PhotoLab DeepPrime vs. others

Hi.

I made a test to compare the PhotoLab DeepPrime with other RAW processors.

Interestingly, in this example the PhotoLab DeepPrime is the worst while JPEG from camera is the best.

Source RAW (Canon EOS R6 II, ISO 100):
https://www.dpreview.com/reviews/image-comparison/download-image?s3Key=9dc4402898fe4882950f3b1ec600daa0.cr3

In this example, you can see many unwanted artifacts in DeepPrime and even more in DeepPrime XD.

Based on the screenshots you provided, Vladimir - I don’t share your conclusion.

The top-right / red-box for the JPG version looks pretty bad (to me): image

2 Likes

Hopefully the images are now in the correct position in the post!!

@Vladimir I am concerned about the quality of your comparison images so I exported the image in

  1. No NR
  2. DP
  3. DP XD

and compared the three image sections (and this is a snapshot and how it will appear in the forum posts I do not know and I am losing the titles when I upload!?)

I then compared your snapshot (forum download) with mine as closely matched as I could (we are now comparing snapshots of a tiny section of a much larger image, one of which is a download to another snapshot)

i.e. this snapshot of your post image

to give

These are snapshots from a small section of an image that are then compared. You may well be correct but a better means of comparison is required I believe, and I have been going round in circles (and the screen quality degrades a little further when you look at the snapshots via the forum)!

I am unclear what is the comparison about exactly? I never found any unwanted artifacts, especially with DeepPrimeXD, so I’m guessing you made some other changes or something. In fact, DeepPrimeXD because of its Demosaicing and noise reduction, even at base ISO produces best and cleanest results of anything else I’ve seen.

Your comparison with the circles is a problem to compare visually since all group together and its symmetrical. You should do a more organic testing, human face or something similar to that. But


I agree with John-M

1 Like

@Vladimir this is an export (100% JPEG) created with DP XD, please identify the artefacts as @MSmithy suggested (the zip contains the full size JPG because the forum automatically reduces the size of a JPG when uploaded).

5M9A0093_PL631R-2.zip (17.2 MB)

1 Like

In these examples you can see what I mean:

For original pictures visit:

But these don’t like like artifacts, but over-sharpened edges. What are your other settings relating to lens sharpness or sharpening? By combining settings between “lens sharpening” and “noise reduction” you can get the look you want. Depending on the profiles for lens or lens sharpness itself you may need to reduce lends sharpness or play with sliders. Also DeepPrimeXD has ability to change how much noise reduction/demosicing you apply as well as what you prioritize. Noise reduction or texture. I doubt that you would not be able to get perfect results by tweaking the settings.

sshot-1858

1 Like

These are not over-sharpened edges. Even if you reduce the sharpness to the minimum, you will see these artifacts that were created by the DeepPrime algorithm.

Of course I tried changing all these values you suggest, but they had no effect on the result. Have you even tried it yourself?

OK , I finally download the RAW file and tried it out. And best I can see what you are calling artifacts is actually moire pattern, common to the type of full frame camera like (Canon EOS R6 II). DP review has actually created the test studio image to make sure any type of defects, including moire pattern can be detected, and that is what you are seeing. Its not a problem of DXO PhotoLab or DeepPrimeXD, its a matter of senor / lens.

I’ll illustrate this with some examples.

sshot-1860

That is more visible because of level of details that is achieved via lens sharpness settings and DeepPrimeXD.

Here is is with lens sharpness off. It is as you would expect less noticeable.

sshot-1859

Furthermore. There are other places in the test image where moire pattern is noticeable.

You can minimize the moire pattern with moire slider but its not ideal.

“MoirĂ© pattern occurs when a scene or an object that is being photographed contains repetitive details (such as lines, dots, etc) that exceed sensor resolution. As a result, the camera produces strange-looking wavy patterns.”

To illustrate that this is not a problem of a DXO , you can use camera with higher resolving power and no moire pattern problem.

Fujifilm GFX 100/100s for example.

DPreview has test file if you like to try it. Same DXO settings as on Canon.

2 Likes

How can this be a matter of sensor / lens when the artifacts are not visible when you change the denoising algorithm or when you see the result of the JPEG from the camera?

The Moiré slider has no effect to this:

1 Like

I imagine its related to demosaicing process.

A demosaicing (also de-mosaicing, demosaicking or debayering) algorithm is a digital image process used to reconstruct a full color image from the incomplete color samples output from an image sensor overlaid with a color filter array (CFA)

JPEG from Camera is shot as RAW and converted to JPEG in camera, based on some propitiatory demosaicing and denoising process from the camera manufacturer. You have no control over it, and I imagine since its done by the people who build the camera, its optimized for that camera.

PRIME denoising is just denoising, while demosaicing is done separately.

Deep Prime is demosaicing + demoseicing with the help of AI, but obviously DeepPrime XD is the same, just with better more accurate processing.

This example you are using is shot at base ISO, so its pretty clean image to start with. If you want to see a more noticeable difference, try higher ISO shots.

Like I said in the previous reply. Try the same scene with a different higher res camera and use the same settings in DXO and you will see what happens. That should demonstrate that if the scene and post processing application remain the same, but camera and lens change it must be the main reason behind what you are seeing.

"Typically, denoising and demosaicing are completed separately. DxO DeepPRIME adopts a holistic approach and combines the two steps into one, leading to exceptional performance.

The results are spectacular: DxO DeepPRIME removes noise and recovers information without losing detail in the image’s colors, creating more even and natural transitions. Compared to DxO PRIME, the DxO algorithm which previously set the standard for denoising in the photo industry, DxO DeepPRIME delivers an improvement of around two ISO stops.

DxO’s massive image database was used to train DxO DeepPRIME. Thanks to billions of carefully selected input and output samples, the artificial intelligence ‘learned’ the two fundamental operations of any RAW photo editing process: demosaicing and denoising." DeepPRIME - L'IA rĂ©volutionne le dĂ©bruitage des images

“DeepPRIME removes noise and recovers information without losing detail, creating more natural transitions and textures. The results are spectacular, and as detailed below, the introduction of DeepPRIME XD pushes the boundaries even further. Compared to conventional noise reduction technology, DeepPRIME delivers the equivalent of an extra two stops of ISO. With DeepPRIME XD, it can be 2.5 stops, sometimes even more.”

Bayer moire
On the analysis of RAW converters
https://www.libraw.org/articles/bayer-moire.html

4 Likes

I personally prefer to take pictures of things more interesting than test charts, and I would wager that DxO trained their ML algorithms on things other than test charts, too.

I am consistently happy with the results, and the ease of achieving them, from PhotoLab (since v3) and not so with Lightroom on both counts.

6 Likes

Comparing de-noising algorithms on a 100 ISO image seems strange to me.

3 Likes

DeepPrime and DeepPrimeXD are also demoseicing alghoritims, so that makes sense, and I always use them even at base ISO because they provide better results. But I agree, using in-camera JPEG processing and the legacy Prime algorithm, which is only denoising, does not make much sense.

Anyone interested investing in a software platform today has the luxury of downloading fully functional RAW processing/editing/converting software on a trial basis. I downloaded 4 programs and compared portrait, landscape and macro images taken on my Fuji X-Pro 2 with two different Fuji lenses. After enlarging the same sections of each photo, Deep Prime XD won out on the sharpest rendering, contrast and tonality in my opinion, which is the only one that counts
for me! It was a marked difference. I would suggest anyone do the same, and select whatever best meets their needs.

5 Likes

I Like the moiré pattern on his belly.
Images are heavely cropped and are at 100% zoom.

RAW

DEEPPRIME XD

I tried 5 denoising softwares on this image test from an overcast cloudy day.
No one could give what deepprime did in photolab.
I always have to reduce sharpening a lot (generally from -1 to 0). Here -0.33.

I don’t do reproduction photography, but when iso raise, deepprimeXD is the master.
And when iso is very low, HQ (or sometime no denoising at all) works very well for me.

3 Likes

I absolutely LOVE Deep PRIME XD. My Night photography, high ISO, shots are made near perfect! Doesn’t this technology work best when processing a Raw file? I think so. Anyway, my urban 16,000 ISO photography Raw photography is greatly blessed to receive a “treatment” from DxO Deep PRIME XD (Digital Camera ISO 16000! - Ed Ruth Photography Instructor in Bakersfield California). I think DxO PhotoLab Elite 6 is truly amazing. And I just discovered the power of the “Perspective,” tool! Where have I been?

2 Likes

DeepPRIME and DeepPRIME XD only work on RAW files.

2 Likes

It was obvious for me that it is a de-mosaicking algorithm, used on RAW files. I, too, use sometimes DP at low ISO. My surprise was that he used low ISO, where noise should be difficult to see, to test it.
I have made tests by comparing the same scene, taken at 100 ISO and 6400 ISO under difficult light condition(*). 100 ISO has been processed w/o NR, 6400 ISO with DP and DPXD. With the 100 ISO file as a reference, you can tell if what you see on the 6400 ISO processed images is the reality or some artefacts created by the NR algorithm.

(*) I used to download these files from the studio scene comparison toll at dpreview.com. Unfortunately, this site is closing on April 10th.

1 Like

Yes. I don’t know how I feel about DPreview though. The site was basically full of people with very low standards who got lot of gear to test from Amazon and basically well paid guaranteed job with no real incentive to improve so ever since Amazon bought them for their own selfish reasons, the quality of posting and testing has been pretty sub par and has no improved virtually at all for many years. Meanwhile there are lot of small independent creators who bust their butt trying to improve all the time, so I feel for the legacy of the DPreview website and its symbolism, but I don’t feel bad for people who might have to work for a living for a change.

That being said, the ability to download RAW files from various lenses and camera to test out was a very useful feature. I will miss that one, for sure.