The algorithm is pretty similar to that of temporal smoother (DNR is faster, but more limited). What it does is quite simple: it compares the value of a pixel to its value in the previous frame. If the difference between them is less than the threshold you set in the configuration dialog, this is considered the result of noise and the pixel's value is substituted with the old value. If the difference is greater, this is considered motion and the pixel is let through (and the reference pixel is updated).
That was my original version. Avery Lee (phaeron) ported it to use MMX instructions (hence the shared credit), but he introduced a bug. This resulted in noise not being locked into place, but blended (when updating the reference value). This turns out to have a quality-enhancing effect, because the reference value has the noise smoothed out of it. It does however mean more changes in the frame and therefore won't help compression as much.
