Your proposal seems to boil down to adding a feature for automatic switchover from dp to dpxd based on an editable threshold iso value per camera model. The camera is detected automatically, iso is in metadata, so all that needs to be added is that threshold.
DxO could provide a default that the user can change in a certain range, say ± 2 EV. This could keep the interface simple in what is DPR’s design.
For which type of photography you found this feature useful to introduce?
I think the choice between DP and DPXD depends mostly on the subject,
rather than ISO. DP(XD) adapts to ISO and does it quite well, it seems.
For indoor sports I would use DP only, as probably most people here.
For indoor architectural details, DPXD can be sometimes noticeably better,
if it doesn’t generate nasty artifacts. For the latter, my experience
is still very limited though, so I would admire your thoughts.
You could probably do it using exiftool and editing DOP files using some scripting,
but I’m not sure if it’s worth the effort.
According to things posted by DxO, DeepPRIME XD is not primarily intended to treat full, but cropped images. DPXD seems to produce less artefacts in cropped images…and my guess is that the machine trained parts of the algorithm get less “derailed” under such (less variable) conditions.
That’s a surprise to me. Do you have a link?
I’m new to PL so I may have missed some relevant forum notes.
I’ve read DeepPRIME - Denoising and RAW conversion with AI (dxo.com), which says that for DEEPPrime you may expect 2 stop improvement and 2.5 stops for DP XD over conventional noise reduction, which I think is the case indeed. I also came across a forum note by someone from DxO staff which implied that DP(XD) was adapting to various ISO by design. But nothing there about cropped images.
During last month I did many comparisons on sport photos in low light
(ISO 10k-30k, D780). In quite many cases (with ISO 20+k), DPXD “magnified” minor skin defects beyond my taste, while DP produced acceptable results. That’s why I’m a bit reluctant to use DPXD for faces. On the other hand there were very few cases where DPXD performed a bit better (e.g. eyebrows), you just had to push slightly the microcontrast to the left. In higher ISOs, DPXD leaves much less luma noise in background, but some may prefer the grain left by DP. For low ISO, DEEPPrime with default setting doesn’t do anything noticeable to the picture, unlike “traditional” NR. Hence DP made my workflow faster than in old LR (Smart Lighting being another factor).
Before PL7, for over 10 years I used LR5.7, where you had to manually adjust NR settings, depending on ISO and shadow pulling. Perhaps OP had something like this in mind.
Actually, I think I was the first person to suggest using XD for cropped images during beta testing. I posted on the subject at the time and again after It was released. I read virtually everything that is posted on this site, and I don’t recall DxO ever commenting on it, much less suggesting that DeepPRIME XD was intended for cropped images.
During beta testing I discovered very early on that when viewing images full screen on my 28 inch 4K monitor, the differences between the results with DeepPRIME and DeepPRIME XD was often subtle. But, when comparing those same images zoomed in, the difference in the amount of fine detail retrieved was very significantly greater with the XD version. While I generally use DeepPRIME and DeepPRIME XD almost equally depending on the images I’m processing, for cropped images I use XD exclusively.
Drifting slightly off the topic, but OP stays quiet.
I think I misunderstood you upon first reading.
At first I thought you meant that cropping and exporting to jpeg
produces better results than cropping the exported jpeg image,
but it seems the two operations do commute.
Now I assume you meant that comparing DP and DPXD full pictures on a monitor
(even 4k) will reveal the differences only on cropped images.
In other words, the differences between uncropped images would be seen
only on large prints, or on the monitor when enlarged to 100%.
It has nothing to do with AI/ML inventing and NOT inventing.
Actually, the first time I used DP(XD), the following differences popped,
when viewing full size pictures (at about 50%, so below the 75% threshold
for some corrections to be displayed, 24MP camera, 27"/4k):
background and shade noise (less for DPXD at high ISO)
microcontrast, acquity perception, details (more for DPXD at high ISO)
color (minor, yet to be investigated, perhaps microcontrast related)
jpeg size (often 20% smaller for DXPD, even at low ISO)
exporting time (1.5-3 times slower for DPXD than DP)
I very often see significant amounts of additional fine detail using DeepPRIME XD when comparing DP and DP XD exports of my 21mp raw files full screen and zoomed in on a 28" 4K monitor. However, the amount of extra fine detail visible is also dependent on the quality of the originally raw file, the lighting, the contrast, the content, and how much an image is cropped or zoomed in.
I also find that images captured with my sharpest lenses at very close focusing distances of under 30cm also often show much greater amounts of fine detail with XD, depending on the sharpness and amount of contrast, without the need for cropping or zooming in.
To recap, in my experience XD may not be the right choice for every image, but for many images the results it provides are cleaner with greater amounts of fine detail when compared to the original version of DeepPRIME. In the end, the appropriate use of DeepPRIME XD requires a lot experimentation and experience which may take a significant amount of time to sort out. Most people may not be willing to spend the time needed to identify the XD best use cases for their images.
For what it is worth, I have had a very extensive amount of experience using DeepPRIME and DeepPRIME XD both during their development and after their release, and have used each one thousands of times with a very large variety of images from different cameras and with varied content. Of course, my opinion is based on my own experiences and may vary from the experiences of other users.
I have posted comparative examples in the past and will post some again when I have the time.