About calculating PSNR and SSIM metrics

[EternalEvan]

Hi, thanks for your great work. I'm reproducting your work and have recovered visually amazing results. However, I find the results of the two metrics PSNR and SSIM not aligned with those in your paper. I'm a newer in blind face restoration task and I can't find the reason about it. Here is the code of my calculation:

def calculate_psnr(img1, img2):

    assert img1.shape == img2.shape, (f'Image shapes are differnet: {img1.shape}, {img2.shape}.')
    img1 = img1.astype(np.float64)
    img2 = img2.astype(np.float64)
    mse = np.mean((img1 - img2)**2)
    if mse == 0:
        return float('inf')

    return 10. * np.log10(255.*255.0 /mse)

which is completely the same with your metric codes and so do SSIM. I send the saved results and the corresponding groundtruth to the function. But my results on CelebA-test is about 5 point higher than your PSNR and 0.2 higher than your SSIM. Can you help me with my problems?