This blue clipping issue is driving me nuts.
I've attached a pixel peep screen shot of a corner detail of a night scape shot of an artificial narrow band light, made with an A7r-II. It's one of the shots I've got in a PM, I took the freedom of posting this size crop without asking, hope it's okay. I normally always ask, but this size from a high MP shot I thought was ok.
The first is without profile (eg identity matrix after WB) that shows the actual tonality in the file. Then to the right is adobe's (before looktable+curve) which is quite close to the identity matrix, that is tonality is mostly kept (they mess up a bit when adding looktable+curve but not too bad).
And in the middle is DCamProf (before looktable+curve). Here we see it's much bluer, so blue in fact that much of the tonality is lost, and a sharp transition into cyan. Why is this? When the matrix is optimized to match the CC24 a very good match can be had, but that requires a great deal of blue subtraction, otherwise the dark blue patch becomes much to light, and other colors as well. In theory the LUT could compensate that, but that leads to strong bends, and when relaxing you end up with quite large errors.
This can be seen with Adobe's profile, which doesn't match CC24 colors that well, and thus is not a very accurate profile. But in return very robust with extreme colors.
It seems like some subtraction is okay, say if the Yb value is -0.10 or larger you can still have good tonality (which it is on my Canon for example), but for the A7r-ii and some other Sony sensors it's like -0.28, and then blue is pushed down to the ground in the extreme range. I don't know why these sensors have this type of blue sensitivity, someone that dare to speculate?
By limiting the matrix optimizer to not have (large) negative blue value I can get about the same result as Adobe, but then accuracy in the normal range is lost (can't compensate with LUT without too strong bending) and I really don't like the idea to mess up the normal case to be robust in an extreme case.
I'm thinking of the idea to blend between two profiles/matrices, one for the normal range and gradually blending into another for the extreme range. It's a quite big amount of work to test if that even works though, so I'm probably going to add an "known limitations" section and put this issue there.
However, I still have some messy tone operator clipping issues in this range too, that needs separate fixing. So it's certainly going to be better than the current 0.9.13 when finished, but I think Adobe is going to win on "extreme blues" even after completion.