Karl_Guttag wrote:
Like others on this forum, I was finding some impressive results using DXO PL4 with DeepPrime on some random photographs. I decided to take a series of photographs with a test setup varying ISO.
I used the R5 with the RF15-35 @19mm f8.0 shooting a bunch of different subjects on the ground in the center of the picture. I side illuminated (only) the test patterns with two 4" by 6" LED arrays.
The test patterns consisted of carpet (to simulate animal fur, including some random white hairs donated by my dog); a screen mesh that is tilted up to the camera by about 45 degrees (propped up on a roll of Duck tape); some converging line targets printed out by color laser printer, an X-Rite Passport color checker, and some old Kodak grayscale and color charts.
It turns out the mesh due to its tilt, has roughly 1 wire per every two pixels horizontally and about 1 wire for every 3 pixels vertically. It is propped up so it will not moire with its own shadow. There is a lot of text, particularly from the Kodak charts that have about 1-pixel wide strokes.
The picture below shows the setup with a dotted line showing the area that will be used in the subsequent comparisons. I set the camera up in aperture priority and used Canon's iPhone AP to step through every ISO from 100 to 51,200 in 1 stop increments without touching the camera.
I have looked at all of the ISOs stacked in layers in Photoshop going back and forth at 100% and 200% and I would subjectively say that DXO4's DeepPrime seems to give at least 1 to 2 stop advantage over DPP4 in terms of noise and sharpness. And DPP4 is much better than what I can get out of Adobe Camera RAW.
THE EXIF DATA that pops-up with the pictures is WRONG. Go by the red text on the pictures. I stacked all the images in a single PDF file.

Below is the Canon's DPP at ISO100 with sharpness settings Strength=3 and Finess=2 (whatever the heck it means). Note you will want to view this at 100%

The exact same CR3 file was converted with DXO PL4 using DeepPrime using its default sharpening of 100% with a radius of 0.5:

What amazed me is that with DXO DeepPrime you can see every wire in the mesh whereas with DPP4, the wires alias. Simple sampling theory says they should alias, but apparently, DXO DeepPrime is doing some significant (AI?) processing to avoid aliasing. You can also see in the converging lines black and white target more detail.
What is also interesting, is that you can still see the grid with DXO processing all the way to ISO6,400 (see below at 100%). Things are getting a little mushier, such as the text at ISO6,400 with DXO, but in terms of straight resolution, in some ways, it is beating DPP4 at ISO100.
I have looked at all of the ISOs stacked in layers in Photoshop going back and forth at 100% and 200% and I would subjectively say that DXO4's DeepPrime gives about a 1 to 2 stop advantage over DPP4. And DPP4 is much better than what I can get out of Adobe Camera RAW. With a well-exposed picture, it is hard to seem much different between ISO100 and ISO800 with DeepPrime (yes you can find a little if you pixel peep).

Now let's go to the far extreme of ISO51,200.
First DPP4. At ISO51,200 is about what you might expect. Very mushy with large blotches and colors fading:

Now for DXO4 DeePrime at ISO51,200. The picture has a lot of fine grain, but has much better resolution and much better colors. Look at the X-rite color squares, they are almost exactly the same as at ISO100 albeit with some fine grain noise. The one "funny" thing is that the various squares in the screen mesh have some false color tint to them. But overall the image is degrading gracefully:

Disclaimers/Comments:
- I have no relationship with DXO whatsoever. I learned about DXO with DeepPrime on this forum a few months ago.
- There are hundreds of knobs to be turned on in the various conversion software packages. I'm looking for one that works well most of the time with minimal effort.
- I am an engineer by training and photography is a hobby.
- I have read a lot of good reviews of DXO PL4 with Deeprime. I was amazed how good some test pictures I took of my dog at ISO51,200 came out with DXO DeepPrime and wanted to try and quantify how much it helped.
Thanks for sharing this good data.