All my sub images are stored as .cr2 (14 bit), .fit (16 bit) or .tif (16 bit). My dslr camera has a 14-bit ADC so all these three formats store them without loss. Only .cr2 store all of Canon's Exif data. After calibration, dark, flat and more I save them as 32-bit per pixel. Either 32-bit floating point fits or maybe 32-bit tiff, it depents which software I use, but all of them handle the information without loss in normal cases, even after stacking. No reason to not save in 32-bit, memories are cheap today. Why save money and destroy your expensive data you paid so much for to get, both money and time.
The 16-bits formats are integers and the 32-bits are floating points in this case.
The trick is then how to process back from 32-bit to a common 8-bit deap (each color) and still see the weak details in a nebula and don't clip the stars and without get no nice looking image. You increase the weak signal and dampen the high levels, gamma function do that something like our eyes do. More advanced processing do different thing on different areas. All this needs high dynamic data input, at least 32-bit. It doesn't matter if you do averaging or sum. It's the small details that differ. Sometimes the noise is heavy and mask the problem.
The software I normaly use are AstroImageJ and Fitswork. Both are free and handle 32-bit floating numbers. Long time ago I used Matlab which can work with 64-bit. But no need to the images I have today, maybe in future.
Update (now when not in the middle of night):
With a 16-bit ccd camera that are calibrated with masterdarks and masterflats maybe you reach in theory the limit when to move from 32-bit floating point to use 64-bit. Let say we calibrate the sub image with an average masterdark that consist of 50 subs. The means you subtract with a resolution of just 1/50 of a bit. To keep it without truncating it you need only for this calibration step 5 or 6 extra bits. The 32-bit floating point format only has 24 bit, the other are used for the expontent -/+ 127. We have now used 21 of the 24 bit we have avaible. Then we can only stack about 10 images before we loss information. In reality we don't even come close to this because we have a lot of noise that mask this problem. Little bit surprised that it was so close with a 16-bit camera, I have a 14-bit camera and then get two extra bit to play with, then I can add about 40 images when stacking without problem, in theory. Correct me if I did something wrong in my estimations. I will do a calculator later on my homepage to find how many bits needed in theory.
This was for deep sky images, videofiles I don't have very much experience from. But from hometheatre I know 10-bit is far better then 8-bit when processing the videosignal.
Here is a list of what software I use:
http://www.astrofriend.eu/astronomy/...nt.html#part05
Lars