Title:One Size Fits All: Hypernetwork for Tunable Image Restoration

Abstract:We introduce a novel approach for tunable image restoration that achieves the accuracy of multiple models, each optimized for a different level of degradation, with exactly the same number of parameters as a single model. Our model can be optimized to restore as many degradation levels as required with a constant number of parameters and for various image restoration tasks. Experiments on real-world datasets show that our approach achieves state-of-the art results in denoising, DeJPEG and super-resolution with respect to existing tunable models, allowing smoother and more accurate fitting over a wider range of degradation levels.