The guys from RAWDigger/FastRawViewer (Illiah if you are around here, I have a couple of suggestions below) have made a nice software to extract two RAW files from a Dual Pixel Canon (Canon 5D IV at the moment?) RAW file, allowing one extra stop of DR (this matches on a Canon 5D IV the Nikon D850 and Sony A7R III DR). The explanation of how it works here:
https://www.fastrawviewer.com/DPRSplitI downloaded a Canon 5D IV Dual Pixel RAW file with a partially saturated sky. With DPRSplit I extracted the embedded two RAW files in DNG format; it works flawlessly. I developed equally both RAW files correcting exposure by 1EV to match them, and compare.
Quite notorious parallax. Actually I don't know how Dual Pixel works and how can it produce increasing parallax away from the focus plane, but 0 parallax over the focus plane:
Regarding noise and sharpness, the -1EV RAW file is noisier (about 1 stop noisier in the shadows) and sharper. I gues the increase of sharpness is due to the half pixel sampling which increases horizontal resolution, but also makes the capture more prone to aliasing. Sharpness freaks,
I wonder if this sharpness increase could be a reason to use Dual Pixel by itself...
Let's go to the
Dynamic range story: since the normal shot has the G channel saturated in the RAW file, but the -1EV RAW file has it intact as can be seen in the RAW histograms:
RAW 1: G clipped
RAW 2: all three channels intact
A standar RAW development (without colour inpaint algorithms) produces a monochrome sky, while the -1EV RAW file has the right colours throughout the entire sky:
An optimum fusion would take the sky from the -1EV RAW file and the rest from the normal RAW. It's important to minimise progressive blending since the parallax effect would ruin sharpness in the overlapping areas. I have calculated the pixel to pixel relative exposure in both RAW files, and the expected gaussian bell splits in two. I cannot interpret this now, but the relative exposure falls well close to 1EV as expected:
An optimal fusion (obtained with
Zero Noise) suggests this fusion map:
Black areas are taken from the regular RAW file and white from the -1EV. Standard HDR blending software could produce issues if a hard blend is not performed. I would suggest simply to develop both RAW files separately, match exposure, and use a hard layer mask to "recover" the clipped highlights.
Two suggestions for the DPRSplit team:
- Its Windows version doesn't allow window resizing and you need at least Full HD vertical resolution to fit the app on the screen. I had to install it on a second Full HD laptop.
- Not sure if this is DPRSplit responsibility, but developing the DNG files with DCRAW, it uses a saturation point value of 12400. A closer inspection of the RAW file I downloaded reveals a more precise saturation point of 14335; log2(12400/14335)=-0,21EV. This means using the DNG sat. point can yield to unnecesarily clip up to 0,2EV in the highlights. In a context where we are trying to bring 1EV back to life, this is not desired.
Regards!