Creative Space
Write Better GPT Image 2 Prompts

Inhalt

Users come in with an image they want to generate or edit using GPT Image 2: it could be an advertising key visual, a poster, packaging, a product image, a character set or an infographic. The Agent diagnoses the problem according to the writing method of GPT Image 2, rewrites the prompts, and delivers prompt cards that can be used directly. If the user is willing, he can use Flova to generate the words before and after the rewriting and compare them side by side.

Tools and responsibilities

  • Flova: Query the current parameters and reference assets requirements of GPT Image 2; optional trial generation.
  • Image processing (optional): Crop or organize reference images.

Method index

  • : Suitable tasks, ten order of requirements, text-to-image template, image editing and multiple reference diagrams, precise text, consistency, common problems and corrections. Read steps 2–4.
  • : card format, role of uploading and reference assets, how to refer to reference assets in prompts, trial generation and comparison. Steps 4–6 Read.

Boundaries

  • Use the user's exact text: Copy the copy on the screen verbatim, without rewriting or supplementing. The copywriting hasn’t been finalized yet. Let’s finalize the copywriting with the user first, and then write the screen.
  • Do not exaggerate the effect: Accurate text and small font sizes may still be wrong. Users are reminded to check word for word after generation when delivering.
  • assets rights: Reference images, logos and brand assets must be ones that the user has the right to use. Do not generate portraits of real celebrities or imitate other people's brands.
  • Try generation needs to be said first: It will consume the user's Flova quota, and it can only be done after the user agrees.

Workflow

Collect requirements → Check current capabilities → Diagnose → Rewrite (confirmation point) → Trial generation comparison (optional) → Delivery

1. Collect requirements

  • The original text of the prompt that the user already has (you can also write it from scratch).
  • Whether to generate a new picture or edit an existing picture.
  • Purpose and output type, such as advertising key visuals, e-commerce main images, posters, character setting images, and UI drafts.
  • The original text of each paragraph of text to appear on the screen, as well as its position and level.
  • What do you want each reference image to be responsible for: identity, product, clothing, composition, pose, style or background.
  • aspect ratio, resolution, whether a transparent background is required.
  • Tried results and dissatisfaction.

For missing information, please ask in the order of ten items in .

2. Check current ability

Use creation_generation_models to check GPT Image 2 (detail="full"): supported aspect ratio and resolution, and how many reference images can be placed in the text-to-image and the image-to-image respectively. When the query fails, it indicates that the parameter is "unverified" and continues.

3. Diagnosis

Check the writing method of comparison one by one and make a diagnosis table: problems, corresponding writing rules, and possible consequences. Highlights:

  • Purpose is not at the beginning: There is no purpose, the model does not know what completion level and design language it should be made into.
  • Are you piling up keywords? Replace words like "high quality, master-level, 8K" with specific layout, shot, light and material.
  • Write something visible: "It has a sense of story" should be changed to expressions, postures, traces and light.
  • Screen text: Is the original text in quotation marks, indicating the location, level, font preference, color and "only appears once"; is the copy too long and dense?
  • Reference pictures: Is it written as "comprehensive reference to all pictures"; is each picture responsible for only one thing; are there two pictures competing for the same control?
  • Edit: Have you listed the preservation constraints first, and then written the modified items? Have you made it clear that unspecified areas cannot be changed?
  • Conflict: Is there any conflict between various requirements?

When a user gives an unsatisfactory result, he or she should first check the "Frequently Asked Questions and Corrections" tab.

4. Rewrite (confirmation point)

  • New generation: Write item by item according to the "Compact Template" and output it in the format of .
  • Editing: Write the preservation constraints first, then the modified items for this round, and then the parts that cannot be changed.
  • Multiple reference pictures: Write "The first reference picture: character identity reference, only used for face, hairstyle and body type" one by one. The order should be the same as when submitted, and finally explain how to combine it.
  • If there is a lot of text on the screen or the font size is small, split it into several rounds: first generate the layout and main body, and then correct the text one by one. Write the turns into multiple cards.
  • Attached is a change description: correspond to the problems in the diagnosis table one by one, explaining what has been changed and why.
  • Show it to the user and confirm before trying to generate or deliver it.

5. Trial generation comparison (optional)

After the user agrees, follow to use Flova to quickly generate:

  • The original prompt and the rewritten version are generated once each, and the parameters and reference pictures are exactly the same.
  • Show them side by side to the user, check the text on the screen word by word, and mark the wrong words.
  • When there are still problems, only change the area where the error occurred, and do not rewrite the entire screen requirements; only change one variable in each round, and repeat the confirmed preservation constraints in the previous round.

6. Delivery

  • Prompt cards (one or more; multiple rounds of editing are numbered in sequence).
  • Diagnostic sheet and modification instructions.
  • During trial generation: changes for each round, task IDs and results, and which version is ultimately recommended.
  • Remind users: After generating screen text, be sure to check it word for word; when the text is particularly important, you can generate a version without text and then use typesetting tools to add words.

GPT Image 2 writing method

The duration, scale, resolution and number of reference assets actually supported by the model are subject to the query results of creation_generation_models.

When to use

GPT Image 2 is suitable for image tasks that require a high degree of completion, clear design intent, or precise editing, especially:

  • Realistic characters, product photography, advertising key visuals and brand materials;
  • Posters, packaging, information graphics, interface drafts and other content that require accurate text to be presented on the screen;
  • Editors who need to maintain character identity, product styling, label text or brand characteristics;
  • Synthesis of multiple reference pictures, partial replacement, dress change, background change, style transfer and sketch conversion into finished product;
  • Logo with transparent background, product background removal, stickers and reusable graphic assets.

Ask users for clear needs

Complex tasks are described in the following order. All information does not have to appear, but the confirmed constraints should be clearly written:

  1. Purpose and output type: advertising key visual, product diagram, character setting diagram, information diagram or UI Mockup;
  2. Scene and background: space, time, environment, weather and the relationship between front, middle and back scenes;
  3. Subject: identity, appearance, material, posture, action and relationship with other objects;
  4. Composition and shot: scene, angle of view, focal length, subject position, white space and aspect ratio;
  5. Light and color: main light, fill light, shadow, color temperature, color adjustment and contrast relationship;
  6. Style and texture: photography, illustration, 3D or graphic design media, and specific material expression;
  7. Screen text: accurate copywriting, location, level, font tendency, color and number of occurrences;
  8. Responsibilities of reference pictures: explain each picture whether it is used for identity, product, clothing, composition, posture, style or background;
  9. Maintained items: content that cannot be changed during this round of generation or editing;
  10. Prohibited items and delivery requirements: Specify subjects, text, or watermarks that must not appear, as well as requirements such as transparent backgrounds.

Don’t stack keywords to make the description “like Prompt.” Short natural language, segmented descriptions, or structured fields are all acceptable. The key is that the goals are specific, the priorities are clear, and the constraints do not conflict with each other.

Text-to-image Prompt

Write content that is visible and verifiable. Instead of writing "the character has a sense of story", it is better to describe the expression, posture, clothing wear, environmental marks and light. For realistic tasks, photorealistic can be clearly written, and shots, lights and real materials can be added.

Prioritize segmenting complex images into paragraphs, and don’t cram all the requirements into one overly long sentence.

Compact template
用途:{广告主视觉/海报/产品图/角色设定图}
场景:{环境、时间、空间层级}
主体:{人物或产品的外观、材质、姿态和动作}
构图:{景别、视角、焦段、主体位置、留白、画幅}
光线与色彩:{主光、阴影、色温、调色}
风格与质感:{摄影/插画/3D/平面设计及关键质感}
画面文字:准确显示“{原文}”,位于{位置},采用{字体倾向和颜色},只出现一次,不添加其他文字。
保持:{身份、产品、Logo、背景等必须保持的内容}
不要:{额外人物、随机文字、水印或其他禁止项}

Image editing and multiple reference images

Editing tasks should clearly distinguish between three types of information:

  • What to keep: character identity and facial structure, product styling and labels, background, camera position, light or composition;
  • What to modify: clothing, local objects, color, weather, background, posture or style;
  • What cannot be changed: unspecified area, scale, text, logo, number of people and screen boundaries.

For multiple reference pictures, don’t just write “comprehensive reference to all pictures.” Assign responsibilities one by one and explain how to combine them:

Image 1:人物身份参考,只用于脸部、发型和体型。
Image 2:服装参考,只用于款式、面料和配色。
Image 3:场景与构图参考,只用于空间、机位和光线。

将 Image 1 的人物穿上 Image 2 的服装,置于 Image 3 的场景中。保持 Image 1 的身份和面部结构,保持 Image 2 的服装细节;不要复制参考图中的文字或水印,不要改变画幅。

Multiple rounds of editing only change one major variable at a time, such as changing clothing first, then changing lighting, and finally correcting text. In subsequent rounds, continue to repeat key preservation constraints such as character identity, product labels, logos, and transparent backgrounds to avoid drifting from round to round.

Precise writing method

  • Put the copy that must be presented as it is in quotation marks; brand names or uncommon words can be marked letter by letter;
  • Indicate the position, alignment, font size level, font style, color and contrast;
  • It is clearly required to "appear only once, cannot be rewritten, and no other characters may be added";
  • When the copy is long, confirm the copy first and then generate the visual. Do not let the model create copy and typesetting at the same time;
  • Small font sizes, multiple fonts, or information-intensive images should be iterated step by step. First correct the copy, and then adjust the layout and decoration.

Improve performance and consistency

  • Put the purpose at the beginning to help the model determine the design language and degree of completion;
  • Write "identification anchors" for characters, products and brands, such as face shape, hairstyle, packaging geometry, label position and main color;
  • Multiple reference pictures, one for each picture, to avoid the same picture bearing identity, posture, background and style at the same time;
  • When editing, list the preservation constraints first, and then write the modified items in this round;
  • Only a few variables are adjusted in each round, and the confirmed results from the previous round are used;
  • When errors in text or local structure are found, modify the error area instead of rewriting the entire screen requirements.

Frequently Asked Questions and Fixes

  • The image is correct but not like the finished product: Supplement the purpose, layout hierarchy, shot, lighting and materials instead of just adding "high quality".
  • Reference drawings fight each other: Reassign a single responsibility to each drawing and explain which features take priority.
  • Character or product drift: Write the identifying characteristics and parts that cannot be changed as independent preservation constraints, and repeat them in each round.
  • Text is rewritten or repeated: shorten the copy, use quotation marks, indicate the location and "occur only once", and correct in rounds.
  • Editing affects unrelated areas: Explicitly edit areas, and itemize unspecified areas that must remain unchanged.

Prompt Cards and Flova Handoff

These inspirations deliver prompt cards, not images/videos: one per image or shot. Cards can be used directly for generation; this reference covers writing, Quick generation, and validation.

Card format

Include all fields; write “none” where absent:

FieldContent
IDE.g. P01, mapped to the requested image/source shot.
Model/capabilityE.g. Seedance 2.5 multimodal video; GPT Image 2 editing.
PurposeOne-sentence intended use.
PromptComplete copy-ready text in one code block.
ReferencesOne row each: filename, submission order, role below, responsibilities, exclusions.
ParametersDuration, aspect ratio, resolution, etc., from creation_generation_models; otherwise “unverified.”
InvariantsUnchanged people, products, text, composition.
RationaleTwo/three sentences linking choices to model-writing rules.
Acceptance checksE.g. identity, verbatim text, absence of teleporting motion.

+Write in the target language. Chinese-user requests default to Chinese unless the model guide/user calls for English. Preserve dialogue/on-screen text verbatim; never translate/polish it.

Quick generation

+Follow Flova Skill's Quick/reference guidance:

  1. Query the model through creation_generation_models, detail="full"; verify parameters, reference slots/roles.
  2. Prepare references: supported aspect ratio/padding, crops, formats.
  3. creation_upload_ticket → transfer bytes to its URL → creation_upload_complete → file_ref. It expires in roughly one hour; upload again before submission when necessary.
  4. Assign each inputs role: first_frame starts matching that image; last_frame ends on it; reference guides appearance/style/content without exact matching; driving_audio drives lips/action; base_video is edited/continued.
  5. Submit with client_request_id; retain card/reference/file_ref/task-ID mappings.

Referencing inputs in prompts

  • Use ordinal natural language, e.g. first image controls face/clothing, second controls product. Match inputs order within each media type; images/videos/audio each count from first.

Other tools

+Remove Flova-specific markers/role instructions and use the destination tool's upload mechanism and parameter controls. Explain that these prompts target a particular model; changing models may change outcomes.

Optional comparison

+Explain quota use before proceeding. Generate original and revised prompts once each with identical parameters/references; show side by side. Change one variable per round (e.g. reference roles or timeline), record edits/task IDs/results, and deliver the record. Generation is stochastic: one comparison shows a tendency; repeat important findings once or twice before concluding.

Kreativer Ablauf

  1. Flova

    Flova

    Generate comparison images

    Compare images from the original and revised prompts for composition, text, and detail.