Blog

On Sep­tem­ber 17th, 2024, DxO intro­duced the lat­est ver­sion 8 of their image pro­cess­ing soft­ware Pho­to­Lab 8*. I already use DxO Pho­to­Lab for years to remove noise from very noisy high-ISO images. For all fur­ther pro­cess­ing of my images (which I shoot exclu­sive­ly in RAW for­mat), how­ev­er, I have been using Adobe Light­room and occa­sion­al­ly Adobe Pho­to­shop since its release in 2007.

So what inter­ests me most about the new Pho­to­Lab 8 ver­sion is the qual­i­ty of the inte­grat­ed and sup­pos­ed­ly fur­ther improved AI-sup­port­ed denois­ing tool, Deep­PRIME, which is now avail­able in ver­sion XD2s.

The fol­low­ing arti­cle deals exclu­sive­ly with the results of the new Deep­PRIME XD2s algo­rithm. I have only briefly tried out the oth­er Pho­to­Lab func­tions. Pho­to­Lab 8 is very com­pre­hen­sive and pow­er­ful, but I don’t pre­sume to have a well-found­ed opin­ion on it.

DeepPRIME

I’ve been using DxO Deep­PRIME, the AI-based denois­ing tech­nol­o­gy first intro­duced in DxO Pho­to­Lab 4, since 2020. I test­ed Pho­to­Lab 4 exten­sive­ly back then and was so impressed by the incred­i­bly effec­tive denois­ing that I bought the pro­gram imme­di­ate­ly afterwards.

As I have been edit­ing, tag­ging and archiv­ing my images in Adobe Light­room for many years, how­ev­er, I have inter­nal­ized the process­es so well that a com­plete switch was not an option for me. Over a peri­od of more than 20 years, I have already archived a good 260,000 images in my Light­room data­base. I have described my cur­rent way of work­ing with Light­room in detail here.

I have there­fore always used Pho­to­Lab via the includ­ed Light­room plug-in.

DeepPRIME XD

AI-based denois­ing has been fur­ther devel­oped by DxO over the years. In 2022, with Pho­to­Lab 6, the new ver­sion Deep­PRIME XD (XD stands for eXtreme Details) was released. After an exten­sive test, I real­ized that this ver­sion could cre­ate images with about one f-stop less noise, so I updat­ed to Pho­to­Lab 6 and Deep­PRIME XD.

How­ev­er, after that Adobe was not idle either. For the first time in ver­sion 12.3, a new­ly devel­oped AI-sup­port­ed denois­ing func­tion was inte­grat­ed into Light­room Clas­sic. I also com­pared the new Light­room denois­ing with that of DxO in detail. This result­ed in a slight advan­tage for DxO. How­ev­er, Deep­PRIME was able to process the RAW files of my DJI drones much bet­ter than the AI-sup­port­ed process inte­grat­ed in Lightroom.

Since then, I have used both Light­room and Deep­PRIME XD to denoise my very noisy images, depend­ing on the sub­ject, where the seam­less inte­gra­tion into Light­room made it eas­i­er to use. How­ev­er, DxO’s Deep­PRIME XD remained reserved for the “hard cas­es” and my DJI drone shots.

I skipped ver­sion 7 of Pho­to­Lab because the Deep­PRIME-XD denois­ing that I use exclu­sive­ly was left unchanged.

Note:
In 2021, DxO also released a slimmed-down ver­sion of Pho­to­Lab called Pur­eRAW* for users who only need PhotoLab’s denois­ing func­tion, like me. Since ver­sion 2, Pur­eRAW also includes a Light­room plu­g­in, but it offers few­er options for denois­ing than Pho­to­Lab. I there­fore con­tin­ue to use the plug-in of the full ver­sion DxO Pho­to­Lab. Pur­eRAW is cur­rent­ly avail­able in ver­sion 4 with Deep­PRIME XD2.

DeepPRIME XD2s

Now DxO has released anoth­er update of Pho­to­Lab, the new ver­sion 8, in which the AI-based denois­ing has been fur­ther improved, accord­ing to DxO. The new algo­rithm is now called Deep­PRIME XD2s and is appar­ent­ly a fur­ther devel­op­ment of Deep­PRIME XD2, which was first intro­duced in Pur­eRAW 4*. DxO itself writes about the new features:

Is Deep­PRIME XD2s now sig­nif­i­cant­ly bet­ter than the Deep­PRIME XD that I used up to now and how does it com­pare to the AI denois­ing in Light­room Classic?

So it’s once again time for a com­par­i­son test.

Comparison test

Once again, I used the very noisy RAW files from my pre­vi­ous com­par­i­son tests. The first test object was again the shot of our tom­cat Tom tak­en at ISO 12,800 with my Canon EOS R5. I was par­tic­u­lar­ly inter­est­ed in the com­par­i­son with the pre­vi­ous ver­sion Deep­PRIME XD and also with the new AI denois­ing func­tion avail­able direct­ly in Light­room Classic.

First of all, here you see the orig­i­nal image again, which has not been denoised. As it was still too dark despite ISO 12,800, I bright­ened it up by anoth­er f-stop in Light­room. So it actu­al­ly cor­re­sponds to a shot with ISO 25,600!

I’m sure we all agree that the pic­ture is use­less. So get rid of it?

The AI is already doing it!

Since the avail­abil­i­ty of AI-based denois­ing meth­ods, I no longer have any prob­lems work­ing with 5-dig­it ISO val­ues. The results that can be achieved in this way still amaze me to this day. Here is an exam­ple in direct com­par­i­son. On the left is the orig­i­nal with­out noise reduc­tion and on the right the ver­sion processed with DxO Deep­PRIME XD2s ( cropped at 200% magnification):

That’s real­ly very impres­sive, isn’t it?

But now to the most inter­est­ing com­par­isons of the dif­fer­ent pro­cess­ing meth­ods. I processed the orig­i­nal .CR3 RAW file from my Canon EOS R5 suc­ces­sive­ly with DxO Deep­PRIME XD, DxO Deep­PRIME XD2s and, for com­par­i­son, also with the in Light­room since ver­sion 12.3 inte­grat­ed AI denoising.

In Deep­Prime XD and XD2s I have set the lumi­nance val­ue to 75 and in Light­room I have also set the lev­el of denois­ing to 75. Both val­ues have worked well for me so far. I have deac­ti­vat­ed all the oth­er options (includ­ing the opti­cal cor­rec­tions) in Pho­to­Lab and will apply them after­wards in Light­room. Then I start­ed the denois­ing with the option “Export to Lightroom/Export as DNG”:

By the way, on my com­put­er (intel i9-11980 HK, GeForce RTX 3080/8GB, 64GB main mem­o­ry), denois­ing took around 15-20 sec­onds per image in all cas­es. A pow­er­ful graph­ics card is still impor­tant, with­out this the process takes sev­er­al min­utes for each image.

After­wards in Light­room I sim­ply adjust­ed the sharp­en­ing for all images to the default val­ue of 40 and acti­vat­ed the lens corrections:

The oth­er set­tings in Light­room were not changed. Here are the images in full view (this is already a por­trait crop from a land­scape image from my Canon EOS R5).

There are no sig­nif­i­cant dif­fer­ences to be seen in the down­scaled ver­sion with 2000 pix­els. In prin­ci­ple, a reduc­tion in size in itself already pro­vides effi­cient reduc­tion of noise. To be able to eval­u­ate the dif­fer­ences that still exist, you have to take a clos­er look. Below I there­fore present crops of the above images enlarged to 100%:

Well, to be hon­est, even at 100% mag­ni­fi­ca­tion, all the crops still look good, the dif­fer­ences are small. So to make the sub­tle dif­fer­ences more vis­i­ble, we need to get even clos­er. I’ve added more direct side-by-side com­par­isons at 200% mag­ni­fi­ca­tion below. The bar in the mid­dle of the pic­tures can be moved to the left and right so that indi­vid­ual areas of the image can be com­pared direct­ly with each other.

DeepPRIME XD vs XD2s

The most inter­est­ing ques­tion at the moment is whether the new Deep­PRIME XD2s deliv­ers even bet­ter results than Deep­PRIME XD. Let’s start by com­par­ing the two meth­ods with my default set­ting (lumi­nance 75, every­thing else switched off)

In my opin­ion, this direct com­par­i­son actu­al­ly shows a fur­ther improve­ment in the new XD2s process. With large­ly iden­ti­cal repro­duc­tion of detail, the noise (par­tic­u­lar­ly vis­i­ble in the area of the white chin at the bot­tom left and next to the cat’s left eye) could be removed even bet­ter. In addi­tion, the pre­vi­ous XD ver­sion appears some­what over­sharp­ened. How­ev­er, the dif­fer­ences are not very great - please also bear in mind that we are look­ing at a 200% enlarge­ment here. Nev­er­the­less, I like the new XD2s ver­sion here better.

DeepPRIME XD2s vs Lightroom KI

But if you are already using the cur­rent ver­sion of Light­room Clas­sic, as I am, how does it com­pare direct­ly with the AI-based denois­ing that has been inte­grat­ed into Light­room CC since ver­sion 12.3?

Here is the direct com­par­i­son in the 200% view:

Hon­est­ly, the dif­fer­ence here is very slight at first, in my opin­ion even less than that between Deep­PRIME XD and XD2s. To be able to see the dif­fer­ence in more detail, you have to zoom in even more. Here is a direct com­par­i­son in the 400% view in Light­room. On the left you can see the Light­room AI denois­ing, on the right the result with DxO Deep­PRIME XD2s:

On clos­er inspec­tion, Deep­PRIME XD2s reveals a lit­tle more detail. You can see this in our cat’s eye, for exam­ple. The fine details in the iris appear sharp­er and more contoured.

But is it worth the addi­tion­al invest­ment in DxO Deep­PRIME XD2s if you already own Lightroom?

More details

But it may be pos­si­ble to do even bet­ter. The DxO Denois­ing Tech­nolo­gies pan­el in Pho­to­Lab 8 also has a Force details slid­er that is sup­posed to bring out even more details in the image. I did anoth­er denois­ing in Pho­to­Lab 8 with the slid­er set to the max­i­mum 100.

Here is the comparison:

And here for even bet­ter clar­i­fi­ca­tion is anoth­er com­par­i­son of both sec­tions of the eye at 400% mag­ni­fi­ca­tion in the Light­room com­par­i­son view. On the left is again the Light­room AI denois­ing, on the right the result with DxO Deep­PRIME XD2s:

Here, the image denoised with Deep­PRIME XD2s with the option Force details: 100 shows sig­nif­i­cant­ly more details and appears sharp­er. How­ev­er, there are also some arti­facts, e.g. on the upper eye­lid. Ulti­mate­ly, this is a mat­ter of taste.

So would I buy Pho­to­Lab 8 today in addi­tion to Light­room for denois­ing? The results are bet­ter, but does the small over­all advan­tage jus­ti­fy the addi­tion­al investment?

Prob­a­bly not, at least not for my sys­tem cameras.

And other cameras?

How­ev­er, as I men­tioned at the begin­ning, the sit­u­a­tion is com­plete­ly dif­fer­ent with my DJI drones. One “prob­lem” with many cur­rent DJI drones is the 1 / 1.3 inch 48 megapix­el sen­sor installed in them. It has a so-called quad-bay­er design. This means that the col­or res­o­lu­tion is sig­nif­i­cant­ly reduced com­pared to its grayscale res­o­lu­tion. I have detailed this here in my test report on the DJI Mini 3 Pro. My DJI Mini 4Pro and also my DJI Air 3 are cur­rent­ly affected.

Unfor­tu­nate­ly, the AI noise reduc­tion inte­grat­ed in Light­room does not work very well with their DNG files. It pro­duces clear­ly vis­i­ble arti­facts. Here is an exam­ple of an image of the Are­nal vol­cano in Cos­ta Rica tak­en with my DJI Mini 4 Pro at an ISO val­ue of 400. Below is the orig­i­nal processed in Light­room with­out noise reduction:

Ich habe in Light­room dazu fol­gende Ein­stel­lun­gen vorgenommen:

The 100% crop reveals the sig­nif­i­cant noise of the rel­a­tive­ly small sensor:

Now back to the com­par­i­son shots of the denois­ing in Light­room and with Deep­PRIME XD2s. DxO Deep­PRIME XD2s per­forms much bet­ter here, even in the full image reduced to 2000 pixels.

The dif­fer­ences can be seen even more impres­sive­ly in the com­par­i­son view in Light­room, enlarged to 400%. On the left you can see the result of the Light­room AI denois­ing, on the right the one cre­at­ed with Deep­PRIME XD2s:

At present, I think the denois­ing in Light­room is unus­able for DJI drone images with the 48 megapix­el Quad Bay­er sen­sors. DxO deliv­ers way bet­ter results with Deep­PRIME XD2s, so I will con­tin­ue to denoise my DJI drone images with it.

After this detailed com­par­i­son, how­ev­er, I would also like to address some­thing very fun­da­men­tal on the sub­ject of AI denois­ing, which has been on my mind for some time now.

Fact or fake?

If you look at the orig­i­nal­ly mas­sive­ly noisy high-ISO images above, you nat­u­ral­ly won­der how the AI can achieve such a good result. Here you can see the com­par­i­son image from above again:

In con­trast to the algo­rith­mic denois­ing mech­a­nisms used in con­ven­tion­al image pro­cess­ing pro­grams, the way AI works is com­plete­ly dif­fer­ent. A com­plex neur­al net­work is trained with a vast num­ber of images. In prin­ci­ple, you can shoot the same sub­ject with low and high ISO val­ues and train the AI with the image pairs so that it learns that the denoised high-ISO ver­sion comes clos­est to the ver­sion with the low ISO val­ue. As you can see, this works very well.

Denois­ing reveals details that can­not be seen with the naked eye in the orig­i­nal even with the strongest intents. I there­fore ask myself more and more often whether the details vis­i­ble in the processed image are real­ly there, or whether the AI is just gen­er­at­ing struc­tures that it thinks are appro­pri­ate. If you are also inter­est­ed in this: I explored this ques­tion fur­ther in my sep­a­rate arti­cle AI denois­ing - fact or fake.

The bot­tom line: many of the details vis­i­ble in the denoised images were only invent­ed by the AI, although it is get­ting bet­ter and bet­ter at it.

And now?

With all the AI hype, I some­times ask myself whether this is still pho­tog­ra­phy. The term pho­tog­ra­phy is com­posed of the ancient Greek φῶς phōs, (“light”) and the ancient Greek γράφειν gráphein (“to draw”) and there­fore means “to draw with light”. A photographer’s mantra is also “It’s just the light - the light makes the pic­ture”. How­ev­er, light seems to be becom­ing less and less important.

Of course, cam­eras have always dis­tort­ed real­i­ty. Three-dimen­sion­al objects become two-dimen­sion­al. In dig­i­tal cam­eras, pix­el col­ors are inter­po­lat­ed by the col­or fil­ters of the neigh­bor­ing pix­els in the Bay­er matrix. Many algo­rithms auto­mat­i­cal­ly process the raw data from the sen­sor, and the image pro­cess­ing pro­gram does the rest. But these algo­rithms are com­pre­hen­si­bly defined and repro­ducible. With AI it is dif­fer­ent, AI is a “black box”.

I’m quite ambiva­lent about this myself. The top­ic extends far beyond denois­ing. In the post-pro­cess­ing of pho­tos, image areas can be replaced with AI based gen­er­at­ed fills, unwant­ed parts of the image can be removed, skies can be replaced and you can even cre­ate pho­to­re­al­is­tic images based sole­ly on text input.

As a self-con­fessed nerd, I have of course already tried all this out by myself and I admit - I am very impressed. But I’ve also lost some of the mag­ic and joy of the “craft of pho­tog­ra­phy”. Mas­ter­ing the tech­nol­o­gy is becom­ing less and less important.

When I think of the excit­ed feel­ing when an image slow­ly mate­ri­al­ized in the red light of the pho­to lab on the white pho­to paper in the devel­op­er bath - but I digress…

Resumée

Back to the topic:

With the lat­est AI-sup­port­ed denois­ing func­tion Deep­PRIME XD2s includ­ed in DxO Pho­to­Lab 8, DxO has achieved a fur­ther increase in qual­i­ty. How­ev­er, the improve­ments since the pre­vi­ous ver­sions have become small­er and users of Adobe Light­room from ver­sion 12.3 already have an alter­na­tive AI-sup­port­ed denois­ing option that also deliv­ers very good results - at least with my sys­tem cameras.

How­ev­er, it’s a dif­fer­ent sto­ry with some DJI drones in par­tic­u­lar. Here, Deep­PRIME denois­ing is clear­ly superior.

I will most like­ly con­tin­ue as before:

Depend­ing on the sub­ject, my very noisy shots are processed with both Light­room and Deep­PRIME XD, with Light­room being more stream­lined and quick­er to use. Deep­PRIME is still reserved for the “hard cas­es” and my DJI drone shots.

How­ev­er, if you don’t already use Light­room, DxO Pho­to­Lab is cer­tain­ly a good alter­na­tive. In addi­tion to excel­lent denois­ing, it offers a com­pa­ra­ble range of edit­ing options to Light­room and has the advan­tage that you don’t have to pay a subscription.

My recommendation:

• If you have no pre­vi­ous expe­ri­ence with Deep­PRIME, I strong­ly rec­om­mend that you down­load the free tri­al ver­sion of Pho­to­Lab 8 (direct­ly here direct­ly at DxO*) and try it out. DxO is extreme­ly gen­er­ous here and allows you to test its pro­grams for 30 days with­out any restrictions.

• For those who - like me - have been famil­iar with anoth­er RAW devel­op­ment pro­gram for many years, the slimmed-down Pur­eRAW 4 is a more cost-effec­tive and rea­son­able addi­tion. A free tri­al ver­sion can also be down­loaded here* and tried out for 30 days.

• How­ev­er, if you have not yet built up a large image data­base and are still com­plete­ly open in this respect, I rec­om­mend that you take a clos­er look at DxO Pho­to­Lab 8*. With this appli­ca­tion, you may have every­thing you need in one program.

• How­ev­er, if you are already using DxO Pho­to­Lab 6 or 7, you need to con­sid­er whether the minor improve­ments of Deep­PRIME XD2s jus­ti­fy an upgrade. Again, I rec­om­mend check­ing this for your­self with the free tri­al version.

Prices and availability

The ESSENTIAL and ELITE edi­tions of DxO Pho­to­Lab 8 (Win­dows and macOS) are now avail­able for down­load from the DxO web­site* at the fol­low­ing prices:

• DxO Pho­to­Lab 8 ESSENTIAL Edi­tion: 229 €
• DxO Pho­to­Lab 8 ELITE Edi­tion: 139 €

The Deep­PRIME XD2s algo­rithm test­ed here is unfor­tu­nate­ly only avail­able in the more expen­sive Elite ver­sion. How­ev­er, own­ers of DxO Pho­to­Lab 6 or 7 can take advan­tage of a spe­cial upgrade price:

• Upgrade to DxO Pho­to­Lab 8 ESSENTIAL Edi­tion: 75 €
• Upgrade to DxO Pho­to­Lab 8 ELITE Edi­tion: 99 €

Wer Pho­to­Lab 8 selb­st testen möchte, If you would like to test Pho­to­Lab 8 for your­self, you can down­load the tri­al ver­sion here* now and try out all func­tions for 30 days with­out any restrictions.

* = Affil­i­ate link