An Uncertainty-Aware Relational Distillation for Real-world Super-resolution in Real-ESRGAN

In recent years, real-world super-resolution technology has made rapid progress. The super-resolution model represented by Real-ESRGAN successfully restores real images full of complex noise and blur through a sophisticated architecture. However, the computational requirements of such models make them very difficult to deploy on resource-constrained edge devices. Although knowledge distillation is an effective means to achieve lightweight models, there is a lack of research on distillation methods specifically designed for real-world super-resolution models such as Real-ESRGAN. Directly applying traditional rigid feature constraints, when faced with complex high-order degradation, often leads to optimization difficulties due to the capacity gap between teacher and student models, which limits the potential of lightweight models to recover high-frequency textures. In this paper, we propose a relational distillation method based on uncertainty modeling, which adaptively adjusts the distillation intensity in the spatial dimension to achieve flexible constraints on complex degraded regions. At the same time, an improved generative adversarial strategy is used to maximize the efficiency of knowledge transfer. Experimental results on two datasets with high-order degradation strategies demonstrate the universality and superiority of our proposed distillation method.