MIRNet

Why we included this project

MIRNet is a convolutional network for low-level vision tasks like image denoising, super-resolution, and general enhancement, and it has become a common reference point for teams working in that area. Its main design idea is to keep high-resolution features spatially precise through the whole network while drawing contextual information from lower-resolution streams, an approach that later restoration models have built on. The repo includes the ECCV 2020 training and testing code in PyTorch, pretrained weights, and evaluation scripts, so reproducing the published results or adapting the architecture to a custom pipeline is straightforward. One thing to weigh before adopting it: the license only permits non-commercial use, which rules it out for most product work.