ai21.net
语言

Sunburst 多轮精确编辑压测:换衫连环到第几轮崩

原图→改 T 恤→再拿输出当下轮输入,连压五轮:别的模型第三轮已噩梦,Sunburst 第五轮还能看。Flare 赶量 / Sunburst 精改——编辑保真硬证据。

原帖未提供效果图

提示词

How many times can you edit an image with AI before it becomes unusable?

GPT-Image 2.5 Sunburst was released, and one of its biggest improvements was being able to make precise edits to the same image over and over without changing anything else.

So I decided to stress test it.

I took a picture for the input, prompted to change the t-shirt, and then used the output as the input for the next prompt.

This video doesn't necessarily make the model look great.

But if you do this with other models, it will barely even make it to round 3 before looking like something out of a nightmare.

Here, even at round 5, the generation isn't too bad.

When editing an image with AI, it's rare that you would go over two rounds anyway before restarting entirely with the OG image. 

GPT-Image 2.5 is quickly becoming my favorite AI image model, and I absolutely love the speed of Flare and the precise editing of Sunburst.

中文译文

你最多能用 AI 编辑同一张图片多少次,它才会变得不可用?

GPT-Image 2.5 Sunburst 发布了,它最大的改进之一就是能够对同一张图片反复进行精确编辑,而不改变其他任何东西。

于是我决定对它做一次压力测试。

我拿一张图作为输入,用提示词要求更换 T 恤,然后把输出作为下一次提示的输入。

这个视频并不一定会让这个模型显得多出色。

但如果你用其他模型这样做,它几乎撑不到第 3 轮就会变得像噩梦里的东西。

而在这里,即使到了第 5 轮,生成结果也不算太糟。

用 AI 编辑图片时,你通常很少会超过两轮,之后就会干脆用原始图片重新开始。

GPT-Image 2.5 正迅速成为我最喜欢的 AI 图像模型,我非常喜欢 Flare 的速度和 Sunburst 的精确编辑。

溯源

原作者
alecwilcock
发布日期
2026-09-17
原始出处
查看原帖 ↗
收录来源
xianyu110/awesome-gpt-image2.5 · 2100555617941307601 · MIT