RealRestorer: Towards Generalizable Real-World Image Restoration with Large-Scale Image Editing Models
Yufeng Yang, Xianfang Zeng, Zhangqi Jiang, Fukun Yin, Jianzhuang Liu, Wei Cheng, jinghong lan, Shiyu Liu, Yuqi Peng, Gang YU, Shifeng Chen
2026-03-28
Summary
This paper focuses on improving how well images can be restored to their original quality when they've been damaged in realistic ways, like through blur, noise, or bad lighting.
What's the problem?
Current image restoration programs often don't work very well on real-world images because they haven't been trained on enough examples of those kinds of damages. While some really good restoration programs exist, they're usually very large, expensive to run, and aren't publicly available for anyone to use or modify.
What's the solution?
The researchers created a large collection of images with nine different common types of real-world damage. They then used this dataset to train a powerful, but openly available, image restoration program. They also created a new set of test images, called RealIR-Bench, and specific ways to measure how well the program removes damage and keeps the image looking consistent.
Why it matters?
This work is important because it provides a strong, open-source tool for restoring images that can be used by anyone, without needing huge amounts of computing power or expensive software. This is especially useful for applications like self-driving cars and object recognition, where clear images are crucial for accurate performance.
Abstract
Image restoration under real-world degradations is critical for downstream tasks such as autonomous driving and object detection. However, existing restoration models are often limited by the scale and distribution of their training data, resulting in poor generalization to real-world scenarios. Recently, large-scale image editing models have shown strong generalization ability in restoration tasks, especially for closed-source models like Nano Banana Pro, which can restore images while preserving consistency. Nevertheless, achieving such performance with those large universal models requires substantial data and computational costs. To address this issue, we construct a large-scale dataset covering nine common real-world degradation types and train a state-of-the-art open-source model to narrow the gap with closed-source alternatives. Furthermore, we introduce RealIR-Bench, which contains 464 real-world degraded images and tailored evaluation metrics focusing on degradation removal and consistency preservation. Extensive experiments demonstrate our model ranks first among open-source methods, achieving state-of-the-art performance.