As is often the case, DxO’'s documentation is incomplete in its explanation. Since PRIME does not use machine learning or a graphic card’s GPU, and processing times for the DeepPRIME variations differs significantly depending on the graphics card used, I believe their performance comparison is based on running exports for both PRIME and DeepPRIME 3 using the CPU only option. In that scenario on my Windows 10 machine exporting a single image with DeepPRIME 3 took almost 3 time longer to process than PRIME, which I believe was DxO’s point.
When processing DeepPRIME 3 using the GPU of my RTX 4060 the same export processed almost five times faster then PRIME which is similar to your results although my timings were slower than yours because you have a more powerful CPU and graphics card then I do.
Arguably I don’t care if there is something lurking in the woodwork that DP XD2s has “glossed over” if it means I can point my Micro Four Thirds G9 with an F3.5 - F6.3 12-200 (24-400) lens at it and rescue an image taken in very low light with the application of a single preset to all the images, finessing the few where the one preset doesn’t work as well as I might like!
@Wlodek I ran some tests on my high own high ISO images (Panasonic G9 ) from the same visit to Brighton Pavilion, a batch of 50 images with ISO from 1600 to 25600, exported as 100% JPGs.
A Stud earring and a drop earring in a pierced ear according to my wife.
Notwithstanding my concerns about the added Chromatic Aberration with DP XDS2 and the “smoothing” of artefacts, which seem to be a “potential” feature of DP XD2s I still think that there is no contest for noisy images, pick the “best” from those shown below
@zkarj For me DP3 offers a smaller JPG than NO NR and no (or less) added CA than DP XD2s for daylight photos and half the export time but I would prefer that DP XD(?), if and when it appears, resolves the CA problem and improves the accuracy at default settings.
However, what might that cost in GPU memory and export time or will it be like for like with DP XD2s!?
@John-M I was a little shocked at the amount of space the new menu took but this is PL7.13,0 versus PL8.3.1 versus PL8.5.0
But the snapshots are minus the NR “Advanced settings”. How are they going to be accommodated in your scheme, answer just the way they have been before, as an item below the options, sorry I answered my own question as I wrote it!?
Personally I would like the PL8.3.1 scheme simply extended to incorporate all the new elements as and when they arrive.
But your drop down list proposal means that the list can increase for as long as DxO like and the available real estate for the rest of the options won’t change.
DxO have decided to do some flag waving, sorry add some explanatory text for those new to the product, in the UI but is it really unnecessary?
Looks quite normal to me. Denoising reduces randomness, so the data packs better, while sharpening, increasing microcontrast or saturation, make compression slightly less effective.
BTW, thank you for helping me in English
@Wlodek I guessed along those lines but typically never save with ‘NO NR’ except when testing, so the size came as something of a shock when I first saw it some months ago and then again earlier.
Just installed Photolab 8.5 and tested Deep Prime 3 against an older jpg obtained with Deep Prime in Photolab 4.
Well, that’s a huge leap BACKWARDS.
First Deep Prime result was wAY better, finest noise grain and better recovered detail.
For the record, I’m talking about (and working with) CR2’s from Canon 6D mark II, with ISO values of 12800 and above.
Hope something will be fixed (or I hope I’m doing something wrong).
Maybe it’s about lens sharpening or microntrast settings?
I don’t have PhotoLab4 to compare, but just checked DP3/PL8.5 and DP/PL7.13 for EOS 6D II Studio Scene CR2 raws, downloaded from Dpreview taken at ISO 12800, 25600, 51200, and 102400. Applied Optical Corrections preset and set rendering to EOS6DII, exported to jpeg with quality=90 (perhaps should use 100% or tiff for comparisons to avoid additional jpeg artifacts).
Results:
At full 4k screen differences between DP and DP3 are hard to find (default settings). At 200%, the most obvious differences are in three text fragments just above the picture centre. At ISO 12k, DP3 looks just a bit better than DP, while at ISO 25k and 52k the difference is slightly bigger in favour of DP3. At ISO 102k the differences are more noticeable, also in other parts of photo, but DP3 starts to invent things, perhaps not always right (too noisy for human to decide what is “right”). So, DP3 looks slightly better than DP in general, but the differences are subtle and seen only in very noisy pictures at somewhat specific places (at flatly lit ISO 1600 photo you will probably hardly find any differences between DP, DP3, XD, and XD2s).
Hi, tomorrow I’m gonna check it out if I’ll find some time.
But, that said… The dop from which I started was the original one from PL4, so I think the only things that could be changed are the default values for new controls/sliders, nothing else.
And the only options I touched for the test I was talking about are the denoise ones.
Found some time right now… I made a fresh start on a copy of the CR2 and I can confirm that nothing else was affecting the development except denoise changes, and denoise values need to be changed with new version of PL8.5 (and this maybe affects also PR5). To get same result as in past version, for instance, force detail slider must be raised for me on images at highest ISO values (12800+ as stated before).
I am not a “pixel peeper” and do not adjust the noise reduction settings. I just use the defaults and find it works well for me. I did some comparisons of high iso bird images between DeepPrime3 and DeepPrime XD2. DeepPrime3 does leave a bit of noise, but I actually think that is a good thing. XD2 does a wonderful job of eliminating noise, but the image can look sterile. I also find that DeepPrime3 has a bit more contrast. It’s a matter of personal taste, but I am changing my default NR to DeepPrime3. It’s great having this choice.
