@IDA_ML
As @a1ex says, there are many misconceptions out there regarding 'noise'. Also, total noise is made up of many components, eg photon shot noise and 'electronic-camera-related noise' etc.
Anyway, as @a1ex says, higher ISOs have less read-noise, as seen in this graph for a 5D3 taken from photonstophotos:
free photo hosting websitesNote log2 scaling, and that once the ISO enters the ISO invariant zone, around 1600-3200, the images become (relatively) ISO noiseless, ie no more noise is added and the electronic noise is at its lowest, eg a few electrons per photo site.
Of course, noise, on its own is not a good metric for IQ. You need signal; and at high ISO signal is low.
As for alignment, unless you are radically moving the camera, the auto align algorithms seem to do well. At least from my experience using HDR Merge in LR: BTW I don't use Photomatix.
Bottom line: This version of the script works reasonably well, but I'm still experimenting with strategies. For example, I just tried a version that took n max ISO images on top of the ISO bracketed ones. My thinking was to exploit the sqrt(n) noise reduction. But the issue is post processing all the differently processed sub-brackets, eg can't process all of them at one go. I need to do some more thinking if I'm to get the last ounce out of the script ;-)
BTW here is another image I processed from my Winchester trip, that I think shows the power of the technique - once again hand held:

POSTSCRIPT: Just tried Photomatix (I have it, but gave up using it in favour of LR-HDR Merge), and in fusion-natural mode it tooks pretty good.