I think the timelapse style stars is a feature and not a defect, although it goes against the original challenge.
Although this article is about the challenge - decent photography with a mobile phone - it does outline how easy it is to layer up lots of images in Photoshop, median everything out and get a long exposure image. Taking out the sensor 'median' was clever too.
So you could use this with DSLR images too, to take better long exposure images whatever the sensor, so long as everything is fully HDR and manual.
I think that I might just give it a go. With PHP Imagemagick so that I can automate the Photoshop part and tweak settings easily.
By "taking out the median" you're talking about what's known as dark frame subtraction. I believe most DSLRs already do this internally for long exposures - try setting your camera for 10s exposure and time how long it takes before it's ready again; I'll bet you dollars to donuts it's about 20s. Smartphones can't do this because they lack an internal shutter.
On advanced models (and hacked cameras like MagicLantern/CHDK), you can turn this off and do it manually, e.g. shoot a dark frame for only the first image in a series so you can get ~95% duty cycle rather than ~45%. Especially useful if you're trying to capture a rare event, e.g. lightening.
What's weird about the technique in TFA is he takes N bright frames and then N dark frames, computes both medians and then subtracts. By interleaving bright and dark frames, and doing the subtraction before stacking, I'm pretty sure they'd get better results.
As for image stacking, there's tons of ways to do it, and quite a few turnkey apps for image stacking for astrophotography as well, e.g. here's a review from last year of a Mac app doing just this:
Doing all the bright frames together and all the dark frames together is surely suboptimal, but it's hardly weird: there's an obvious reason for it. (Namely, that the transition between bright and dark requires covering the camera lens with black tape and it's easier to do and undo that once than 64 times.)
If the sensor is stable over the period when the frames are shot, why would subtracting the mean of the black frames from the mean of the exposures be worse than subtracting individual black frames from individual exposures? Aside from rounding errors and possible overflows the two procedures should be equivalent:
(e1 + e2 + ... en) / n - (b1 + b2 + ... bn) / n = (e1 - b1 + e2 - b2 + ... en - bn) / n
The suggestion isn't about what order you do the arithmetic in, but about what order you do the captures in. ++++---- versus +-+-+-+-. The former has the disadvantage that if something that affects the images (temperature, say) is varying gradually, or changes abruptly at a particular point, it's easier for it to have substantially different effects on the light and the dark frames.
It's possible that something like Thue-Morse (+--+-++--++-+--+ etc.; one way to define this is to look at the parity of the number of 1-bits in the binary representation of the frame number) might be better than alternating. If whatever disturbances you might worry about are smooth in the right sort of way, it gives you more exact cancellation than alternating.
Although this article is about the challenge - decent photography with a mobile phone - it does outline how easy it is to layer up lots of images in Photoshop, median everything out and get a long exposure image. Taking out the sensor 'median' was clever too.
So you could use this with DSLR images too, to take better long exposure images whatever the sensor, so long as everything is fully HDR and manual.
I think that I might just give it a go. With PHP Imagemagick so that I can automate the Photoshop part and tweak settings easily.