Details or Artifacts: A Locally Discriminative Learning Approach to Realistic Image Super-Resolution (CVPR2022)