There's no increased resolution; the difference may be in extreme shadows (if you ever got posterization). The most obvious difference is with
my first fake_dual_iso test (which is quite extreme, 6 EV between the two exposures).
If you feel you are close to the 16-bit limit, try --soft-film.
Quick sample (fake dual ISO 6EV):
1. 16-bit engine (522eb2e), developed with +10 EV in ufraw-mod, --clip=digital
2. 20-bit engine (b13fcd5), developed with +10 EV in ufraw-mod, --clip=digital
3. 20-bit with --soft-film=4, developed with +6 EV in ufraw-mod, --clip=digital
The --soft-film option compresses the highlights and bakes an exposure compensation in the DNG. With this trick, it uses more bits for shadow detail and you get a brighter image to start with. The baked curve looks like this:
For baking a soft film curve, you need to know the WB in advance, at least roughly (read the multipliers from ufraw, for example). In all the examples I used 4000K/green=0.8, which has the multipliers 1.899, 1, 2.173. Since the default WB for these baked curves is 2,1,2, this fits under the engineering definition of Pi (which is 3), so I didn't bother changing the multipliers for now.
Let's play with it a bit more:
4. 20-bit with --soft-film=8, developed with +2 EV in ufraw-mod, --clip=digital
5. 20-bit with --soft-film=9, developed with +1 EV in ufraw-mod, --clip=digital
6. 20-bit with --soft-film=10, developed with +0 EV in ufraw-mod, --clip=digital
Now, let's try to bake the same WB as the one used for developing, to see how much it matters:
7. 20-bit with --soft-film=10 --wb=1.9,1,2.17, developed with +0 EV in ufraw-mod, --clip=digital
Notice the subtle WB change in highlights.
Alright, so we got pretty good highlight/shadow detail now. How does it compare with ufraw-mod's --clip=film?
8. 16-bit engine, developed with +10 EV in ufraw-mod, --clip=film
9. 20-bit with --soft-film=4, developed with +6 EV in ufraw-mod, --clip=film
In image 8, notice how ufraw-mod sacrifices 2 shadow bits in order to get better highlight detail (compare with 1). That's a limitation of its 16-bit processing engine. Floating-point raw processors would not suffer from this => consider the highlight compression as a trick to bypass the bit depth limits in either the DNG container or in the raw processor itself.
So, you have noticed this curve is exactly the soft-film curve from
ufraw-mod, and now you have the option to bake it in the DNG (you can consider it as a pre-grading). You can bake it completely (images 6,7) or just a part of it (9). Experiment and see what works best.
Experimental binary:
cr2hdr-20bit.exe
This example was extreme (I chose it to push the code beyond reasonable limits). In practice, the improvement of the 20-bit engine is probably minor, I'm not sure if it's worth the small speed penalty and increased memory usage. You tell me (if you find differences, just post some examples).
Happy pixel peeping!
