Incorrect PSNR for uint8 images

[ymtoo]

🐛 Bug

PeakSignalNoiseRatio silently produces incorrect results when applied to two randomly generated uint8 images, likely due to an overflow error. However, converting the images to float32 resolves the issue and yields a reasonable PSNR value.

import torch
from torchmetrics.image import PeakSignalNoiseRatio

img1 = (torch.rand(3, 20, 30) * 255).to(torch.uint8)
img2 = (torch.rand(3, 20, 30) * 255).to(torch.uint8)
psnr = PeakSignalNoiseRatio(data_range=255.0)

psnr(img1, img2)
# tensor(27.8814)
psnr(img1.to(torch.float), img2.to(torch.float))
# tensor(7.8037)

Expected behavior

Shouldn't the function raise an error when provided with uint8 inputs?

Environment

python                    3.10.14
torchmetrics              1.4.1
torchvision               0.19.0

[github-actions[bot]]

Hi! thanks for your contribution!, great first issue!