Generative augmentation for blind image super-resolution

Blind image super-resolution (SR) has made significant progress; however, its development remains constrained by the lack of paired datasets containing real low-resolution (LR) and high-resolution (HR) images. Existing methods typically rely on synthetic degradations generated from simplified models, where noise levels, blur kernels, and compression artifacts are randomly applied to HR images. Unfortunately, such handcrafted degradations fail to reflect the true complexity of real-world imaging pipelines, leading to a substantial domain gap between synthetic training data and real-world test conditions. To overcome this limitation, we propose a generative augmentation framework that learns real degradation characteristics directly from real-world images. Our method generates realistic LR counterparts from clean HR inputs, enabling the construction of large-scale, degradation-aware training datasets that closely align with practical scenarios. Extensive experiments show that models trained on our generated data achieve superior performance on multiple real-world benchmarks, effectively reducing the domain gap and outperforming state-of-the-art blind SR approaches.