Psnr seems weird in error metric

When comparing source image and compressed texture, PSNR in Error Metrics seems higher than my own calculation. My result keeps the same with astcenc.

After some digging (making two one-pixel image example with (0,0,0) and (1,2,3), I found PVRTexTool seems relying on RMSE rather than MSE. Wikipedia shows 10*log_{10}(255*255/MSE), but when I changes to 10*log_{10}(255*255/sqrt(MSE)) I can get just the same result with PVRTexTool.

Hi qiankanglai,

Thanks for your detailed post. After reviewing the issue with the Tools Team, they’ve confirmed you’re right.

We can provide you an internal version of PVRTexTool with the issue fixed. You just need to request it through our support portal ( Also please specify the Operating System you’re using.

Best regards,