How-toSelected

MIRAGE algorithm reduces noise in AI-generated images

Why we picked it

This is a practical tool for anyone using AI image generation, as it helps prevent the common problem of image degradation during editing.

A new algorithm called MIRAGE, developed by a team from Shanghai Jiao Tong University and the company 小山 (Luoxiaoshan), effectively reduces noise in AI-generated images.

Body

The algorithm addresses issues like repeated textures and grain that appear in images after multiple edits.It works on models such as Image 2.

0 and Image 2.5, and the team has released the code and research paper to help improve image quality.

Original · 向阳乔木

实用啊,AI垫图多了就容易出现噪点。 小山和上海交大合作搞的算法能非常有效降低这种情况。 让你生成的图片质量更高,尤其GPT 2.5 image 也上线到Codex了。 这个技巧和方法可以用起来。 洛小山 @luoxiaoshan_ai 相信你对这种崩坏应该不陌生吧。 第一张 AI 图明明很惊艳。你只是想改一下头发、换个背景,或者微调构图。 可到了第三轮、第四轮,画面突然开始失控了。 头发变成层层叠叠的鳞片,树林和墙面长出密集、重复的颗粒,整张图还会蒙上一层细密格纹,看起来又脏又假。 最让人崩溃的是:当你继续编辑之后,问题往往越来越重;直接降噪,正常的发丝、树叶、笔触和材质也会一起消失。 而且换成新模型后,这个问题并没有彻底消失。 我们在 Image 2.0 和 Image 2.5 的产物中都观察到了类似现象。 为了能为 Alice 提供更好地生成效果,保障用户体验。 我们做了 MIRAGE:先诊断问题属于哪一种,再选择精确频谱陷波、结构保护处理,或净化垫图后的重新生成。 在我们的实验中,MIRAGE 对 Image 2.0 和 Image 2.5 都取得了不错的修复效果:明显减轻格纹、鳞片和重复纹理,同时比普通降噪保留更多真实细节。 现在,我们把代码、论文、测试、完整对比和 Agent Skill 全部开源。 希望下一次图片在反复编辑后崩坏时,你不用再从头开始。 GitHub: github.com/miyang-ai/mira… 🔗 View Quoted Tweet📊 1 ⚡

Industry Coordinates & Source Verification

This report is categorized under AITOP's "How-to" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include MIRAGE, denoising, prompt engineering.

Source attribution: curated from 向阳乔木 (@vista8).