Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data

22 Jul 2021  ·  Xintao Wang, Liangbin Xie, Chao Dong, Ying Shan ·

Though many attempts have been made in blind super-resolution to restore low-resolution images with unknown and complex degradations, they are still far from addressing general real-world degraded images. In this work, we extend the powerful ESRGAN to a practical restoration application (namely, Real-ESRGAN), which is trained with pure synthetic data. Specifically, a high-order degradation modeling process is introduced to better simulate complex real-world degradations. We also consider the common ringing and overshoot artifacts in the synthesis process. In addition, we employ a U-Net discriminator with spectral normalization to increase discriminator capability and stabilize the training dynamics. Extensive comparisons have shown its superior visual performance than prior works on various real datasets. We also provide efficient implementations to synthesize training pairs on the fly.

PDF Abstract
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Video Super-Resolution MSU Super-Resolution for Video Compression Real-ESRGAN + x264 BSQ-rate over ERQA 5.58 # 26
BSQ-rate over Subjective Score 0.335 # 4
BSQ-rate over VMAF 0.698 # 5
BSQ-rate over PSNR 7.874 # 42
BSQ-rate over MS-SSIM 0.881 # 16
BSQ-rate over LPIPS 0.733 # 4
Video Super-Resolution MSU Super-Resolution for Video Compression Real-ESRGAN + x265 BSQ-rate over ERQA 6.328 # 28
BSQ-rate over Subjective Score 0.64 # 10
BSQ-rate over VMAF 1.464 # 30
BSQ-rate over PSNR 8.113 # 44
BSQ-rate over MS-SSIM 5.393 # 61
BSQ-rate over LPIPS 12.689 # 68
Video Super-Resolution MSU Super-Resolution for Video Compression Real-ESRGAN + uavs3e BSQ-rate over ERQA 7.225 # 37
BSQ-rate over Subjective Score 1.417 # 23
BSQ-rate over VMAF 2.122 # 48
BSQ-rate over PSNR 15.144 # 73
BSQ-rate over MS-SSIM 4.612 # 56
BSQ-rate over LPIPS 2.633 # 28
Video Super-Resolution MSU Super-Resolution for Video Compression Real-ESRGAN + aomenc BSQ-rate over ERQA 11.584 # 49
BSQ-rate over Subjective Score 1.398 # 22
BSQ-rate over VMAF 2.712 # 52
BSQ-rate over PSNR 15.144 # 73
BSQ-rate over MS-SSIM 6.857 # 74
BSQ-rate over LPIPS 11.957 # 67
Video Super-Resolution MSU Super-Resolution for Video Compression Real-ESRGAN + vvenc BSQ-rate over ERQA 6.712 # 32
BSQ-rate over VMAF 3.8 # 58
BSQ-rate over PSNR 14.561 # 71
BSQ-rate over MS-SSIM 5.95 # 68
BSQ-rate over LPIPS 12.744 # 69
Video Super-Resolution MSU Video Super Resolution Benchmark: Detail Restoration Real-ESRnet Subjective score 3.697 # 26
ERQAv1.0 0.598 # 27
QRCRv1.0 0 # 21
SSIM 0.824 # 21
PSNR 27.195 # 16
FPS 1.019 # 13
1 - LPIPS 0.871 # 17
Video Super-Resolution MSU Video Super Resolution Benchmark: Detail Restoration Real-ESRGAN Subjective score 5.392 # 13
ERQAv1.0 0.663 # 16
QRCRv1.0 0 # 21
SSIM 0.774 # 29
PSNR 24.441 # 29
FPS 1.01 # 14
1 - LPIPS 0.895 # 11
Video Super-Resolution MSU Video Upscalers: Quality Enhancement RealEsrgan-F PSNR 28.82 # 29
LPIPS 0.185 # 5
SSIM 0.850 # 13
Video Super-Resolution MSU Video Upscalers: Quality Enhancement RealEsrnet-F PSNR 30.01 # 22
LPIPS 0.280 # 20
SSIM 0.868 # 21
Video Super-Resolution MSU Video Upscalers: Quality Enhancement RealEsrgan-V PSNR 25.52 # 48
LPIPS 0.333 # 27
SSIM 0.795 # 1
Video Super-Resolution MSU Video Upscalers: Quality Enhancement RealEsrnet PSNR 30.52 # 17
LPIPS 0.296 # 23
SSIM 0.878 # 26
Video Super-Resolution MSU Video Upscalers: Quality Enhancement RealEsrgan PSNR 29.14 # 26
LPIPS 0.181 # 3
SSIM 0.855 # 18
Video Super-Resolution MSU Video Upscalers: Quality Enhancement RealEsrgan-A PSNR 28.71 # 30
LPIPS 0.244 # 16
SSIM 0.830 # 4

Methods