A Appendix

Following the methodology in R. Zhang et al. (2018), we did not select our final model based on a This process is illustrated in Figure 1. The first row shows that the ImageNet-C Gaussian noise corruptions (with standard deviations of 0.08, 0.12, 0.18, 0.26, 0.38) exactly aligns with our additive Gaussian noise for all four models, indicating that the evaluation was correctly calibrated. The abscissa reflects the scale of the corresponding metric; e.g., MS-SSIM ranges from 0 to 1, where 1 means the two images are identical, and 0 means the two The other three models are Euclidean distances (i.e., 0 means that More image examples are provided in Figs. Zoom corruptions discussed in Section 3.3.