SourceOn the Analysis of GAN-based Image-to-Image Translation with Gaussian Noise Injection
The paper evaluates the released HiFaceGAN on a 16x face super-resolution task (32 to 512 pixels) using 1000 test images from FFHQ. Under five noise types at six intensity levels, FID degrades sharply with noise. For Gaussian noise, FID rises from 34.83 (clean) to 320.41 (s6), while PSNR drops from 22.09 to 15.11 dB. The degradation pattern is consistent across all noise types, with color noise producing the worst FID (328.49 at s6).