Blog

As you can see from my pre­vi­ous arti­cles, I have been using DxO Pho­to­Lab since ver­sion 4 as a Light­room CC plu­g­in for noise reduc­tion of my high-ISO images. The AI-based Deep­PRIME noise reduc­tion process intro­duced with DxO Pho­to­lab 4 deliv­ers much bet­ter results than the denois­ing algo­rithms imple­ment­ed in Adobe Light­room for high-ISO images with a lot of noise. Deep­PRIME also per­formed bet­ter in my com­par­i­son with the com­peti­tor Topaz DeNoise AI.

Since then, the DxO Pho­to­Lab Light­room plu­g­in ( cur­rent­ly in ver­sion 5) has become an inte­gral part of my work­flow for demand­ing high-ISO shots. How­ev­er, I do not use the oth­er exten­sive func­tions of Pho­to­Lab 6 (so far). All fur­ther pro­cess­ing (expo­sure cor­rec­tion, col­or cor­rec­tions, crop­ping, tag­ging, export) after noise reduc­tion with DxO Deep­PRIME is still done in Adobe Light­room. I par­tic­u­lar­ly enjoy using Light­room’s exten­sive cat­a­log fea­tures. My cat­a­log dates back to 2003 and con­tains more than 200,000 images.

There­fore, in this arti­cle I will deal exclu­sive­ly with the aspect of noise reduc­tion with the Pho­to­Lab 6 plugin.

DxO has added the advanced denois­ing process Deep­PRIME XD (XD stands for “eXtreme Detail”) to the new Pho­to­Lab 6 Elite edi­tion. Accord­ing to DxO, the results should be much bet­ter than before.

For Deep­PRIME XD, DxO says it made use of a new neur­al net­work that has been trained with bil­lions of image sam­ples and is now said to allow improve­ments of more than 2.5 f-stops. This means that after pro­cess­ing with DxO Deep­PRIME XP, an image with ISO 12,800 will have about as much noise as one with ISO 2,000-2,500 before. In my own test, DxO Pho­to­Lab 4’s Deep­PRIME algo­rithm already showed an improve­ment of about 3 f-stops with respect to the vis­i­ble image noise of high-ISO shots.

Since Deep­PRIME XD is an inte­gral part of the RAW con­ver­sion process, it should also be able to fur­ther increase the dynam­ic range. How­ev­er, the orig­i­nal Deep­PRIME mode remains avail­able in Pho­to­Lab 6. It is intend­ed for faster opti­miza­tions or for images that sim­ply require few­er cor­rec­tions. Unfor­tu­nate­ly, this sug­gests that the new algo­rithm will be even more elab­o­rate and require even more pro­cess­ing pow­er. I was able to ver­i­fy this, in my exam­ples Deep­PRIME XP need­ed about 1.5 times as much time as Deep­PRIME, but this is still with­in rea­son­able lim­its - at least if your com­put­er hard­ware is suf­fi­cient­ly potent.

Oth­er changes in DxO Pho­to­Lab 6 con­cern col­or man­age­ment, retouch­ing, and pho­to library man­age­ment. How­ev­er, these are of no rel­e­vance to me for the time being, since, as already men­tioned in my review of ver­sion 4, I only use Pho­to­Lab as a Light­room plug-in.

Being ini­tial­ly very skep­ti­cal whether the already excel­lent result of DxO’s Deep­PRIME could be fur­ther improved, I imme­di­ate­ly down­loaded the tri­al ver­sion - and after a short test, imme­di­ate­ly pur­chased the update:

DxO DeepPRIME XD has actually gotten even better

For the test I again used the pho­to of our tom­cat Tom, which I already used for the test of Deep­PRIME in Pho­to­Lab 4. The pic­ture rep­re­sents a real endurance test for the noise reduc­tion. On the one hand it was tak­en with ISO 12,800 (I lim­it­ed the Auto-ISO range of my cam­era to 100-12,800) and was still under­ex­posed by a f-stop, so that it is basi­cal­ly a shot at ISO 25,600!

Image editing

As usu­al, I first processed the image in Light­room (Expo­sure +1.05, Con­trast +60, High­lights -100, Depths +100) and denoised it (Details|Luminance 60). The oth­er set­tings remained at their defaults. Then I trans­ferred the CR3 file to Pho­to­Lab 6 and processed it with the default set­tings (though lumi­nance was set to 75, which gave me the best results) first with the Deep­PRIME algo­rithm known from the pre­vi­ous ver­sions and then with the new Deep­PRIME XD method. Then the images were passed back to Light­room Clas­sic as DNG files with­out any fur­ther changes using the Export to Light­room func­tion in Pho­to­Lab 6.

The pro­cess­ing time for the image of our tom­cat on my lap­top (i9-11980 HK, Geforce RTX 3080/8GB, 64GB main mem­o­ry) was a bit increased with the new algorithm:

The per­for­mance of the Deep­PRIME algo­rithms seems to depend heav­i­ly on the graph­ics card. Final­ly, I test­ed this again on my lap­top (i9-11980 HK, GeForce RTX 3080/8GB, 64GB main mem­o­ry), this time with a 48 megapix­el image from my DJI Mini 3 Pro. To ana­lyze it, I let DxO Pho­to­Lab 6 ren­der the image first using the GeForce GPU and then the CPU.

With the GeForce 3080 lap­top GPU/OpenCL turned on:
- Deep­PRIME: 0:22min
- Deep­PRIME XD: 0:22min (real­ly this time no dif­fer­ence)

With the CPU ( run­ning with 8 cores at 3.5GHz):
- Deep­PRIME: 1:39min
- Deep­PRIME XD: 9:47min (almost 6times slower)

Since the DNG files cre­at­ed by DxO Pho­to­Lab 6 looked a bit dark­er and had a min­i­mal magen­ta col­or cast after re-import­ing them into Light­room, I com­pen­sat­ed this in Light­room with Expo­sure +0.5 and Tint -4. The oth­er set­tings were applied as spec­i­fied in the Light­room pro­cess­ing above (the noise reduction/luminance was set to 0, of course).

Below is an overview of the images cre­at­ed in this way:

Admit­ted­ly, in the resized for the web overview images, the dif­fer­ences are hard­ly notice­able. Reduc­ing the size of images is already a very effec­tive way of reduc­ing noise. Any­one who only presents his pic­tures in small size on the web there­fore does­n’t real­ly need any denoising 😊.

It is dif­fer­ent if the images are intend­ed to be print­ed or dis­played on a larg­er scale as well. Then the dif­fer­ence becomes obvi­ous. I have there­fore cre­at­ed a few more crops from the image above in the fol­low­ing, which demon­strate the dif­fer­ences much more impressively.

DeepPRIME

I’ll start with a detail of the right eye of our tom­cat. I have enlarged it to 200% for bet­ter vis­i­bil­i­ty. Here is a com­par­i­son of the devel­op­ment in Light­room Clas­sic 11.5 (left) with the addi­tion­al denois­ing using the old­er Deep­PRIME method (right):

Next, anoth­er detail of our cat’s chin:

DxO Deep­PRIME’s noise reduc­tion is already much bet­ter com­pared to the result obtained with Light­room’s built-in func­tion, as I already found out in my review of DxO Pho­to­Lab 4. How­ev­er, some fin­er details seem to be lost in the pro­cess­ing, which can just be sus­pect­ed in the noisy ver­sion devel­oped in Light­room. This can be seen in par­tic­u­lar in the low­er shot at the base of the throat. The area looks some­what unnat­u­ral­ly washed out.

But now to the actu­al­ly inter­est­ing topic:

Comparison between DeepPRIME and DeepPRIME XD

How do these areas now look in com­par­i­son between the famil­iar Deep­PRIME and the new Deep­PRIME XD? For this pur­pose, I demon­strate below the two sec­tions com­par­ing the two meth­ods. The image on the left has been denoised using the above set­tings in the DxO Pho­to­Lab 6 plug-in with Deep­PRIME, while the one on the right has been denoised using the new Deep­PRIME XD. At first, the eye:

And here is the detail of the chin of our cat:

In the crops, you can now clear­ly see that the new algo­rithm actu­al­ly reveals even more details. In the pre­vi­ous­ly emp­ty areas at the base of the neck, hair details are now vis­i­ble again. In the eye, the pat­tern of the iris is clear­ly more detailed. The dif­fer­ences are cer­tain­ly small­er than those between Light­room and the old­er Deep­PRIME process, but they are still vis­i­ble and lead to a more nat­ur­al-look­ing appear­ance of the image.

File size vs noise

The bet­ter noise reduc­tion of Deep­PRIME XP com­pared to the pre­vi­ous Deep­PRIME ver­sion can also be esti­mat­ed from anoth­er val­ue, name­ly the file size.

Cur­rent cam­eras per­form image com­pres­sion even on RAW files. With my Canon cam­eras and also many oth­ers this is loss­less, i.e. no infor­ma­tion is lost. This is sim­i­lar to .ZIP files.

Of course, the com­pres­sion rate depends very much on the sub­ject. Large uni­form areas (e.g. the blue sky) can be com­pressed pret­ty well. On the oth­er hand, a lot of details lead to low­er com­pres­sion rates. Noise, how­ev­er, is par­tic­u­lar­ly dif­fi­cult to com­press. The extreme case, the so-called “gray noise”, can­not be com­pressed at all.

The poor com­press­ibil­i­ty of the noise also means that RAW files from our cam­eras with high ISO val­ues are larg­er than files with low ISO val­ues. Thus, the file size also gives an indi­ca­tion of the degree of noise.

The rule is:

• low noise - small file
• strong noise - large file

So I took anoth­er look at the file sizes of the above exam­ple image of our tom­cat in File Explor­er where I found the fol­low­ing file sizes:

Orig­i­nal CR3 file	57.993 KB
Deep­PRIME DNG file	126.286 KB
Deep­PRIME XD-DNG file	111.653 KB

Gen­er­al­ly, the DNG files export­ed with Pho­to­Lab are always sig­nif­i­cant­ly larg­er than the orig­i­nal RAW files. But the file cre­at­ed with the new Deep­PRIME XD process is a good 12% small­er than the file cre­at­ed with the old­er Deep­PRIME process. My oth­er com­par­isons of the two process­es show a sim­i­lar ratio. In each case, the Deep­PRIME XD file is 10-12% small­er than the Deep­PRIME file gen­er­at­ed from the same orig­i­nal, which actu­al­ly gives anoth­er mea­sur­able indi­ca­tion of the improved noise reduc­tion with the new process.

But now back to the top­ic and the prac­ti­cal results:

Example 2

Aber es geht noch extremer:

Dur­ing our safari in Kenya in Decem­ber 2021, we had noticed some move­ment in the grass in half dark­ness one evening on the way back to our camp in the Masai Mara and stopped. It turned out to be a ser­val walk­ing around in the dim­ness. With the naked eye it could be seen only very schemat­i­cal­ly. Since I still had my equip­ment ready (accord­ing to the safari slo­gan “always be pre­pared”) I was able to take some shots of the cat with my 400mm f/2.8 on the R5.

It was already very dark, so the fol­low­ing shot was tak­en at 1/125s, f/2.8 and ISO 12,800. It was still under­ex­posed by 2 f-stops. I have cal­cu­lat­ed it: this cor­re­sponds to an light­ing inten­si­ty of 2-3 lux, i.e. the light of 2-3 can­dles at a dis­tance of 1m!

First of all, here is the RAW file devel­oped in Light­room Clas­sic WITHOUT any denoising:

This does­n’t look bad at all, does it?

In the down­sized image this is indeed the case, but the 100% crop reveals the bru­tal truth. Below on the left is the image in Light­room Clas­sic with­out any denois­ing and on the right with mas­sive noise reduc­tion (lumi­nance 70, col­or 60):

The noise reduc­tion in Light­room is quite effec­tive at remov­ing the grainy noise, but unfor­tu­nate­ly it also removes the details. So the image as a whole is unus­able for any­thing except great­ly reduced pre­sen­ta­tion on the web. But you don’t real­ly need it for that either, as the overview image above shows. But there’s still DxO Deep­PRIME. Here’s a com­par­i­son of the image above, this time devel­oped with DxO Deep­PRIME (left) and Deep­PRIME XD (right):

Yes, this is real­ly the same shot! It’s amaz­ing what DxO Deep­PRIME can get out of that image. And the new Deep­PRIME XD is actu­al­ly even better!

Below is a more illus­tra­tive com­par­i­son of the unprocessed image with the image treat­ed with Deep­PRIME XD:

The fol­low­ing is a direct com­par­i­son between Deep­PRIME and Deep­PRIME XD:

Prices and availability

The ESSENTIAL and ELITE edi­tions of DxO Pho­to­Lab 6 (Win­dows and macOS) are now avail­able for down­load* from the DxO web­site at the fol­low­ing prices:

• DxO Pho­to­Lab 6 ESSENTIAL Edi­tion: 139 €
• DxO Pho­to­Lab 6 ELITE Edi­tion: 219 €

The Deep­PRIME XD algo­rithm test­ed here is unfor­tu­nate­ly only avail­able in the more expen­sive Elite ver­sion. How­ev­er, own­ers of DxO Pho­to­Lab 4 or 5 can take advan­tage of a spe­cial upgrade price:

• Upgrade price for DxO Pho­to­Lab 6 ESSENTIAL Edi­tion: €75
• Upgrade price for DxO Pho­to­Lab 6 ELITE Edi­tion: €99

If you would like to test Pho­to­Lab 6 for your­self, you can down­load the tri­al ver­sion here* now and try out all the func­tions for 30 days with­out any restrictions.

Resumée

For me, the result of my tests caused me to buy the update imme­di­ate­ly, despite the rather high price of €99. Still, DxO’s Deep­PRIME algo­rithm is, in my opin­ion, excep­tion­al­ly good. But good things can indeed get even bet­ter. Sub­jec­tive­ly, I esti­mate the improve­ment of the new ver­sion at about anoth­er f-stop.

Admit­ted­ly, the exam­ples shown here are absolute extremes. I have already pho­tographed inten­sive­ly in ana­log film days where all ISO val­ues above 400 were large­ly unus­able. The extreme was B/W films pushed to ISO 1600, where the coarse grain struc­ture was pre­sent­ed as an aes­thet­ic fea­ture. Apart from the large film grains, details were hard­ly recognizable.

With the cur­rent dig­i­tal cam­eras, on the oth­er hand, ISO val­ues up to 1,600 can now be used with­out any prob­lems. Thanks to sophis­ti­cat­ed soft­ware - and here I’m cur­rent­ly think­ing of DxO in par­tic­u­lar - the lim­its have now been pushed even fur­ther. Those who already use Deep­PRIME in Pho­to­Lab 4 /5 or Pur­eRAW 1/2 will have to decide for them­selves whether the improve­ments of Deep­PRIME XD com­pared to the pre­vi­ous ver­sion are worth the rather high surcharge.

I have decid­ed to do this…

If you don’t use Deep­PRIME or Deep­PRIME XD yet and you fre­quent­ly use high ISO val­ues, I strong­ly rec­om­mend you to try the tri­al ver­sion for yourself.

I look for­ward to com­ments and feed­back from you.

*=affil­i­ate link