最近 GPT 全新 AI 生图模型 GPT-image-2 上线,一经发布就引爆全网。
所以问题来了——它到底有多强?和 Gemini、豆包、万象(千问的生图模型)比起来差距有多大?
今天我们就来一场**「公平对决」**:四家模型,两轮对决,同样的提示词,看看谁才是 AI 生图的"真·王者"。
❝
顺便声明:豆包用的是 Seedream 5.0 Lite,千问用的是万象 2.7,都是各家最新模型,不存在"故意拿旧模型来打"的情况。
❞
本次测评的提示词参考了 awesome-gpt-image-2-prompts 仓库中的高质量 Prompt(要是让我自己写,大概就是:"帮我生成一个复古报纸,中间是阿丽塔"……所以专业的还是得看专业的)。
第一轮:复古报纸风格(阿丽塔)
「提示词:」
Create the most realistic front page design of a vintage newspaper featuring Alita, the cyborg warrior from the movie "Alita: Battle Angel". The layout should be made in the style of a real printed newspaper with a cinematic black-and-white aesthetic.
The main photo should be prominently placed in the center, framed, like the image in the title of the article. The subject in the photo should remain unchanged and clearly distinguishable in natural light and slightly increased contrast in order to match the spectacular editorial style.
Create a bold, attention-grabbing headline at the top (create a unique title that matches the spirit of the photo - it can be romantic, mysterious, dramatic, or action-packed). Add a smaller subtitle under it, which will look like a real newspaper caption.
Add realistic newspaper elements:
- Columns of small text (in the style of lorem ipsum, but framed like real news)
- At the top is the fictitious name of the publication (for example, The Daily Prompts, AI Times or similar - think creatively)
- Date, issue number and location
- Decorative lines, dividers, and vintage typography
- Small additional articles or captions to the main image
- Optional stamps, doodles, or editorial notes to add personality.
Style:
- Black and white or slightly faded monochrome paper
- Fine paper texture, grain, and ink defects
- Small shadows and creases that mimic real printed paper
- The aesthetics of a clean but slightly worn vintage newspaper
- Mood: Give the design personality, expressiveness and plot, as if the plot is part of the main article.
- Aspect ratio: 4:5 or 1:1
- High-detail, ultra-realistic hybrid of editorial photography and print design.
GPT
人物还原度很高,排版也很规整。中文有部分乱码,但整体瑕不掩瑜。
Gemini
画面氛围不错,人物还原也还可以。但报纸中的中文乱码比较明显,比 GPT 差了一截——看来"中文文字生成"还是各家共同的难题。
豆包
如果看过《阿丽塔:战斗天使》的小伙伴,应该能看出来……这和主角阿丽塔基本没有关系。人物还原度完全不在线,文字同样是乱码。
万象(千问)
说实话,生成的东西不行就算了,关键生成图片还需要**「排队」**。不开通会员或者不用积分加速,等的时间非常长。就这个水平还收费,属实有点说不过去。
❝
💡 有个有意思的细节:第一轮提示词是英文的,所以生成的报纸文字也是英文。我后来让 GPT 和 Gemini 在其他条件不变的情况下把报纸文字改成中文——至于为什么豆包和万象没有再次生成,我想大家心里已经有判断了。
❞
「🏆 第一轮结果:GPT 胜出。」
第二轮:苹果发布会现场
「提示词:」
A hyper-realistic amateur smartphone photo captured from the audience seating at Apple Park, during the iPhone 20 product launch keynote. Tim Cook is standing on stage in front of a massive LED display showing the new iPhone, gesture towards the screen with his signature smile. The massive auditorium is filled with thousands of attendees, dim theater lighting with blue stage lights illuminating the stage. Large "20" logo visible on the screen behind him. Audience heads and shoulders visible in the foreground bokeh. Professional camera flash reflections on the stage surface. News media cameras in the background. Authentic event atmosphere, candid moment, slight motion blur on fast-moving stage lights, authentic Apple keynote event documentation --ar 16:9
GPT
人物还原没问题,但仔细看会发现手部还是有点奇怪——经典的 AI 生图"手部问题"依然存在。这一轮 GPT 表现中规中矩。
Gemini
不得不说,这轮 Gemini 真的太强了。还原度高不说,注意看前排观众举起手机录屏的画面——手机屏幕上居然还有内容!这个细节处理可以说是拉满了。
豆包
单说人物还原度就已经出局了,和 Tim Cook 没什么关系……环境氛围也差得比较远。
万象(千问)
生成速度太慢,没有积分用来加速,直接淘汰。
「🏆 第二轮结果:Gemini 胜出。」
第三轮:UI 设计图
「提示词:」
A sleek modern UI design system displayed on a MacBook Pro screen, glassmorphism cards with subtle blur and transparency, frosted glass effect on navigation sidebar, modern gradient buttons in purple and blue, clean typography, data dashboard with charts and graphs, minimalist SaaS interface, soft ambient lighting reflecting on the screen, shallow depth of field, photorealistic product photography style, ultra detailed --ar 16:9
GPT
Gemini
这一轮我真的纠结了——两个生成得都不错,UI 布局、配色、质感都在线,很难分出高下。大家觉得呢?
总结
两轮下来,GPT 和 Gemini 各赢一轮,综合实力明显领先——无论是画面质感还是细节处理,都远超预期。豆包和万象作为国产模型还有很大的进步空间,希望之后有机会赶上来。
如果觉得这篇文章对你有帮助,欢迎转发给身边的开发者朋友——毕竟,帮别人省钱也是一种美德。 📢 本文首发于公众号「小贺前端笔记」,专注前端+AI 实战。