Current image restoration tools struggle with real-world dirt and scratches because their training data is too limited or mismatched to actual conditions. While powerful closed-source models like Nano Banana Pro can fix these issues seamlessly, they demand immense computational power that is difficult for many users to justify. Researchers address this gap by building a massive new dataset covering nine common degradation types and fine-tuning an open-source state-of-the-art model to match the performance of expensive alternatives. To prove its efficacy, they created RealIR-Bench, a specialized benchmark comprising 464 real-world degraded images designed to test both cleaning ability and image consistency. The results are striking: this new open-source approach now ranks first among all open-source methods, effectively closing the gap with proprietary systems without requiring massive hardware upgrades.

Source: "RealRestorer: Towards Generalizable Real-World Image Restoration with Large-Scale Image Editing Models" by Yufeng Yang et al., https://arxiv.org/abs/2603.25502