Remember this link about PSNR?

It’s expressed in dB, a log function. In theory, a difference of 0.5 dB MAY be visible under a controlled environment. Since the differences between the top 7 shrink programs are within plus/minus 0.5 dB, I doubt that anyone could see the improvement in real-world condition.

Funny how Nero expresses the PSNR up to three decimal places. The variations from test-to-test can change the PSNR by 0.1 dB.

Running Recode in High Quality mode improves the PSNR by less than 0.5 dB over Fast Mode. To me, this is a measurable but not visible improvement.