这一篇解决什么问题

图像生成的失败，多数不是「模型不够强」，而是需求没有写清楚：主体是什么、用在哪里、哪些必须保留、哪些不许出现。这篇文章给出一套可复用的写法：先把结果定义成一幅可以被检查的图，再逐项写出约束，最后按清单验收。

核心思路只有一句：先说你需要的图，再描述主体、构图、风格与限制；编辑时说清改什么、保留什么；一次只改一件事，并且每次都检查结果。

先选模型

同一个系列里有两个选择：

两者都支持生成、编辑与透明背景。选择顺序应当是先确认质量，再优化延迟：如果 Sunburst 满足质量要求，再用同一组提示词、参考图和尺寸去测 Flare；只有质量仍可接受且延迟下降，才切换。

迁移时的对照：

一个容易忽略的点：同样的 quality 标签，在不同模型上不等于同样的画质和延迟。速度提升必须在你自己的工作负载上测，不能从别人的结果外推。

参数与提示词分开

API 参数不写进提示词，单独设置：

常见尺寸：1024x1024（方）、1536x1024（横）、1024x1536（竖）、2048x2048（2K 方）、2048x1152（2K 横）、3840x2160（4K 横）、2160x3840（4K 竖）。

自定义分辨率写成 宽x高，必须满足：

• 每条边不超过 3840 像素。
• 两边都是 16 的倍数。
• 长边与短边之比不超过 3:1。
• 总像素在 655,360 到 8,294,400 之间。
• 超过 3,686,400 像素（即 2560x1440）属于实验性范围。

质量参数怎么调：先按用途选模型，再调 quality。第一次比较时，把 quality、提示词、参考图、尺寸全部固定，只换模型。输出不达标就升一档；达标后再往下降，看能否在可接受的质量下降低延迟。xhigh 与 max 只在确有未满足的质量要求、且延迟预算允许时使用——更高的档位不保证每个提示词都更好。

需要透明素材时，必须显式写 background="transparent" 并使用 PNG 或 WebP；output_compression 只对 JPEG 或 WebP 有效，对 PNG 不要设置。

迁移已有工作流的六步

1. 保存基线。 收集有代表性的生产提示词与参考图，要包含难例：精确文字、人脸、产品几何、透明素材。记录当前模型、参数与结果。
2. 选第一个候选。 旧模型已达标就从 Flare 起测延迟；复杂用例达不到质量就从 Sunburst 起测质量。第一轮比较时提示词、参考图、尺寸、输出格式都保持不变。
3. 检查完整结果。 对照指令遵循、身份与产品保留、文字准确度、非预期改动、透明度五项。重复请求以测稳定性；编辑流程要把整条编辑链和单个步骤分别测。
4. 质量通过后再测延迟。 若从 Sunburst 起步且它已达标，再用同一套要求测 Flare；只有质量仍可接受且延迟下降才切换。
5. 一次只改一个设置。 先比 quality 再改提示词。记录典型与最慢响应、失败、重试与「每张被接受图片」的成本。价格要查当前报价，不要假设快模型一定更便宜。
6. 按工作流分批放量。 达到验收标准后先切一小部分流量，监控同样指标，再逐步扩大；旧模型保留可回滚。

一个必须接受的现实：多轮编辑仍可能改动你本想保留的细节。 每次都要重申约束并检查结果；如果某个区域必须像素级不变，就把通过验收的编辑合成回原图，而不是指望提示词。

写提示词的七条原则

1. 定义结果。 说出主体与用途（产品图、广告、示意图），指定构图、宽高比与重要的位置约束。复杂需求按「场景、主体、细节、约束」分段。
2. 选可维护的格式。 短句、段落、类 JSON 结构、指令、标签都能表达同一件事；选最容易读和改的那种，不要依赖特殊语法。
3. 描述可见细节。 写材质、光线、颜色、视觉媒介。需要照片感就明说 photorealistic 或 real photograph，并描述取景与质感。相机参数只是外观线索，不是物理模拟的保证。宽幅、电影感、低光、雨夜、霓虹这类场景，要写尺度、氛围与颜色，不能只堆情绪词。
4. 写清人物与动作。 说明取景范围、相对大小、视线与物体的交互。像「全身可见，包含脚」「低头看打开的书」「双手自然握住车把」这样的句子，能显著减少歧义。
5. 精确文字要加引号。 给出必须出现的文案、位置与字体风格。不常见的词或品牌名逐字母拼出。同时要求「不要额外文字」，并在输出后检查拼写与可读性。小字、密集信息、多字体时对比 medium 与 high。
6. 把改动与约束分开。 编辑时写「只改 X」，再列出要保留的细节：身份、几何、版式、光线、标签。还要写排除项，例如不要多余文字、logo、水印。做局部精修时，把饱和度、对比度、箭头、机位、周围物体也列为不变项。
7. 给参考图分配角色。 按编号说明每张输入图的作用：主体、风格、服装、背景；再说明它们如何组合、哪个元素要移到哪里。

迭代方式：把上一轮输出作为下一轮输入，一次只要求一个改动，并重复要保留的细节。像「和之前一样的风格」这种指代能携带上下文，但结果一旦漂移，就要把关键约束重新写全。

生成类场景

控制风格与光线

把照片拆成主体、取景、光线、质感四层，并显式排除不想要的处理。

Create a photorealistic candid photograph of an elderly sailor on a small fishing boat.
Weathered skin with visible pores and sun texture; faded traditional tattoos.
Shot like a 35mm film photograph, medium close-up at eye level, 50mm lens.
Soft coastal daylight, shallow depth of field, subtle film grain.
Honest and unposed. No glamorization, no heavy retouching.

用图解释流程

写清流程名称、受众、要传达的信息。信息图与示意图除了外观，还要核对标签与事实关系。

Create a detailed infographic of the flow inside an automatic coffee machine:
bean basket → grinding → scale → water tank → boiler.
Make the technical and visual flow understandable.

渲染精确文字

把文案放进引号，说明出现次数；只写与文案相关的指令，不要混入无关要求。

A campaign image for a youth streetwear brand called Thread.
Render the tagline "Yours to Create." exactly once, clearly and legibly.
No extra text, no watermarks, no unrelated logos.

设计可复用 logo

描述品牌与构成标志的形状，要求在不同尺寸下都清晰；用 n 一次生成多个方案。

Original, non-infringing logo for Field & Flour, a local bakery.
Warm, simple, timeless; clean vector-like shapes, strong silhouette, balanced negative space.
Flat design, minimal strokes, no gradients unless essential.
Fully transparent background, centered, generous padding, clean alpha edges,
no backdrop, no checkerboard, no watermark.

历史与真实场景

给出地点与日期来锚定时代，然后逐项检查服装、陈设与环境是否符合史实。

A realistic outdoor crowd scene in Bethel, New York on August 16, 1969.
Photorealistic, period-accurate clothing, staging, and environment.

故事转漫画

把叙事拆成一组清晰的视觉节拍，一格一个动作；描述要具体、以动作为中心。

Create a short vertical comic-style reel with 4 panels.
Panel 1: The owner leaves through the front door...
Panel 2: The door clicks shut. Silence breaks...
Panel 3: The house transformed...
Panel 4: The door opens...

界面预览

把产品当作已经存在来描述，聚焦布局、层级、间距与真实控件；避免概念图语言，否则结果会像草图而不是可用界面。

A realistic mobile app UI mockup for a local farmers market.
Header, vendor list with small photos and categories, a "Today's specials" section,
location and hours. White background, subtle natural accents, clear typography.
Place the mockup in an iPhone frame.

科学教育图

按教学设计的思路写：受众、学习目标、视觉形式、必需标签、科学约束。要求干净扁平的图标体系、一致的风格、清晰箭头与足够留白。标签密集时用 quality="high"。

A simple biology diagram titled "Cellular Respiration at a Glance" for high school students.
Show glycolysis, the Krebs cycle, and the electron transport chain.
Label glucose, pyruvate, ATP, NADH, FADH2, CO2, O2, H2O.
White background, simple icons, clear labels, readable text. Avoid tiny text and decoration.

幻灯片与图表

把提示词当成「成品规格」而不是插画请求：说清交付物（幻灯片、流程图、图表、页面）、画布、层级，直接给出真实文字与数据。演示类用横向尺寸，含小字、图例、坐标轴或脚注时用 quality="high"。

One pitch-deck slide titled "Market Opportunity", in the style of a real Series A deck.
TAM/SAM/SOM concentric circles with values $42B / $8.7B / $340M.
A bar chart of growth from 2021 to 2026; footnotes "AGI Research, 2024" and "Internal analysis".
Clean white background, Inter-like sans-serif, crisp minimal layout.
Avoid clip art, stock photography, gradients, shadows, decorative elements.

编辑类场景

编辑用 client.images.edit，需要精确局部时使用遮罩。

透明抠图的典型写法：

Extract the product from the input image and isolate it on a fully transparent background.
Centered product, crisp silhouette, no halos or fringing.
Preserve product geometry and label legibility exactly.
Add only light polishing. Do not add a solid backdrop, checkerboard, scenery, or shadow.

注意两点：画出来的棋盘格不是透明；后续编辑时要重复「保持透明背景」的要求。

多轮精修

从一张输出开始，检查后再作为下一轮输入。每次的后续请求都要窄，这样才能看出是哪一步起了作用。

典型链路：先做一张广告牌效果图（文案逐字引用、要求只出现一次），再只改天气（「变成飘雪的冬夜」），最后保持角色一致性——把角色的外观、比例、服装、气质写成可复用的参考，后续场景重复这些约束，并明确「不要重新设计角色」。

验收清单

出图后逐条检查：

• 必需文字是否准确、可读？图表的标签与关系是否正确？
• 身份、产品形状、标签与参考细节是否完整保留？
• 编辑是否只改了要求改的部分？
• 需要透明时，文件里是否有真正的 alpha 通道，而不是画出来的背景？
• 更换提示词或模型时，是否在有代表性的输入上比较了质量、延迟与成本？

理解检验

1. 为什么第一次比较两个模型时，要把 quality、提示词、参考图和尺寸全部固定？

参考答案：因为这些因素都会影响画质与延迟。只换模型而固定其余变量，才能把差异归因于模型本身。如果同时改了 quality，就无法判断结果变化是模型带来的还是参数带来的；而且同样的 quality 标签在不同模型上并不等价，混在一起比较没有意义。

2. 自定义分辨率有哪些硬性限制？举一个不合法的例子并说明原因。

参考答案：每条边不超过 3840 像素；两边都是 16 的倍数；长边与短边之比不超过 3:1；总像素在 655,360 到 8,294,400 之间。例如 1000x1000 不合法，因为 1000 不是 16 的倍数；4000x1000 既不满足边长上限，比例也正好卡在 3:1 边界且总像素超限。

3. 为什么说「画出来的棋盘格不是透明」？在 API 层要怎么做？

参考答案：棋盘格只是把「透明」画成了灰白方格的像素，文件本身仍是不透明的 RGB 图像，放到别的背景上会出现白框。真正的透明需要图像格式带 alpha 通道：请求时显式设置 background="transparent"，输出用 PNG 或 WebP，并且不要对 PNG 设置 output_compression。

4. 编辑时「只改 X」还不够，为什么还要列出保留项？

参考答案：因为模型对「不要改」的理解是模糊的。明确列出身份、几何、版式、光线、标签等不变项，等于给出了可检查的约束；否则模型可能为了完成 X 而顺带调整相邻区域。多轮编辑尤其如此——每轮都可能累积漂移，所以每次都要重申约束并在输出后检查。

5. 多轮精修时，为什么每次只要求一个改动？

参考答案：为了定位因果。一次只改一件事，结果变好或变坏才能归因到那个改动；同时改多处，失败时无法判断是哪一处引起的，也无法决定保留哪个方向。把上一轮输出作为下一轮输入，再配合重复的保留约束，可以逐步逼近目标。

6. 从旧模型迁移到新模型时，什么时候应该停在 Sunburst，不再切到 Flare？

参考答案：当 Flare 无法在可接受质量下满足需求时。流程是先确认 Sunburst 达到质量要求，再用同样的提示词、参考图和尺寸测 Flare；只有当 Flare 的质量仍可接受、且延迟确实下降时才切换。如果 Flare 在任何可接受的质量档位下都不达标，就保留 Sunburst——质量优先于延迟。

与前后内容的关系

这篇讲的是「生成与编辑单张图」的写法。把它用于批量素材时，真正的难点会转移到验证与复用：如何确认每张图都带 alpha 通道、如何在图表里核对数据、如何把同一套提示词稳定复现。这些内容在同系列的「透明素材」一篇里展开。

The problem this solves

Most image failures are not the model being weak; they are underspecified briefs: what the subject is, where the image will be used, what must be preserved, and what must not appear. The working method is to define the result as an image someone could inspect, state the constraints one by one, and check the output against a list.

The core instruction is one sentence: start with the image you need, then describe subject, composition, style, and constraints; for edits, say what changes and what must stay; change one thing at a time and inspect every result.

Choose the model first

There are two choices in the family:

Both support generation, editing, and transparent backgrounds. The order is establish quality first, then optimize latency: if Sunburst meets requirements, test Flare with the same prompts, references, and dimensions; switch only if quality remains acceptable and latency improves.

Migration guidance:

One subtlety: the same quality label does not imply the same image quality or latency across models. Speed gains must be measured on your own workload; they cannot be extrapolated from someone else's results.

Keep parameters out of the prompt

API parameters are set separately from the prompt text:

Common sizes: 1024x1024 (square), 1536x1024 (landscape), 1024x1536 (portrait), 2048x2048 (2K square), 2048x1152 (2K landscape), 3840x2160 (4K landscape), 2160x3840 (4K portrait).

Custom resolutions use WIDTHxHEIGHT and must satisfy:

• Each edge no more than 3,840 pixels.
• Both edges multiples of 16.
• Long edge to short edge no more than 3:1.
• Total pixels between 655,360 and 8,294,400.
• More than 3,686,400 pixels (2560x1440) is experimental.

How to tune quality: pick the model for the job first, then adjust quality. For the first comparison, hold quality, prompt, references, and dimensions fixed and change only the model. If output falls short, raise the setting; once it passes, test lower settings to reduce latency. Use xhigh or max only when they fix an unmet requirement within your latency budget — a higher setting does not guarantee a better result for every prompt.

For transparent assets, explicitly request background="transparent" and use PNG or WebP. output_compression applies only to JPEG or WebP; do not set it for PNG.

Migrating an existing workflow in six steps

1. Save a baseline. Collect representative production prompts and references, including difficult edits, exact text, faces, product geometry, and transparent assets. Record the current model, settings, and results.
2. Choose the first candidate. If the old model meets quality, start with Flare and test latency. If it falls short on a complex case, start with Sunburst and establish quality first. Keep prompt, references, dimensions, and output format unchanged for the first comparison.
3. Check the complete result. Compare instruction following, identity and product preservation, text accuracy, unwanted changes, and transparency. Repeat requests to measure consistency; for editing, test the full sequence as well as individual steps.
4. Test for latency only after quality passes. If you started on Sunburst and it meets requirements, evaluate Flare against the same requirements. Switch only if quality stays acceptable and latency improves.
5. Tune one setting at a time. Compare quality levels before rewriting the prompt. Record typical and slow responses, failures, retries, and cost per accepted image. Check current pricing rather than assuming the faster model is cheaper.
6. Roll out by workflow. Once the candidate passes acceptance, move a small share of traffic, monitor the same measures, and expand gradually. Keep the previous model available for rollback.

One reality to accept: repeated edits can still change details you meant to preserve. Restate the constraints and inspect each result. If a region must remain pixel-identical, composite the approved edit into the original instead of relying on prompting.

Seven prompting principles

1. Define the result. Name the subject and intended use (product photo, ad, diagram), and specify composition, aspect ratio, and placement constraints. For complex requests, organize the prompt as scene, subject, details, and constraints.
2. Choose a maintainable format. Short prompts, paragraphs, JSON-like structures, instructions, and tags can all express the same intent. Pick the one that is easiest to read and update; do not rely on special syntax.
3. Describe visible details. Name materials, lighting, colors, and medium. Ask for "photorealistic" or "real photograph" explicitly when that is the goal, and describe framing and texture. Camera specs are cues for appearance, not physical simulation. For wide, cinematic, low-light, rainy, or neon scenes, specify scale, atmosphere, and color instead of piling up mood words.
4. Specify people and actions. Describe framing, relative scale, gaze, and interaction with objects. Lines such as "full body visible, feet included", "looking down at the open book", or "hands naturally gripping the handlebars" remove ambiguity.
5. Quote exact text. Put required copy in quotes and describe position and typography. Spell unusual words or brand names letter by letter when needed. Ask for no extra text, and check spelling and legibility afterwards. Compare medium and high quality for small text, dense information, or multiple fonts.
6. Separate changes from constraints. For edits, say "change only X", then list what must be preserved: identity, geometry, layout, lighting, labels. State exclusions (unwanted text, logos, watermarks). For precise local edits, also freeze saturation, contrast, arrows, camera angle, and surrounding objects.
7. Assign roles to references. Identify each input by number and purpose: subject, style, clothing, background. Explain how they combine and which element moves where.

For iteration: feed the previous output in as the next input, request one change, and repeat the details to preserve. References like "same style as before" carry context, but if the result drifts, restate the critical constraints in full.

Generation scenarios

Control style and lighting

Split a photograph into subject, framing, light, and texture, and explicitly exclude processing you do not want.

Create a photorealistic candid photograph of an elderly sailor on a small fishing boat.
Weathered skin with visible pores and sun texture; faded traditional tattoos.
Shot like a 35mm film photograph, medium close-up at eye level, 50mm lens.
Soft coastal daylight, shallow depth of field, subtle film grain.
Honest and unposed. No glamorization, no heavy retouching.

Explain a process visually

State the process, the audience, and the information to convey. For diagrams and infographics, verify labels and factual relationships, not just appearance.

Create a detailed infographic of the flow inside an automatic coffee machine:
bean basket → grinding → scale → water tank → boiler.
Make the technical and visual flow understandable.

Render exact text

Put the copy in quotes and state how many times it should appear. Keep the prompt focused on the copy and the visual treatment.

A campaign image for a youth streetwear brand called Thread.
Render the tagline "Yours to Create." exactly once, clearly and legibly.
No extra text, no watermarks, no unrelated logos.

Design a reusable logo

Describe the brand and the shapes that define the mark; require legibility at any size; use n to request variations.

Original, non-infringing logo for Field & Flour, a local bakery.
Warm, simple, timeless; clean vector-like shapes, strong silhouette, balanced negative space.
Flat design, minimal strokes, no gradients unless essential.
Fully transparent background, centered, generous padding, clean alpha edges,
no backdrop, no checkerboard, no watermark.

Historical and real-world settings

Name the place and date to anchor the period, then inspect clothing, staging, and surroundings for accuracy.

A realistic outdoor crowd scene in Bethel, New York on August 16, 1969.
Photorealistic, period-accurate clothing, staging, and environment.

Story to comic strip

Break the narrative into clear visual beats, one per panel, with concrete action-focused descriptions.

Create a short vertical comic-style reel with 4 panels.
Panel 1: The owner leaves through the front door...
Panel 2: The door clicks shut. Silence breaks...
Panel 3: The house transformed...
Panel 4: The door opens...

Interface preview

Describe the product as if it already exists. Focus on layout, hierarchy, spacing, and real controls; avoid concept-art language or the result will look like a sketch rather than a shipped interface.

A realistic mobile app UI mockup for a local farmers market.
Header, vendor list with small photos and categories, a "Today's specials" section,
location and hours. White background, subtle natural accents, clear typography.
Place the mockup in an iPhone frame.

Scientific and educational visuals

Prompt it like an instructional design brief: audience, learning objective, visual format, required labels, scientific constraints. Ask for a clean, flat icon system with consistent style, clear arrows, and enough whitespace. Use quality="high" for dense labels.

A simple biology diagram titled "Cellular Respiration at a Glance" for high school students.
Show glycolysis, the Krebs cycle, and the electron transport chain.
Label glucose, pyruvate, ATP, NADH, FADH2, CO2, O2, H2O.
White background, simple icons, clear labels, readable text. Avoid tiny text and decoration.

Slides and charts

Write the prompt as an artifact spec, not an illustration request: name the deliverable (slide, workflow diagram, chart, page), the canvas, and the hierarchy, and give the real text and data. Use landscape sizes for deck-style output and quality="high" when there is small text, legends, axes, or footnotes.

One pitch-deck slide titled "Market Opportunity", in the style of a real Series A deck.
TAM/SAM/SOM concentric circles with values $42B / $8.7B / $340M.
A bar chart of growth from 2021 to 2026; footnotes "AGI Research, 2024" and "Internal analysis".
Clean white background, Inter-like sans-serif, crisp minimal layout.
Avoid clip art, stock photography, gradients, shadows, decorative elements.

Editing scenarios

Use client.images.edit for edits, and a mask when the local region must be exact.

A typical transparent cutout prompt:

Extract the product from the input image and isolate it on a fully transparent background.
Centered product, crisp silhouette, no halos or fringing.
Preserve product geometry and label legibility exactly.
Add only light polishing. Do not add a solid backdrop, checkerboard, scenery, or shadow.

Two reminders: a drawn checkerboard is not transparency, and subsequent edits must repeat the requirement to preserve the transparent background.

Refining across turns

Start from one output, inspect it, then pass it in as the next input. Keep each follow-up narrow so you can tell which change helped.

A typical chain: produce a billboard mockup (quote the copy verbatim and require it to appear once), change only the weather ("make it a snowy winter evening"), then keep a character consistent by writing the appearance, proportions, outfit, and personality as reusable constraints and repeating them in each new scene with "do not redesign the character".

Acceptance checklist

Check each output before using it:

• Is required text accurate and legible? Are diagram labels and relationships correct?
• Are identities, product shapes, labels, and reference details intact?
• Did the edit change only what was requested?
• If transparency is required, does the file carry a real alpha channel rather than a painted background?
• When changing prompts or models, were quality, latency, and cost compared on representative inputs?

Comprehension check

1. Why hold quality, prompt, references, and dimensions fixed during the first model comparison?

Reference answer: All of them affect image quality and latency. Changing only the model lets you attribute differences to the model itself. If quality changes at the same time, you cannot tell whether the result came from the model or the setting; and because the same quality label is not equivalent across models, mixing them makes the comparison meaningless.

2. What are the hard constraints on a custom resolution? Give an invalid example and explain why.

Reference answer: Each edge at most 3,840 pixels; both edges multiples of 16; long-to-short edge ratio at most 3:1; total pixels between 655,360 and 8,294,400. 1000x1000 is invalid because 1000 is not a multiple of 16; 4000x1000 violates the edge limit, sits exactly on the 3:1 boundary, and exceeds the pixel budget.

3. Why is a drawn checkerboard not transparency? What must be done at the API level?

Reference answer: A checkerboard is just pixels painted to look transparent; the file itself is an opaque RGB image and will show a white box on another background. Real transparency requires an alpha channel: set background="transparent", output PNG or WebP, and do not apply output_compression to PNG.

4. Why is "change only X" insufficient without a list of things to preserve?

Reference answer: Because "don't change" is vague to the model. Explicitly listing identity, geometry, layout, lighting, and labels gives checkable constraints; otherwise the model may adjust adjacent regions while completing X. This matters even more across multiple edits, where drift accumulates, so constraints must be restated and results inspected every round.

5. Why request only one change per refinement round?

Reference answer: To establish causality. With a single change, an improvement or regression can be attributed to it; with several at once, a failure gives no way to tell which change caused it or which direction to keep. Feeding the previous output in and repeating the preservation constraints moves the result toward the target step by step.

6. When migrating, when should you stay on Sunburst instead of switching to Flare?

Reference answer: When Flare cannot meet the requirements at an acceptable quality. The procedure is to confirm Sunburst meets quality, then test Flare with the same prompts, references, and dimensions; switch only if Flare's quality stays acceptable and latency actually drops. If Flare fails at every acceptable quality level, keep Sunburst — quality outranks latency.

How this connects

This note covers producing and editing a single image. Applying it to batches moves the hard part to verification and reuse: confirming every asset has an alpha channel, checking chart data against the source, and reproducing the same prompts reliably. Those are covered in the transparent assets chapter.