Originally posted by clackers Well, there isn't noise between pixels, Rob, there are only pixels and their values, what's this thing do differently, do you guess?
Conventional noise reduction involves neighbour sampling ... you see one green pixel surrounded by red pixels, so you replace it with a red pixel also. It's lossy.
Yes, my thoughts as well...for me, even if it's "just" a really smart NR algo that is better than most commercial ones, baked in, "cooking" the raws but on a very restrictive scale resulting in what appears to be better high-iso raws, that could still be a good feature, but still not something requiring hardware (other than for faster processing).
The 14-bit data should be able to house enough DR (and a fair bit of noise) without giving the internal processor any advantage in acting on the "unprocessed" raw data ....so I'm not sure what's left that it can do...perhaps things like sensor temperature can be incorporated into the NR to give a superior result compared to traditional NR?...or some sneaky blackframe-ish subtraction or similar based on previous images or knowledge of some kind of noise bias of individual pixels over time or some other tricks like that? I guess it's their secret sauce =), but if anyone knows and can "prove" that it's not just good built-in PP, I'd be all the happier =)
(But I guess it will be easier to compare this time; electronics and other vary quite a bit between the k3 and kp, so attributing all noise improvements to the accelerator may be hard to prove, but since they even offer an upgrade this time without a sensor change, I guess mk1+ software pp and mk2 would be easier and more reliable)