WHAT NEW IN X10.30 FRI

AI collections from X · 22 collections · 60,637 posts

gpt-imageARTICLE

10,000-word, 67-image test of GPT Image 2.5 image generation

A long-form test compares GPT Image 2.5 with Doubao and other models across 67 images, concluding GPT's output is still the hardest to detect as AI-generated.

梦老湿 Dream AI
@jackzhang123vipArchived 2026.09.14Originally posted 2026.09.11

What this article says

这是一篇自称「史上最全」的 GPT Image 2.5 生图实测长文,全文约万字、配 67 张图。作者围绕 GPT Image 2.5 展开多维度测试,并与国内模型(文中点名豆包)做横向对比,核心主张是 GPT Image 2.5 生成的图像在真实感上已经强到「能欺骗一切」——即肉眼难以分辨真假,在细节、光影和材质还原上明显领先,因此被作者称为豆包「最严厉的父亲」。文章以大量样图作为证据,覆盖不同题材与难度,结论倾向于:在追求照片级真实感的场景下,GPT Image 2.5 目前是更可靠的选择。适合想了解该模型实际出图水平、以及它与国产模型差距的读者参考。

史上最全共万字67图测试GPT Image 2.5,AI生图能欺骗一切,豆包最严厉的父亲还得是GPT https://x.com/i/article/2098405180471681024

梦老湿 Dream AI @jackzhang123vipOriginally posted 2026.09.11
Open original on X
x.com/jackzhang123vip/status/2098414576031481953Archived 2026.09.14 19:33