · GPT Image 2.5 Sunburst single official output

Case 12: Change Clothing While Preserving Identity and Proportions (Preserve identity and change clothing)

• Single-Run Generation Output Examples (Official single samples for comparative observation, not stable fidelity baselines):
• Editing Parameters: size="1024x1536", quality="medium"
• Bilingual Prompt:

【中文】
编辑这张图片，使用提供的服装图片为这位女士换装。切勿以任何方式改变她的面部、五官、肤色、身形、姿势或身份特征。完全保留她原有的容貌长相、表情、发型与身体比例。仅替换服装，使衣物根据她现有的姿势与身体几何形态自然贴合，并呈现逼真的面料动态。使光照、阴影和色温与原照片保持一致，让整套服装写实自然地融入画面，毫无粘贴拼接感。不要改变背景、相机视角、画面构图或图像质量，也不要添加任何配饰、文字、Logo 或水印。

【English】
Edit the image to dress the woman using the provided clothing images. Do not change her face, facial features, skin tone, body shape, pose, or identity in any way. Preserve her exact likeness, expression, hairstyle, and proportions. Replace only the clothing, fitting the garments naturally to her existing pose and body geometry with realistic fabric behavior. Match lighting, shadows, and color temperature to the original photo so the outfit integrates photorealistically, without looking pasted on. Do not change the background, camera angle, framing, or image quality, and do not add accessories, text, logos, or watermarks.

• Core Mechanism Breakdown:
  • Explicit Preservation Constraints and Checklist Verification: The prompt sets out an exhaustive preservation inventory covering facial features, skin tone, body shape, posture, hairstyle, anatomical proportions, background environment, and camera framing. These represent explicit retention requests; post-generation verification against the source image is necessary to ensure facial likeness and body geometry have not warped.
  • Physical Conformance and Lighting Integration: Requiring garments to drape and wrinkle naturally across the subject's geometry (realistic fabric behavior) while matching original environmental color temperature and shadows helps reduce pasted-on artifacts; manual checks remain essential to confirm contact seams and shadow falloff appear physically authentic.

Official examples | Inputs and outputs (single official samples for inspection, not stable model rankings):

Input / reference	Input / reference
Figure：Input / reference · woman in a museum	Figure：Input / reference · beige jacket

Input / reference	Input / reference
Figure：Input / reference · white tank top	Figure：Input / reference · gray boots

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Result comparison · GPT Image 2.5 Flare single official output	Figure：Result comparison · GPT Image 2.5 Sunburst single official output

Case 13: Multi-Photo Subject Extraction and Compositing (Combine references)

• Input Images: Two photos (Image 1: Woman in street setting; Image 2: Photo of woman holding dog)
• Editing Parameters: size="1024x1536", quality="medium"
• Bilingual Prompt:

【中文】
将第二张图片中的狗放置在图 1 的场景中，紧挨着那位女士身旁，采用与图 1 相同的光影风格、构图和背景。不要改变其他任何内容。

【English】
Place the dog from the second image into the setting of image 1, right next to the woman, use the same style of lighting, composition and background. Do not change anything else.

• Core Mechanism Breakdown:
  • Cross-Image Migration Intent and Lighting Acceptance: Explicitly establishes the cross-image transfer route, pulling the dog from Image 2 and compositing it beside the woman in Image 1 to inherit Image 1's lighting, framing, and background setting. Despite prompting to exclude all other edits, composited outputs must be audited for scale proportions, contact shadows, and illumination consistency with the street environment.

Official examples | Inputs and outputs (single official samples for inspection, not stable model rankings):

Input / reference	Input / reference
Figure：Input / reference · woman in a street scene	Figure：Input / reference · woman with a dog

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Result comparison · GPT Image 2.5 Flare single official output	Figure：Result comparison · GPT Image 2.5 Sunburst single official output

Case 14: Transparent E-Commerce Product Cutout for Shampoo Bottle (Create a transparent product cutout)

• Single-Run Generation Output Examples (Official single samples for comparative observation, not stable fidelity baselines):
• Editing Parameters: size="1024x1536", quality="medium", background="transparent", output_format="png"
• Bilingual Prompt:

【中文】
从输入图片中提取产品主体，并将其孤立放置在完全透明的背景上。
输出要求：产品居中，轮廓清晰利落，无光晕/杂色边缘。
严格保持产品的几何形态与标签的可读性。
仅做轻度润饰修整。不要添加纯色背景、棋盘格、场景或阴影。
不要重新设计产品风格；移除背景并保持干净的 Alpha 透明通道。

【English】
Extract the product from the input image and isolate it on a fully transparent background.
Output: centered product, crisp silhouette, no halos/fringing.
Preserve product geometry and label legibility exactly.
Add only light polishing. Do not add a solid backdrop, checkerboard, scenery, or shadow.
Do not restyle the product; remove the background and preserve clean alpha transparency.

• Core Mechanism Breakdown:
  • Edge Suppression and Alpha Channel Demands: The prompt specifies crisp silhouette, no halos/fringing with the intent to eliminate matte fringing and color bleeding while exporting a clean alpha channel.
  • Geometric Fidelity Demands and Channel Verification: Although the instruction explicitly requires preserving product geometry and label legibility, transparent cutout outputs must still be verified item-by-item: confirm whether the alpha channel is genuinely clean without background residue, whether bottle silhouettes have edge burrs or fringing, whether geometric proportions remain symmetrical, and whether bottle typography remains legible (single samples serve only to illustrate cutout texture, not as verbatim text validation or stable fidelity guarantees).

Official examples | Inputs and outputs (single official samples for inspection, not stable model rankings):

Figure：Input / reference · original shampoo product photograph

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Result comparison · GPT Image 2.5 Flare single official output	Figure：Result comparison · GPT Image 2.5 Sunburst single official output

Case 15: Convert Line Drawing into Realistic Landscape Render (Turn a drawing into a realistic image)

• Input Image: Simple river valley line sketch (drawings.webp)
• Editing Parameters: size="1024x1536", quality="medium"
• Bilingual Prompt:

【中文】
将这幅手绘草图转换为写实照片级图像。
完全保留原有的构图布局、比例与透视关系。
选用与草图意图一致的真实材质与光照。
不要添加新元素或文字。

【English】
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.

• Core Mechanism Breakdown:
  • Compositional Guidance and Topological Verification: The sketch layout, proportions, and perspective act as guiding targets, instructing the model to fill in realistic rock textures, water reflections, and vegetation shadows in accordance with sketch intent. However, the model does not guarantee invariant spatial topology; mountain contours, river curvature, and depth perspective must be inspected for shifts or distortion.

Official examples | Inputs and outputs (single official samples for inspection, not stable model rankings):

Figure：Input / reference · river-valley line drawing

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Result comparison · GPT Image 2.5 Flare single official output	Figure：Result comparison · GPT Image 2.5 Sunburst single official output

Case 16: Localized Object Removal (Remove an object)

• Input Image: Man in blue hat holding flowers
• Editing Parameters: size="1024x1536", quality="medium"
• Bilingual Prompt:

【中文】
移除男子手中的花朵。不要改变其他任何内容。

【English】
Remove the flower from man's hand. Do not change anything else.

• Core Mechanism Breakdown:
  • Concise Directives and Inpainting Verification: When edit goals are strictly localized, concise single-sentence directives help minimize unintended shifts in unrelated areas. However, one cannot assume the model will seamlessly reconstruct hand anatomy after removing the flowers; inspections must focus on knuckle articulation, palm geometry, sleeve cuff contours, and inpainting artifacts in the background.

Official examples | Inputs and outputs (single official samples for inspection, not stable model rankings):

Figure：Input / reference · man in a blue cap holding a flower

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Result comparison · GPT Image 2.5 Flare single official output	Figure：Result comparison · GPT Image 2.5 Sunburst single official output

Case 17: Insert Portrait Subject into High-Action Thriller Scene (Insert a person into a scene)

• Input Image: Portrait of woman in museum (woman-in-museum.webp)
• Editing Parameters: size="1024x1536", quality="medium"
• Bilingual Prompt:

【中文】
生成一个高度逼真的动态场景：画面中的人物正在逃离一只袭击露营地的大型写实棕熊。画面看起来应该像某人现场抓拍的真实照片，而非过度渲染或带有电影感的电影海报。
她位于画面中心但视线背对镜头，身穿户外露营服装，脸上带有泥土，衣物上有破损撕痕。她明显心怀恐惧但一心专注于逃生，在身后的棕熊摧毁营地时全力奔跑逃离。
露营地位于优胜美地国家公园（Yosemite National Park），带有真实可信的自然细节。时间为黄昏，采用自然光线与逼真色彩。一切都应显得真实、纯粹、未经刻意造型，宛如抓拍到的现实瞬间。避免电影感布光、戏剧化调色或风格化构图。

【English】
Generate a highly realistic action scene where this person is running away from a large, realistic brown bear attacking a campsite. The image should look like a real photograph someone could have taken, not an overly enhanced or cinematic movie-poster image.
She is centered in the image but looking away from the camera, wearing outdoorsy camping attire, with dirt on her face and tears in her clothing. She is clearly afraid but focused on escaping, running away from the bear as it destroys the campsite behind her.
The campsite is in Yosemite National Park, with believable natural details. The time of day is dusk, with natural lighting and realistic colors. Everything should feel grounded, authentic, and unstyled, as if captured in a real moment. Avoid cinematic lighting, dramatic color grading, or stylized composition.

• Core Mechanism Breakdown:
  • Anti-Cinematic Style Steering and Realism Verification: Repeatedly emphasizing not an overly enhanced or cinematic movie-poster image and Avoid cinematic lighting, dramatic color grading seeks to steer the output toward authentic field-snapshot lighting, facial grit, and torn clothing. Such anti-cinematic phrasing serves as stylistic steering and does not guarantee genuine photographic realism; character identity consistency, anatomical plausibility of running postures for both the woman and bear, and campsite perspective must be rigorously verified.

Official examples | Inputs and outputs (single official samples for inspection, not stable model rankings):

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Result comparison · GPT Image 2.5 Flare single official output	Figure：Result comparison · GPT Image 2.5 Sunburst single official output

Scenario 4: Multi-Turn Progressive Refinement and Character Consistency

Case 18: Single-Variable Refinement of Outdoor Billboard from Sunset to Snowy Night (Refine an image across turns)

• Turn 1 (Generate Sunset Poster):

【中文】
在日落时分的公路场景中，为这款洗发水创建一个逼真的户外广告牌样机。
广告牌文字（精确逐字，无多余字符）：
"Fresh and clean"
排版样式：粗体无衬线字体，高对比度，居中对齐，字间距规整清晰。
确保文字仅出现一次且清晰易读。
无水印，无 Logo。

【English】
Create a realistic billboard mockup of the shampoo on a highway scene during sunset.
Billboard text (EXACT, verbatim, no extra characters):
"Fresh and clean"
Typography: bold sans-serif, high contrast, centered, clean kerning.
Ensure text appears once and is perfectly legible.
No watermarks, no logos.

• Turn 1 Single-Run Output Example (Official single sample for comparative observation, not a stable baseline):
• Turn 2 (Single-Variable Environmental Change Using Turn 1 Output as Input Image):

【中文】
将其调整为雪花纷飞的冬日傍晚景象。

【English】
Make it look like a winter evening with snowfall.

• Turn 2 Single-Run Output Example (Official single sample for comparative observation, not a zero-drift benchmark):
• Core Mechanism Breakdown:
  • Stepwise Control and Drift Inspection: Breaking complex prompts into sequential stages—locking billboard structure and typography first, then altering season and weather in a concise sentence—reduces control complexity per generation. However, stepwise refinement does not guarantee zero drift on original subjects; billboard lettering spelling, typography alignment, and bottle geometry must be carefully inspected across turns for inadvertent distortion.

Official examples | Inputs and outputs (single official samples for inspection, not stable model rankings):

Figure：Input / reference · original shampoo product photograph

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Turn 1: sunset billboard · GPT Image 2.5 Flare single official output	Figure：Turn 1: sunset billboard · GPT Image 2.5 Sunburst single official output

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Turn 2: snowy-evening billboard · GPT Image 2.5 Flare single official output	Figure：Turn 2: snowy-evening billboard · GPT Image 2.5 Sunburst single official output

Case 19: Preserving Children's Picture Book Character Consistency Across Scenes (Keep a character consistent)

• Stage 1: Establish Character Identity and Baseline Visuals:

【中文】
创作一幅介绍主角的儿童绘本插画。

角色：
一个年轻的童话故事风格小英雄，灵感源自森林小侠客，
身穿简朴的绿色连帽束腰上衣，脚穿柔软的棕色靴子，佩戴随身小腰包。
角色神情善良，目光温柔，举止勇敢而温暖。
随身携带一把仅用于助人、从不伤人的小木弓。

主题：
该角色保护并救助松鼠、小鸟与野兔等森林小动物。

风格：
儿童绘本插画，手绘水彩质感，
柔和的轮廓线，温暖的大地色系，充满童趣且亲切友好。
适合图画书的人体比例（头部略大，面容富有表现力）。

约束：
- 原创角色（无受版权保护的角色）
- 无文字
- 无水印
- 质朴的森林背景，以清晰展示角色主体

【English】
Create a children’s book illustration introducing a main character.

Character:
A young, storybook-style hero inspired by a little forest outlaw,
wearing a simple green hooded tunic, soft brown boots, and a small belt pouch.
The character has a kind expression, gentle eyes, and a brave but warm demeanor.
Carries a small wooden bow used only for helping, never harming.

Theme:
The character protects and rescues small forest animals like squirrels, birds, and rabbits.

Style:
Children’s book illustration, hand-painted watercolor look,
soft outlines, warm earthy colors, whimsical and friendly.
Proportions suitable for picture books (slightly oversized head, expressive face).

Constraints:
- Original character (no copyrighted characters)
- No text
- No watermarks
- Plain forest background to clearly showcase the character

• Stage 2: Use Baseline Image as Input to Develop New Story Plot:

【中文】
使用同一个角色继续展开儿童绘本故事。

场景：
冬日风暴过后，这位森林小英雄正温柔地帮助一只受惊的松鼠
从倒下的树干中脱困。
角色跪在松鼠身旁，给予它安慰与安全感。

角色一致性：
- 相同的绿色连帽束腰上衣
- 相同的面部特征、比例与调色板
- 相同温柔而勇敢的个性

风格：
儿童绘本水彩插画，
柔和的光线，积雪覆盖的森林环境，
温馨治愈的氛围。

约束：
- 切勿重新设计角色
- 无文字
- 无水印

【English】
Continue the children’s book story using the same character.

Scene:
The same young forest hero is gently helping a frightened squirrel
out of a fallen tree after a winter storm.
The character kneels beside the squirrel, offering reassurance.

Character Consistency:
- Same green hooded tunic
- Same facial features, proportions, and color palette
- Same gentle, heroic personality

Style:
Children’s book watercolor illustration,
soft lighting, snowy forest environment,
warm and comforting mood.

Constraints:
- Do not redesign the character
- No text
- No watermarks

• Core Mechanism Breakdown:
  • Consistency Declarations and Drift Verification: When passing the baseline image in Stage 2, explicitly structuring a Character Consistency block that enumerates "same green hooded tunic, same facial features, proportions, and color palette" alongside Do not redesign the character anchors retention goals. While this constraint structure lowers feature divergence probabilities, it does not entirely prevent drift; facial features, hair, and costume accessories must still be audited manually against baseline references for continuity.

Official examples | Inputs and outputs (single official samples for inspection, not stable model rankings):

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Stage 1: character baseline · GPT Image 2.5 Flare single official output	Figure：Stage 1: character baseline · GPT Image 2.5 Sunburst single official output

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Stage 2: snowy story continuation · GPT Image 2.5 Flare single official output	Figure：Stage 2: snowy story continuation · GPT Image 2.5 Sunburst single official output

Scenario 5: Commercial Retail and Interior Furnishing Swaps (Cases 20-22)

Case 20: Interior Furniture Replacement (Change furniture in a room)

• Objective: Perform surgical localized replacement (surgical realism), modifying only target objects while explicitly retaining camera angle, room lighting, floor shadows, and surrounding environment to deliver authentic photographic realism rather than total reconstruction.
• Editing Settings: size="1536x1024", quality="medium"
• Bilingual Prompt:

【中文】
在这张房间照片中，仅将白色的椅子替换为木质椅子。
保留相机视角、室内采光、地面阴影与周围物体。
保持图像的所有其他方面完全不变。
呈现逼真写实的接触阴影与面料材质纹理。

【English】
In this room photo, replace ONLY the white chairs with chairs made of wood.
Preserve camera angle, room lighting, floor shadows, and surrounding objects.
Keep all other aspects of the image unchanged.
Photorealistic contact shadows and fabric texture.

Key Takeaways: Explicitly restricts modification to "ONLY the white chairs" while listing angles, lighting, and peripheral objects to be preserved; Photorealistic contact shadows and fabric textures articulate contact shadow requirements, and post-generation inspection must check whether furniture floats or adjacent areas experience drift.

Official examples | Inputs and outputs (single official samples for inspection, not stable model rankings):

Figure：Input / reference · original kitchen with white chairs

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Result comparison · GPT Image 2.5 Flare single official output	Figure：Result comparison · GPT Image 2.5 Sunburst single official output

Case 21: Vintage Teddy Bear Holiday Greeting Card (Design a holiday card)

• Objective: When conceptualizing holiday cards, comprehensively specify scene composition, emotional tone, materials, lighting, and exact copy.
• Generation Settings: size="1024x1536", quality="medium"
• Bilingual Prompt:

【中文】
创作一幅圣诞节日贺卡插画。

场景：
一个温馨的圣诞场景：一只旧泰迪熊坐在纪念盒中，绒毛略有磨损，带有细柔的缝补针脚，放置在窗户旁，窗外飘着雪花。场景暗示当年的孩子已经长大，但美好回忆犹存。

情绪：
温暖、怀旧、温柔、深情。

风格：
高级节日贺卡摄影质感，柔和的电影级布光，
写实质感，浅景深，
雅致的背景虚化光斑，高印刷品质的构图。

约束：
- 仅限原创艺术作品
- 无商标
- 无水印
- 无 Logo

仅包含此卡片文字（逐字一致）：
"Merry Christmas — some memories never fade."

【English】
Create a Christmas holiday card illustration.

Scene:
a cozy Christmas scene with an old teddy bear sitting inside a keepsake box, slightly worn fur, soft stitching repairs, placed near a window with falling snow outside. The scene suggests the child has grown up, but the memories remain.

Mood:
Warm, nostalgic, gentle, emotional.

Style:
Premium holiday card photography, soft cinematic lighting,
realistic textures, shallow depth of field,
tasteful bokeh lights, high print-quality composition.

Constraints:
- Original artwork only
- No trademarks
- No watermarks
- No logos

Include ONLY this card text (verbatim):
"Merry Christmas — some memories never fade."

Key Takeaways: Breaks down scene elements (old teddy bear in a keepsake box, worn fur, soft stitches, falling snow outside window), emotional mood, and photographic style (soft cinematic lighting, shallow depth of field, bokeh), adds originality and trademark-free constraints, and explicitly restricts copy via Include ONLY this card text (verbatim):; verbatim inspection remains necessary after generation.

Official examples | Inputs and outputs (single official samples for inspection, not stable model rankings):

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Result comparison · GPT Image 2.5 Flare single official output	Figure：Result comparison · GPT Image 2.5 Sunburst single official output

Case 22: Blister-Pack Collectible Toy Airplane (Design collectible merchandise)

• Objective: Use commercial product photography prompting (materials, packaging form factor, print clarity) to explore merchandise and packaging concepts while maintaining original, non-infringing designs.
• Generation Settings: size="1024x1536", quality="medium"
• Bilingual Prompt:

【中文】
创作一款复古风格玩具螺旋桨飞机的收藏级模型，带有圆润机翼、前置旋转螺旋桨、微带磨损的掉漆边缘以及经典的童年玩具比例，设计为吸塑包装内的怀旧节日收藏品。

概念：
一款充满怀旧情调的节日收藏品，灵感来自孩子们在冬日假期里玩耍的简朴玩具飞机。
唤起温暖、想象力与童年探索心。

风格：
高端玩具摄影质感，逼真的塑料与烤漆金属纹理，
影棚布光，浅景深，
清晰锐利的标签印刷，高规格零售陈列质感。

约束：
- 仅限原创设计
- 无商标
- 无水印
- 无 Logo

仅包含此包装文字（逐字一致）：
"Christmas Memories Edition"

【English】
Create a collectible action figure of a vintage-style toy propeller airplane with rounded wings, a front-mounted spinning propeller, slightly worn paint edges, classic childhood proportions, designed as a nostalgic holiday collectible, in blister packaging.

Concept:
A nostalgic holiday collectible inspired by the simple toy airplanes
children used to play with during winter holidays.
Evokes warmth, imagination, and childhood wonder.

Style:
Premium toy photography, realistic plastic and painted metal textures,
studio lighting, shallow depth of field,
sharp label printing, high-end retail presentation.

Constraints:
- Original design only
- No trademarks
- No watermarks
- No logos

Include ONLY this packaging text (verbatim):
"Christmas Memories Edition"

Key Takeaways: Explicitly dictates blister packaging (in blister packaging), weathered paint edges (slightly worn paint edges), studio lighting, and high-end retail presentation, while strictly defining packaging copy via Include ONLY this packaging text (verbatim):; verbatim verification remains necessary after generation.

Prompts are Starting Points; Delivered Results Require Verification: Official guidelines emphasize that prompts serve only as starting points for workflow exploration; output images must be audited against business requirements before production use:

1. Text and Diagram Accuracy: Is required text verbatim-accurate and legibly rendered? Are diagram labels and relational mappings correct?
2. Subject and Detail Integrity: Are subject features, product geometry, package labeling, and reference details properly preserved?
3. Strictly Controlled Edit Bounds: Did the model modify only specified localized areas without unintended drift across backgrounds and peripheral elements?
4. Genuine Alpha Channel: When transparency is requested, does the exported image contain a true alpha channel rather than a rendered solid faux-checkerboard backdrop?

Official examples | Inputs and outputs (single official samples for inspection, not stable model rankings):

GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Figure：Result comparison · GPT Image 2.5 Flare single official output	Figure：Result comparison · GPT Image 2.5 Sunburst single official output

Part 4: Developer Production Deployment and Complete Python SDK Implementation

1. Official Python SDK Baseline Example

The official end-to-end runnable code example remains anchored to gpt-image-2 as its baseline. Developers can reference it to validate core flows for generating 4 transparent logo variations and extracting transparent product assets from real photography:

Setup and Prerequisites:

• Install the official OpenAI Python SDK and configure the OPENAI_API_KEY environment variable.
• Download the official product photo in advance and save it to the local path input_images/shampoo.webp.
• Initiating live requests incurs standard API usage fees.
• This example represents official baseline reference code rather than an officially recommended invocation for GPT Image 2.5. When evaluating GPT Image 2.5 (such as Flare or Sunburst), select candidate models independently and review their currently supported request parameters and configuration specs carefully.

import base64
from pathlib import Path

from openai import OpenAI

client = OpenAI()


prompt = """
Create an original, non-infringing logo for a company called Field & Flour, a local bakery.
The logo should feel warm, simple, and timeless. Use clean, vector-like shapes, a strong silhouette, and balanced negative space.
Favor simplicity over detail so it reads clearly at small and large sizes. Flat design, minimal strokes, no gradients unless essential.
Fully transparent background. Deliver a single centered logo with generous padding, clean alpha edges, and no solid backdrop, scenery, checkerboard, or watermark.
"""

result = client.images.generate(
    model="gpt-image-2",
    prompt=prompt,
    size="1024x1536",
    quality="medium",
    background="transparent",
    output_format="png",
    n=4,  # Generate 4 versions of the logo
)

# Preserve the returned PNG bytes, including the alpha channel.
for index, item in enumerate(result.data, start=1):
    Path(f"logo-generation-{index}-gpt-image-2.png").write_bytes(
        base64.b64decode(item.b64_json)
    )

# Extract a product from a reference image.
prompt = """
Extract the product from the input image and isolate it on a fully transparent background.
Output: centered product, crisp silhouette, no halos/fringing.
Preserve product geometry and label legibility exactly.
Add only light polishing. Do not add a solid backdrop, checkerboard, scenery, or shadow.
Do not restyle the product; remove the background and preserve clean alpha transparency.
"""

result = client.images.edit(
    model="gpt-image-2",
    image=[
        Path("input_images/shampoo.webp"),
    ],
    prompt=prompt,
    size="1024x1536",
    quality="medium",
    background="transparent",
    output_format="png",
)

Path("extract-product-gpt-image-2.png").write_bytes(
    base64.b64decode(result.data[0].b64_json)
)

2. The Reality of Drift Prevention in Production and Engineering Mitigations

The official guide provides a frank engineering warning:
Repeated edits may still alter subtle features intended to be preserved. If your business scenario strictly demands "pixel-identical" consistency across localized regions, do not rely solely on prompts; adopt an engineering or frontend masking approach that composites verified AI-generated patches back onto the original source image.

Part 5: Practical Ready-to-Copy Toolkit (Takeaway Assets)

Asset 1: Modular Engineering Prompt Blueprint (Prompt Master Blueprint)

When drafting complex commercial or editing prompts, use the structure below as a modular starting template. A prompt template does not guarantee single-shot success; adapt it to specific requirements and verify results item-by-item in actual production:

【中文】
[任务目标与形态]
生成 [核心主体与场景] 的 [写实照片 / 矢量 Logo / 信息图 / UI 原型 / 幻灯片]。

[主体与动作]
[详细描述人物外观、服装、视线方向、肢体动作，或工业产品的外形与关键机械部件]。

[媒介风格与物理质感]
[指定材质、光影、色温、相机镜头感（如 35mm 镜头、柔和海滨日光、自然皮肤颗粒噪点）]。

[确切排版文本 (可选)]
仅渲染此文本（逐字一致，仅渲染一次）："[必须精确渲染的文字内容]"
排版要求：[字形要求，如粗体无衬线体、居中对齐、规整字间距]。

[严格约束与负向排除]
- 仅修改 [在编辑模式下填入修改项，生成模式下略去]
- 保留 [必须锁定的属性：五官、比例、背景、原有光影角度]
- 带有清晰 Alpha 边缘的全透明背景 (若需透明背景)
- 无多余文字、无无关 Logo、无水印、无重度磨皮修图、无假棋盘格。

【English】
[Goal & Format]
Create a [photorealistic photo / vector logo / infographic / UI mockup / slide] of [core subject and scene].

[Subject & Interaction]
[Describe the person's appearance, clothing, gaze direction, and body movement, or the industrial product's shape and key mechanical components].

[Style & Physical Textures]
[Specify materials, lighting, color temperature, and camera cues such as a 35mm lens, soft coastal daylight, and natural skin grain].

[Exact Typography (optional)]
Render ONLY this text (verbatim, exactly once): "[exact text to render]"
Typography: [type requirements such as bold sans-serif, centered, clean kerning].

[Strict Constraints]
- Change ONLY [the requested edit; omit in generation mode]
- Preserve [locked attributes: facial features, proportions, background, and original lighting direction]
- Fully transparent background with crisp alpha edges (when transparency is required)
- No extra text, no unrelated logos, no watermarks, no heavy retouching, no fake checkerboard.

Asset 2: Multi-Image Editing "Role Assignment" Quick Reference

Asset 3: Production Migration and Launch Acceptance Checklist

Before switching API traffic to a new model, verify item-by-item:

By mastering structured intent expression, precise localized editing constraints, and genuine alpha transparency extraction—while establishing evaluation baselines for quality, retries, and comprehensive cost per qualified image on representative use cases—teams can unlock the engineering value of image generation and editing models more stably and efficiently in production environments.