离了大谱的GPT-Image-2模型,真的太强了(附案例)

昨天凌晨,OpenAI正式发布GPT-Image-2,生图模型新王正式上位。

image

T2I Arena 中GPT-Image-2断崖领先:

 

image

 

image

 

当然,不仅是Benchmark,实战上也是断崖式领先,之前一眼惊艳的Nano Banana Pro/2在GPT-Image-2面前就是个弟弟,对华人用户尤其友好的是:GPT-Image-2对于中文文字的处理也是独一档的存在!

以下附上实战案例以及一些心得:

不要过于信任GPT-Image-2默认知识能力,最好在生图前先让ChatGPT把knowledge先整理好,然后再生图。

{主题} = 红楼梦金陵十二钗
{风格} = 工笔画
{Task} = 为{主题}绘制一幅横轴长卷(萱草纸质混合羊皮卷轴),需要包含{主题}中所有人物形象、名字、判词、生平,归宿
{图片格式要求} = 比例9:21, 4K分辨率;
{HERO TITLE} = 在横轴中部上方,用汉字软笔书法行书撰写的 {主题} 

下面这张图也一眼惊艳,但仔细一看:很多判词都有文字瑕疵(点击放大):

image

这种情况,你应该先让ChatGPT先把判词都整理出来,然后再生图:

image

然后,就靠谱多了(点击放大):

image

只需要很少的提示词,就能出质量非常高的图

image
image

经验条、技能槽、聊天区、状态栏、BOSS血条,以及那个“未来已来,你来不来?”的霓虹灯牌…笑出鼻涕泡

但如果你的Prompt是英文,GPT-Image-2出来的就是英文版游戏了:

BTW,这辆特斯拉不错

image

直接丢个网页就能生图

image
image

一句话转小红书风格

image
image

时事新闻一键出图:

image

人物形象还原度很高!包括这个苹果新CEO特努斯…

image

GPT-Image-2非常“程序化思维”

Role: World-class Scientific Encyclopedia Illustrator & Knowledge Graph Architect.

Task:Generate a highly detailed, extremely intricate, and visually stunning "Universal Illustrated Encyclopedia Science Infographic" in a classic, unbranded (NO logos) scientific encyclopedia style.
Subject Matter:[People].
Specific Subject:[Einstein].
Style:Fine, detailed scientific illustration on a retro, aged beige paper background. Delicate linework. Intricately complex and professional.

Key Visual Requirements:
1.  Lifelike 3D Effect (The Central Subject):The central subject in the "C position" must be rendered with extraordinary realism and dynamism. Create a dramatic sense of depth where the character, plant, or animal appears to break the frame, leaping or bursting out of the flat paper towards the viewer (an effect similar to anamorphic 3D or dynamic pop-out, with high-precision realism).
2.  Layout & Strategic White Space:
* Central Subject:Dominates the center, with intentional "strategic white space" around it to enhance the popping-out effect and make the figure a clear focal point.
* Surrounding Modules:The surrounding area (left, right, top, bottom, and corners) must be filled with 6-8 distinct, highly organized knowledge modules, depending on the subject. There should be a sense of organized density, not random clutter. The modules themselves must have clear borders, headers, and extensive, detailed content.
3.  Connections:Use a complex, logical network of fine leader lines, arrows, brackets, dotted lines, and small connection points to link the central figure to all surrounding modules, and interconnect the modules themselves into a cohesive knowledge web.
4.  Text & Annotation (Hard Requirement - Must be CLEAR Chinese):
* Main Title:A large, prominent, beautifully executed Chinese calligraphy (书法体) of the specific subject's name [e.g., "大王乌贼"].
* Calligraphic Accents:Scattered throughout the main content and module titles, use beautiful, clear Chinese calligraphy for important terms.
* Standard Chinese Text:All other descriptive text, handwritten notes (大量清晰中文手写注释), module content, and annotations must be clear, legible Chinese characters (简体中文), not gibberish or unreadable symbols. Ensure text clarity is prioritized.
* Leader Line Annotations:Every single small component, detail, submodule, diagram, or illustration within the modules must have detailed leader line annotations (拟解剖图) pointing directly to it for maximum professionalism and educational value. Every part should be labeled.

Subject-Specific Module Structure (Example for general reference):
A. For Humans [People]:
-Module 1:Anatomy & Skeletal Structure (w/ magnified cross-sections)
-Module 2:Physiological Processes (e.g., Circulatory/Nervous System)
-Module 3:Historical Context & Timeline (Key Achievements)
-Module 4:Major Contribution Diagram (Detailed breakdown)
-Module 5:Cognitive Process / Psychological Insight
-Module 6:Genetic Profile / Evolution
-Module 7:Global Influence & Cultural Impact
-Module 8:Cultural Representations / Legacy

B. For Animals:
-Module 1:Full External Sketch & Anatomy (w/ microscope magnified detail circular windows)
-Module 2:Behavioral Patterns & Lifecycle (e.g., Mating/Migration, Flowchart style)
-Module 3:Digestive & Skeletal System
-Module 4:Habitats & Distribution Map (with environmental details)
-Module 5:Unique Adaptations (e.g., camouflage, hunting tools)
-Module 6:Evolutionary History & Relatives
-Module 7:Symbiotic Relationships / Ecosystem Role
-Module 8:Conservation Status & Human Interaction

C. For Plants:
-Module 1:Full Plant Sketch & Anatomy (w/ magnified leaf/root details)
-Module 2:Photosynthesis & Lifecycle Flow (w/ icons for environment)
-Module 3:Cellular Structure (Magnified circular views)
-Module 4:Medicinal Properties / Practical Applications (as in original original prompt)
-Module 5:Environmental Adaptations / Unique Features
-Module 6:Distribution Map & Environmental Context
-Module 7:Genetic Variations & Cultivation
-Module 8:Historical Usage & Folklore

Overall Composition:Extremely dense with information, organized into 6-8 structured modules, but balanced with strategic empty space around the center to allow the main, hyper-realistic figure to pop. Hard-core, professional, academic, but visually engaging due to the dynamic 3D central figure. No branding from any specific encyclopedia (e.g., no "DK" logos). All annotations must be legible. All handwritten notes must be clear. Main titles in Chinese calligraphy. Aspect Ratio:3:4.

我从X上找到一个很牛x的prompt模板(应该是X上的 @GeekCatX 原创),可以看到这里面其实是有People、Animals、Plants 三种选项的,但每次在Prompt中选择一种生成,GPT-Image-2对Prompt的follow也非常好:

image

桌面版Codex也可以使用GPT-Image-2了

image
image

电商主图

image

游科不给出四妹的DLC,就只能自己造了:

请生成《黑神话·四妹》的PC版动作游戏的进入游戏前的界面,采用《黑神话·悟空》的视觉风格。

主屏人物为《黑神话·悟空》中的”四妹“,参考图1。

界面风格参考图2。

image

再制作一个”四妹“在战斗中的游戏界面(可以参考附图)

image
image

给袁崇焕和崇祯皇帝搞个朋友圈

1.
请生成一个崇祯皇帝的朋友圈,比例为9:21,时间为李自成攻破北京前夜
2.
再生成一个袁崇焕的朋友圈,时间为宁锦大捷
image
image

绘画风格Gallery

请绘制一张16:9的Gallery view用于展示50种不同的漫画绘画风格

image

《山海经》

生成5个《山海经》中神兽的图片,要求:

1.
极繁主义的工笔画;
2.
宽高比9:16;
3.
分辨率4K高清;
4.
图上标记有神兽的名字和相关描述;
image
image
image
image
image

此文发布时,OpenRouter也刚刚上架了GPT-Image-2的API,你也可以在应用中随时调用GPT-Image-2来生图了:
离了大谱的GPT-Image-2模型,真的太强了(附案例)

实战分享

【万字剖析】抛开 RAG 谈蒸馏.skill,大概率是形式主义

2026-4-22 8:28:38

AI情报

来了!Google最新模型Gemma 4手机端使用教程

2026-4-4 9:49:04

0 条回复 A文章作者 M管理员
    暂无讨论,说说你的看法吧