Yeah, if it is captured on digital, then a single image isn’t going to be helped, as every source will be from the digital master.

With prints scanned to film however, I think you end up with the same result, by stacking images correctly, you throw out the noise and are left with the information. With enough samples the chances of only that ‘single’ pixel being correct is basically nil, or at least, less likely to happen than the chances of, for example, incorrectly identifying the centre of a dye cloud.

I’d be interested to see this play out though, I’ll go through my siggraph library to see if anyone is already doing this, there was a lot of research into this kind of thing in the late 90s and early 2000s, there might be something in there. From a programming point of view, it wouldn’t be difficult to implement.