Update model card

README.md

@@ -18,9 +18,10 @@ datasets:
 - marklpd/PSRGAN-Sentinel-2
 base_model: xinntao/Real-ESRGAN
 metrics:
+- lpips
+- vgg
 - psnr
 - ssim
-- lpips
 model-index:
 - name: PSRGAN
   results:
@@ -32,15 +33,18 @@ model-index:
       name: PSRGAN Sentinel-2 Super-Resolution (test split)
       split: test
     metrics:
+    - type: lpips
+      value: 0.1971
+      name: LPIPS (AlexNet)
+    - type: vgg
+      value: 5.645
+      name: VGG19 perceptual loss
     - type: psnr
       value: 21.23
       name: PSNR (dB)
     - type: ssim
       value: 0.5556
       name: SSIM
-    - type: lpips
-      value: 0.1971
-      name: LPIPS (AlexNet)
 ---
 ![PSRGAN grid](psrgan_grid.jpg)
 
@@ -124,6 +128,8 @@ All numbers are on the test split of [PSRGAN Sentinel-2](https://huggingface.co/
 
 PSRGAN beats Real-ESRGAN in PSNR on 98% of test patches and in LPIPS on 99.9%. Compared with the likelihood stage, the adversarial stage gives up 0.60 dB PSNR and changes LPIPS from 0.216 to 0.197 (lower is better). Bicubic upsampling keeps a higher PSNR than the final model on 74% of patches, because PSNR rewards smooth averages over sharp but slightly misplaced detail. LPIPS (lower is better), which follows perceived similarity, is lower for PSRGAN than for bicubic on 100% of patches.
 
+In contrast to LPIPS, PSNR and SSIM are rather low, partly because of how the data was made. The HR targets were resampled with nearest neighbour from the 10 m Sentinel-2 grid onto a latitude-longitude grid (about 10 m × 15.6 m per pixel), so edges in the targets jump by up to one pixel in an irregular pattern that the 4x downsampling erases from the input. Plus, the random blur removes detail that no model can recover, so sharp detail a model adds rarely sits exactly where it is in the target, which both metrics count as error.
+
 On the 128 test patches from Herzogenaurach, Wolfsburg and München:
 
 | method | PSNR (dB) ↑ | SSIM ↑ | LPIPS ↓ | VGG ↓ | OT ↓ |