HyperRes: Efficient Hypernetwork-Based Continuous Image Restoration

Continuous image restoration attempts to provide a model that can restore images with unseen degradation levels during training at inference time. Existing methods are limited in terms of either the accuracy of the restoration, the range of degradation levels they can support, or the size of the model they require. We introduce a novel approach that achieves the optimal accuracy of multiple dedicated models for a wide range of degradation levels with the same number of parameters as a single base model. We present a hypernetwork that can efficiently generate an image restoration network to best adapt to the required level of degradation. Experiments on popular datasets show that our approach outperforms the state-of-the-art for a variety of image restoration tasks, including denoising, DeJPEG, and super-resolution.