56 条提示词含花括号模板变量，18 条含 [BRAND NAME] 类占位符，不替换会被模型画进图里；提示词是自然语言而非模型专用语法，作者并未在 GPT Image 2.5 上实测；GPT Image 2.5 尚未全量开放，需有可用账号才能运行
awesome-gpt-image-2-5-prompts
youart-open-source活跃维护收录 150 条 GPT Image 2.5 社区提示词，覆盖 9 类使用场景，每条都标注原作者和出处，可直接拿来生成图片。
GPT Image 2.5 prompts, verbatim and credited: 150 community prompts across 9 use cases, each with the author who wrote it and their post. Free to run.
暂无版本573star42fork同步于 2026-11-07 07:30 UTC
无需安装，直接打开 prompts/gpt-image-2-5-<use-case>-prompts.md 复制提示词；也可从 data/prompts.json 读取全部 150 条 硬件要求按 README 说明
README 未说明
150 条带作者署名的 GPT Image 提示词合集，按用途分类，可直接复制使用
收录 150 条 GPT Image 2.5 社区提示词，覆盖 9 类使用场景，每条都标注原作者和出处，可直接拿来生成图片。
核心能力
- 收录 150 条提示词，覆盖 9 类用途，每条保留原作者与出处链接
- 提示词原文照录，不翻译不改写，保证模型渲染引号内文字准确
- 按成品类型组织：海报、信息图、UI 稿、产品图、角色设定等
- 25 条以 JSON 结构书写，长提示词按规格逐项描述布局
- 附六条写作模式：先定成品形态、锁布局、引号写死文字、指定镜头
容易踩的坑
内容由模型根据 README 生成，可能不完整或不准确，以官方文档为准。看原文 README
暂无 release 数据（可能没有发版，或尚未同步）。
用户点评
登录用户的真实使用记录，可按版本和硬件筛选共 0 条
还没有人点评。第一个说出真实体验的人最有价值。
Awesome GPT Image 2.5 Prompts
150 production prompts for OpenAI's GPT Image models, including GPT Image 2.5. Every one is reproduced verbatim from the person who published it, with their account and a link to the post it came from.
150 prompts · 110 creators credited · 9 use cases · 25 written as JSON · updated 2026-09-10
GPT, GPT Image and ChatGPT are trademarks of OpenAI. YouArt is an independent platform and is not affiliated with or endorsed by OpenAI. GPT Image 2.5 is still rolling out and is not yet open to every account; when it opens, these prompts move across unchanged.
What this is
A curated, credited library of prompts that people published after running them on OpenAI's GPT Image models. They are grouped by what you are trying to make, not by style, because that is how the demand is shaped. Every prompt keeps its original wording, its author, and a link to the post it came from.
Where these came from, plainly: they were published on X by the 110 people credited here, and collected by two open-source repositories we drew from. 144 of the 149 datable posts predate GPT Image 2.5's release on 2026-09-08. These are natural-language prompts, not model-specific syntax, so a prompt written for GPT Image 2 runs unchanged on 2.5 — we would rather say that than restyle a GPT Image 2-era library as a 2.5-native one, and we do not claim that anyone here ran these on 2.5.
Every number in this file — prompts, creators, use cases, characters — is computed from data/prompts.json when the file is built. Not one of them is typed by hand, so none can drift from the corpus: change the data and the sentence changes with it.
How to use it
- Pick a use case below and open its file.
- Copy the whole prompt out of the code block. Do not paraphrase it — the quoted strings are there so the model renders them literally.
- Change the parts that should be yours: the subject, the copy, the palette. Then run it.
56 of these prompts still carry the template tokens their author wrote in curly braces. Replace each one with your own value before you run the prompt, or the model will render the placeholder as text.
A further 18 prompts use the other placeholder convention their authors wrote in — bracketed capitals such as [BRAND NAME] — and those need the same treatment. Swap in your own value, or the model will draw the placeholder into the picture, spelled correctly.
The bodies live in prompts/gpt-image-2-5-<use-case>-prompts.md, one file per use case. All 150 are in data/prompts.json as well, if you would rather work from the data than from the page.
GPT Image 2.5 image examples
Every example below was generated on YouArt with GPT Image 2.5. The prompts and editable templates are available on the GPT Image 2.5 prompts page.
The example images are reproduced from YouArt for demonstration. Image rights remain with their respective creators; this repository grants no additional licence for them. The third-party prompts retain their separate authorship and terms as well.
Browse by use case
The prompts themselves are not translated. They stay in the words their author wrote, because a prompt is input to a model rather than prose — this model's headline skill is rendering the exact quoted string into the picture, and translating it changes the output.
- Posters — Movie teasers, event flyers, travel posters and album covers. These prompts spend most of their length on layout, because a poster is a composition problem before it is a style problem: where the title sits, how much negative space is above it, what occupies the lower third. Open the 17 prompts
- Text in image — Legible words rendered inside the picture is the capability this generation of the model is known for. These prompts quote the exact string to render, name the typeface character, and place the text in the frame instead of hoping for it. Open the 17 prompts
- Infographics — Charts, timelines, exploded views and explainer slides. The long ones read almost like a spec: how many callouts, on which side, with what label text. That is the reason they work, and the reason many of them are written as JSON. Open the 17 prompts
- UI mockups — Dashboards, mobile screens, landing pages and stream overlays. A UI prompt has to name its components, because the model will invent a plausible interface if you do not tell it which one you meant. Open the 17 prompts
- Product shots — Studio product shots, packaging renders and catalog imagery. These are the prompts that pay for themselves: they describe a lighting setup, a surface and a camera position precisely enough to be repeatable across a whole product line. Open the 17 prompts
- Ad creative — Campaign key visuals, banners and social ad creative. Most of these carry both an image direction and a piece of copy to render, which is why they belong to this model rather than to a generation that could not spell. Open the 17 prompts
- Character design — Turnaround sheets, expression grids, equipment breakdowns and mascots. The useful pattern here is asking for a sheet rather than a picture: several views of one character in a single image, which is how you get a design you can keep using. Open the 16 prompts
- Portraits — Editorial portraits, headshots and film-look photography. These live or die on camera direction: the format, the focal length, the light source and the grade do more work than any adjective about mood. Open the 16 prompts
- Illustration — Anime, watercolour, line art, 3D render, isometric and pixel art. Style prompts are the shortest in the library, and the ones most worth editing: swap the subject and keep the style clause, and you have a series. Open the 16 prompts
How to write a GPT Image 2.5 prompt
Six patterns that show up again and again across the library. They are not rules about wording; they are decisions the model will make for you if you do not make them yourself, and the difference between a prompt that works once and one you can reuse.
Name the artefact, not the subject
Start with what the finished thing IS. "A 16:9 movie teaser poster" and "a cinematic scene of a woman in a chamber" produce different images from the same description, because the first one commits to a format, a purpose and a set of conventions. Almost every prompt in this library opens this way.
Lock the layout before the style
Say where things sit: which third holds the subject, what fills the header, how many panels, what is in the lower band. Style adjectives are cheap to change afterwards; a composition you did not specify is the thing you will regenerate five times trying to fix.
Put the exact words in quotation marks
If a string has to appear in the image, write it in quotes and say where it goes. Describing it ("a bold title at the bottom") gets you a bold title with the wrong words. This model can spell, so the failure mode has moved from garbled glyphs to text you did not ask for.
Direct the camera, not just the scene
Format, angle, focal length, light source, and the grade. "35mm film, harsh frontal flash, low-angle three-quarter view" is a specification. "Cinematic, moody" is a hope. This is the single biggest difference between the portrait prompts that reproduce and the ones that do not.
Constrain the palette
Three or four named colours beat any mood word. It is also what makes a prompt reusable as a template: a travel poster prompt that says "coastal cities use aqua, coral and cream; mountain cities use alpine blue and snow white" becomes a series instead of one image.
Say what must not appear
A short exclusion list at the end does real work: no watermark, no extra text, no visible sun in the corner, no plastic skin. Keep it short and specific. A long list of negatives starts competing with the description for the model's attention.
For an edit, name the change before the constraints
Describe the one thing that should change, then list what must survive it — subject, pose, camera angle, lighting, background, style. Reversing that order buries the instruction, and asking for three changes at once leaves you unable to tell which one moved the result. Match the new element to the original perspective, shadows and texture, or the edit reads as pasted on.
Say what a reference image is for
This model takes up to sixteen reference images, and "match this" is not an instruction. Name what the reference supplies — the subject, the silhouette, the colour palette, the layout, the material — and what should carry through to the new image. Then describe the setting or the change you actually want, which the reference does not contain.
Vary one thing, not everything
Hold the product, character or message constant and change one variable: two lighting directions, two compositions, two styles. Comparing a set that differs in one respect tells you which decision did the work; comparing four unrelated images tells you nothing. Keep the prompts that reproduce — those become the template for the next campaign.
What the library measures
Those patterns are not a house style; they are what this corpus does. The median prompt here is 2,587 characters — the shortest 272, the longest 4,648, 367,657 characters of prompt text in all. That median is the most useful number on this page: a prompt that reproduces for this model is closer to a short brief than to a sentence, and most people arrive expecting the opposite.
The long ones are not longer sentences. The longest quarter of the library — 38 prompts, from 2,973 characters up — runs to a median of 11 lines each, against 1 in the shortest quarter, and 88 of the 150 are broken across lines at all. They get long by adding parts rather than clauses: roughly one line per decision.
What that length is spent on is measurable, and it is not adjectives. 73 of the 150 name an explicit aspect ratio or page size. 74 of the 125 written as prose quote at least one literal string for the model to render. 57 name a lens, an aperture, a camera angle or a depth-of-field instruction, and they cluster where a result has to be repeatable rather than merely good — heaviest in the Portraits file, at 12 of 16. 73 end on an explicit exclusion clause. And colour is named in words rather than values: 0 of the 150 give a hex code.
Licence and credit
The build scripts are MIT — fork them, ship them commercially, credit us and our part is done. Everything else we assembled is not offered under an open licence yet; ask if you want to reuse a piece of it.
The prompts and their titles are their authors' work, not ours. Each row credits the person who wrote it and links the post it came from, and we did not change the text. We have not traced the chain of rights behind each one, so if you want to build on a prompt for something that matters, that credit line tells you whose it is.
Wrote one of these and want it gone? Ask, and it goes — no explanation needed.
LICENSE · ATTRIBUTION.md · TAKEDOWN.md
The JSON prompt format
Some of the prompts in this library are a JSON object rather than a sentence. That is not a special API mode or a hidden feature: the model reads it as text like anything else. What the structure buys you is precision, and an easier edit.
Reach for it when the image has parts that need to stay separate: a header, a centrepiece, a numbered set of callouts, a footer. Prose blurs those together and the model has to guess the hierarchy. Keys keep them apart, and they make the prompt a template, because changing one value does not risk rewriting the sentence around it. For a portrait or a single-subject illustration, prose is shorter and works just as well.
25 of the 150 prompts here are written that way, and JSON is the longer format when it is used: a median of 3,360 characters against 2,487 for the 125 written as prose. The bias grows with size — 17 of the 38 longest prompts in the library are JSON objects. Paste the whole object, braces and all; the model reads it as text, and a fragment of one reads as a fragment.
Every prompt
All 150. The Prompt link opens the full text in this repository; Run opens it on YouArt, where you can generate from it.
GPT Image 2.5 poster and key-art prompts
GPT Image 2.5 typography and in-image text prompts
GPT Image 2.5 infographic and diagram prompts
GPT Image 2.5 app and web UI mockup prompts
GPT Image 2.5 product photography and packaging prompts
GPT Image 2.5 advertising and campaign prompts
stable-diffusion-webui
AUTOMATIC1111·Python
基于 Gradio 的 Stable Diffusion 网页界面，把文生图、图生图、局部重绘等能力做成可视化操作面板，不用写代码就能跑图像生成。
图像生成+4图像生成进阶
165290+4431kv1.10.12026-03-02 07:00 UTC
ComfyUI
Comfy-Org·Python
基于节点图的扩散模型图形界面与后端，支持本地快速推理和 API 调用，方便搭建和调试图像生成工作流。
扩散模型+4图像生成进阶
146237+75117kv0.45.02026-11-11 05:07 UTC
OpenMontage
OpenMontageInc·Python
把 Claude、Cursor 等 AI 编程助手变成完整视频制作工作室，内置 12 条制作流水线和 100+ 工具，覆盖脚本、配音、图像生成到剪辑全流程。
视频制作+4图像生成硬核
79488+83411k2026-11-09 06:23 UTC
OpenMontage
calesthio·Python
把 Claude、Cursor 等 AI 编程助手变成完整的视频制作工作室，内置 12 条制作流水线和 100+ 工具，覆盖从生成到剪辑的全流程。
视频制作+4图像生成硬核
7786710k2026-11-03 17:42 UTC