DxO DeepPrime versus Prime, results and controls

Austinian · Oct 22, 2020

Breakfastographer 2 said:
Austinian said:

Breakfastographer 2 said:

Austinian said:

Breakfastographer 2 said:

Thanks for adding the SOOC jpeg. I'm still curious about the untampered output from DxO (no PRIME, DeepPRIME or HQ denoising) compared to PRIME and DeepPRIME, but seeing the SOOC jpeg gives a first idea.

I would be doing my own tests if I currently had a machine that could run it, so I appreciate your efforts.

Click to expand...

I don't have a PRIME, but here's a no-NR image vs. DeepPrime. Heavily edited and shadows brightened.

No NR

DeepPrime

Bad as the half-size JPEG is, the full-size No NR TIFF looked worse.

Click to expand...

Thank you for posting this. What I take from this is that DeepPRIME is rather rigorous about ditching shadow detail in its quest to return a nice-looking image. So the positive is that it doesn't seem to want to hallucinate detail where there is none. On the other hand, it's rather decidedly cutting down on dynamic range.

I can't help but wonder if that's a big part of how it works - just ditch lots of shadow detail... as a result, the well-lit parts look super crisp while the shadows look a bit like an oil painting.

I have to confess this one section has me in awe, and a little suspicious (sample reproduced in the name of expediency - hope this is acceptable):

Click to expand...

Sure, glad to see some skepticism; I find some of what DeepPrime does suspiciously good.

The image was from Carlsbad in 2015 with a APS-C camera; one exposure from what was originally a 5-shot HDR series, -2 EV from what the AE wanted.

BTW, that colored lighting no longer exists.

RANT ON.

the last time I was there the park had replaced it with boring, near-tungsten-like LEDs! They told me this was 'more natural'. WTH is "natural" in a cave hundreds of feet below the surface, except total darkness! Grrr! The old lighting was beautiful.

RANT OFF.

Click to expand...

The green does make it look more like a pirate cave.

Yeah, that one big rock looks a bit like a giant skull too.

Breakfastographer 2 said:
Who doesn't like pirates? I guess we know the answer now - people who live in Carlsbad! j/k

Breakfastographer 2 said:

Breakfastographer 2 said:

[ATTACH alt=""Before" on the left, probably obvious!"]2596385[/ATTACH]
"Before" on the left, probably obvious!

Were they really able to get that detailed texture back without "help"? Hmm...

Click to expand...

All editing was done on the entire image, with and without NR as the last step.

I did clone out a couple of blown areas near the middle large, bright stalagmite, and also a number plaque and information sign where the right-side railing makes a sharp left turn.

I hope this is at least plausible. :-D

Click to expand...

Right, I wasn't doubting your hand in it, just wondering at how much texture detail DeepPRIME managed to pull back from a lot of green and red dots. It would be easy to think the AI pulled some more info from one of the other, better lit walls and used this to reconstruct the area in question, similar to what context dependent fill does.

Of course, I'm hoping it didn't do that and it's really just using the local data very efficiently.

That's a very good question. When "AI" is involved, there's often no traceability of exactly how the AI arrived at the final result. So no way to tell AFAIK.

Maybe the only way to be happy with AI like this is a thing I saw on TV in a very different, NSFW context: "We know it's fake, but we don't care."

Tootsall · Oct 22, 2020

I ran a 6 set batch of photos through four times. No noise reduction, 5.6 sec average. HQ reduction, default settings, 6.0 seconds. Prime, default, 42.6 sec. Deep Prime, default, 35.3 sec. These on a late 2017 build iMac 27, 8 GB ram, i5 processor & 3.2 GHz. Shots from a Nikon Z50 with Sigma 17-50 f2.8 lens in raw files ~21MB each. I would think folks with newer computers built for photo processing could get a lot faster.

Just to remind...you don’t have to process every shot in Prime or Deep Prime...only those that need it due to light and ISO.

--
(Ed)

Sagittarius · Oct 22, 2020

Downloaded trial version of PL4 and ran an image I always check noise reduction with. This image from Nikon D500 at 51,600 ISO.

Original

Denoise AI

DXO Deep Prime

DXO Deep Prime + Denoise AI

IMO DXO + Denoise looks the best. What do you think.

--
Best regards

NAwlins Contrarian · Oct 22, 2020

There was a request that I post a version with no noise reduction, premised in part on the concern that DeepPrime was somehow generating fake detail in the tree near the top right of the crop (https://www.dpreview.com/forums/post/64483304). I decided to expand the comparison from two versions to four, adding not only noise reduction turned off but also DxO HQ noise reduction, which is the least sophisticated / powerful DxO noise reduction (and the only type available for non-raw files or in the less-expensive "Essential" version of PL). Here the results are presented in three different ways, so hopefully everyone can find the preferred way of viewing them. Again, these are all 100% crops from a file that I had previously processed to taste in PhotoLab 3, using heavy noise reduction, and that I merely opened the raw + sidecar combination in PhotoLab 4 and changed only the type of noise reduction (none, HQ, Prime, or DeepPrime).

This is the two new versions.

Without noise reduction, there's so much noise that I have no idea how you evaluate what is 'real' detail in the tree just to the right of the stop sign. Remember, click on "original size" and view at 100%.

This is the two versions previously presented.

And for those of you who want to view all four versions together and have a high enough screen resolution to do so:

And here's the quad with all four versions in one file, entirely viewable at 100% on a QHD or higher-resolution monitor.

By all means, draw your own preliminary conclusions, understanding that this is an extreme case, and after my initial processing in PL 3 with Prime noise reduction, I did not try to tweak the noise reduction settings for best results. But my take is that this is what you'd expect.

Van Griff · Oct 22, 2020

DXO Deep Prime

DXO Deep Prime + Denoise AI

IMO DXO + Denoise looks the best. What do you think.

S

I'm going to go with . . . DxODP . . . it has a little more grit to it than the version with DAI added which I think is slightly too soft or smooth.

Great example and awesome performance by DxODP!

Thanks for sharing!

Best,

V G

Jacques Cornell · Oct 22, 2020

AgoP5 said:
And here hte results with my desktop.

Here with DxO Photolab v3

Here with v4, by default set to Auto. As you can see this way DeepPRIME is unable to process the image.

Here with v4 set to the GPU, an AMD RX 550 with 4GB of VRAM.

The difference in time it's not much but it's there.

The difference in image quality is way better!!!

Interesting. I wasn't aware one could select CPU or GPU for processing. Here's another data point:

2013 Mac Pro 3.3GHz 8-core, 64GB RAM, 1TB SSD, dual FirePro D500 GPUs exporting from GX7 16MP ISO 12,800 RAW file, DeepPrime processing

Auto = 28 seconds
CPU = 25 seconds
GPU = 25 seconds

--
"I learned long ago, never to wrestle with a pig. You get dirty, and besides, the pig likes it." - George Bernard Shaw

Landscape & Wedding Photographer - Jacques Cornell

Jacques Cornell Photography - landscape, travel, portrait and art reproduction

jacquescornell.photography

Happening Photos - Event Photography in New York City

Event Photography in New York City and Around the World - conferences, sales meetings, fundraisers and more.

happening.photos

Breakfastographer 2 · Oct 23, 2020

Sagittarius said:
Downloaded trial version of PL4 and ran an image I always check noise reduction with. This image from Nikon D500 at 51,600 ISO.

Original

Denoise AI

DXO Deep Prime

DXO Deep Prime + Denoise AI

IMO DXO + Denoise looks the best. What do you think.

Similar to Austinian's example, there seems to be less tonality in the lion's eyes. But now I don't know if that's DxO doing it, or DeepPRIME.

--
https://breakfastographer.wordpress.com

Breakfastographer 2 · Oct 23, 2020

NAwlins Contrarian said:
There was a request that I post a version with no noise reduction, premised in part on the concern that DeepPrime was somehow generating fake detail in the tree near the top right of the crop (https://www.dpreview.com/forums/post/64483304). I decided to expand the comparison from two versions to four, adding not only noise reduction turned off but also DxO HQ noise reduction, which is the least sophisticated / powerful DxO noise reduction (and the only type available for non-raw files or in the less-expensive "Essential" version of PL). Here the results are presented in three different ways, so hopefully everyone can find the preferred way of viewing them. Again, these are all 100% crops from a file that I had previously processed to taste in PhotoLab 3, using heavy noise reduction, and that I merely opened the raw + sidecar combination in PhotoLab 4 and changed only the type of noise reduction (none, HQ, Prime, or DeepPrime).

This is the two new versions.

Without noise reduction, there's so much noise that I have no idea how you evaluate what is 'real' detail in the tree just to the right of the stop sign. Remember, click on "original size" and view at 100%.

This is the two versions previously presented.

And for those of you who want to view all four versions together and have a high enough screen resolution to do so:

And here's the quad with all four versions in one file, entirely viewable at 100% on a QHD or higher-resolution monitor.

By all means, draw your own preliminary conclusions, understanding that this is an extreme case, and after my initial processing in PL 3 with Prime noise reduction, I did not try to tweak the noise reduction settings for best results. But my take is that this is what you'd expect.

Thanks for posting the additional renderings. As you say, it really is an extreme case, and that makes it hard to tell what is going on. I think I see some of the recovered detail in the original, but I'm also seeing some detail in the original that PL4 recovered differently from how my brain is recovering it.

I feel slightly better about it.

--
https://breakfastographer.wordpress.com

Raist3d · Oct 23, 2020

I think when doing multiple passes of machine learning algorithms there's kind of a "style" that emerges from using them, and I think the double AI pass using both apps starts to look a bit plasticky. I what seems to be a few artifacts in the ceiling in upper right even if it looks "cleaner."

Also worth trying deepPrime + increase luminance noise reduction and then make a higher radius /stronger unsharp mask directly on Dxo. In one image it helped a lot for web size.

Raist3d · Oct 23, 2020

STOP sign area. Look at the StopSign and the mesh immediate to its left.

DXO deepPrime shows a clear superiority here.

onlyfreeman · Oct 23, 2020

Looks like I need DeepPrime, probably the first new software in ages that I think I might actually want.

I just want the DeepPrime function however, unfortunately it's not standalone.

NAwlins Contrarian · Oct 23, 2020

Member said:
STOP sign area. Look at the StopSign and the mesh immediate to its left.

DXO deepPrime shows a clear superiority here.

Although I agree the stop sign is a good place to compare, IMO one of the more telling elements of the first photo for comparing noise reduction types is the HC monogram on the left hip of the defender in the foreground on the left side of the larger crop from the photo.

100% crops: no noise reduction, DxO HQ, DxO Prime, and DxO DeepPrime, respectively

CrisL · Oct 24, 2020

In case we still need to be convinced we can look at this video from PhotoJoseph Don't known if it was already posted but I post it because this is exactly what I am seeing with my pictures

Digital Nigel · Oct 24, 2020

CrisL said:
In case we still need to be convinced we can look at this video from PhotoJoseph Don't known if it was already posted but I post it because this is exactly what I am seeing with my pictures

Thanks, that's a very informative, succinct video.

It strikes me that all the changes are surprises. Despite all the speculation and wish lists, DxO seems to have been able to maintain an effective embargo. I certainly didn't guess what would be in PL 4. For, example, I thought there might be improvements to the local adjustments, or NIK features incorporated, but I don't think either happened.

Another thought: now that PL 4 makes Topaz DeNoise AI almost redundant (I don't think I'll be paying for enhanced versions of it), will PL 5 do the same to Sharpen AI?

CrisL · Oct 24, 2020

Digital Nigel said:
CrisL said:

In case we still need to be convinced we can look at this video from PhotoJoseph Don't known if it was already posted but I post it because this is exactly what I am seeing with my pictures

Click to expand...

Thanks, that's a very informative, succinct video.

It strikes me that all the changes are surprises. Despite all the speculation and wish lists, DxO seems to have been able to maintain an effective embargo. I certainly didn't guess what would be in PL 4. For, example, I thought there might be improvements to the local adjustments, or NIK features incorporated, but I don't think either happened.

Another thought: now that PL 4 makes Topaz DeNoise AI almost redundant (I don't think I'll be paying for enhanced versions of it), will PL 5 do the same to Sharpen AI?

Trying to compare Topaz AI D and DeepPRIME I am finding that DeepPrime achieve the same level of denoising than Topaz AI D low light mode without introducing the colors artifacts. It looks we can push the luminance slider without loosing too much details . So it looks me very remarkable for their first version and quite promising for the future.

Digital Nigel · Oct 24, 2020

CrisL said:
Digital Nigel said:

CrisL said:

In case we still need to be convinced we can look at this video from PhotoJoseph Don't known if it was already posted but I post it because this is exactly what I am seeing with my pictures

Click to expand...

Thanks, that's a very informative, succinct video.

It strikes me that all the changes are surprises. Despite all the speculation and wish lists, DxO seems to have been able to maintain an effective embargo. I certainly didn't guess what would be in PL 4. For, example, I thought there might be improvements to the local adjustments, or NIK features incorporated, but I don't think either happened.

Another thought: now that PL 4 makes Topaz DeNoise AI almost redundant (I don't think I'll be paying for enhanced versions of it), will PL 5 do the same to Sharpen AI?

Click to expand...

Trying to compare Topaz AI D and DeepPRIME I am finding that DeepPrime achieve the same level of denoising than Topaz AI D low light mode without introducing the colors artifacts. It looks we can push the luminance slider without loosing too much details . So it looks me very remarkable for their first version and quite promising for the future.

DxO, no doubt, spent a lot of time evaluating DeNoise AI. Given that PRIME was one of the key reasons for buying PhotoLab, it was in danger of losing a lot of sales if DeNoise was deemed to be better.

So DxO needed to either beat it, or partner with/acquire Topaz. I doubt that DxO has the financial resources to acquire Topaz, so its primary objective would have been to beat DeNoise AI, obviously by taking full advantage of the GPU (something DxO hasn't done much before). That's why I suspect it'll try and beat Sharpen AI in a future release.

Sagittarius · Oct 24, 2020

I do not think Topaz will seat still, they will further develop Denoise AI and Sharpen AI. Competition is always good for customers.

NicholasD · Oct 24, 2020

poppyjk said:
Here is an ISO 51200 comparison of SOOC, DeepPrime plus Topaz Denoise AI, DeepPrime alone with other adjustments to taste.

75% crop, Canon R5 (in crop mode), 1/800, f5, ISO 51200, 227mm(100-400 II). You can see from the settings I'm not an experienced bird photog.

I have frequently used Topaz DeNoise AI to 'polish' a DxO Prime NR photo. DeepPrime alone is awesome and the DeNoise AI 'polish' makes an ISO 51200 photo actually quite usable.

(snip)

Interesting - had never considered pairing Topaz w/ PL. If you don't mind, can you expand on preferred workflow approach, situations that work well (and those that don't), etc. I have PL3 (Prime) and Topaz AI. Topaz was a recent acquisition and still learning the program, although there's not much to learn.... I use LR as my primary editor, only using DxO and Topaz for noise reduction on select images. Haven't decided on Deep Prime - looks promising but taking my time to make a purchasing decision. So appreciate you (and others) taking the time to post comparisons. I sometimes "polish" images in LR (i.e. basic edits in LR, run out to PL for noise reduction, and then return to LR for sharpening and any final WB changes etc.) I find LR luminance NR to be fine and sometimes I'll limit Prime to chrominance NR only.

The thought of relying on DxO for the heavy lifting and then Topaz for fine tuning is intriguing - any observations appreciated - thanks.

IanYorke · Oct 24, 2020

NicholasD said:
poppyjk said:

Here is an ISO 51200 comparison of SOOC, DeepPrime plus Topaz Denoise AI, DeepPrime alone with other adjustments to taste.

75% crop, Canon R5 (in crop mode), 1/800, f5, ISO 51200, 227mm(100-400 II). You can see from the settings I'm not an experienced bird photog.

I have frequently used Topaz DeNoise AI to 'polish' a DxO Prime NR photo. DeepPrime alone is awesome and the DeNoise AI 'polish' makes an ISO 51200 photo actually quite usable.

Click to expand...

(snip)

Interesting - had never considered pairing Topaz w/ PL. If you don't mind, can you expand on preferred workflow approach, situations that work well (and those that don't), etc. I have PL3 (Prime) and Topaz AI. Topaz was a recent acquisition and still learning the program, although there's not much to learn.... I use LR as my primary editor, only using DxO and Topaz for noise reduction on select images. Haven't decided on Deep Prime - looks promising but taking my time to make a purchasing decision. So appreciate you (and others) taking the time to post comparisons. I sometimes "polish" images in LR (i.e. basic edits in LR, run out to PL for noise reduction, and then return to LR for sharpening and any final WB changes etc.) I find LR luminance NR to be fine and sometimes I'll limit Prime to chrominance NR only.

The thought of relying on DxO for the heavy lifting and then Topaz for fine tuning is intriguing - any observations appreciated - thanks.

Polishing PRIME results with Topaz has become a common technique. It certainly looks like with DeepPrime this will not be necessary. Using Topaz Sharpen AI will still be viable.

I used LR from V1 until V6 when I switched to C1 and haven't looked back

That is personal choice, if you rely on DAM or need all-in-one, stick with LR, DXO has some useful dam features but is not a comprehensive dam like LR. DAM not significant to me so none issue, but as I say personal choice.

You can use your LR DCP profiles with DXO if you wanted to.The DXO local editing in layers is so much better than the brush and pin UI in LR. LR uses the same layer system but hides the UI because the lead engineer was "anything but PS". That's why the crop works the opposite way round to PS and every other program

Choices are good

Ian

Ian

Digital Nigel · Oct 24, 2020

NicholasD said:
poppyjk said:

Here is an ISO 51200 comparison of SOOC, DeepPrime plus Topaz Denoise AI, DeepPrime alone with other adjustments to taste.

75% crop, Canon R5 (in crop mode), 1/800, f5, ISO 51200, 227mm(100-400 II). You can see from the settings I'm not an experienced bird photog.

I have frequently used Topaz DeNoise AI to 'polish' a DxO Prime NR photo. DeepPrime alone is awesome and the DeNoise AI 'polish' makes an ISO 51200 photo actually quite usable.

Click to expand...

(snip)

Interesting - had never considered pairing Topaz w/ PL. If you don't mind, can you expand on preferred workflow approach, situations that work well (and those that don't), etc. I have PL3 (Prime) and Topaz AI. Topaz was a recent acquisition and still learning the program, although there's not much to learn.... I use LR as my primary editor, only using DxO and Topaz for noise reduction on select images. Haven't decided on Deep Prime - looks promising but taking my time to make a purchasing decision. So appreciate you (and others) taking the time to post comparisons. I sometimes "polish" images in LR (i.e. basic edits in LR, run out to PL for noise reduction, and then return to LR for sharpening and any final WB changes etc.) I find LR luminance NR to be fine and sometimes I'll limit Prime to chrominance NR only.

The thought of relying on DxO for the heavy lifting and then Topaz for fine tuning is intriguing - any observations appreciated - thanks.

It's often been discussed here, and many of us came to the conclusion that DeNoise AI was a good way to 'polish' some of the noisiest images, after PRIME. The two together were better than either alone, but if you only used one, PRIME was the better choice. So I use PRIME on every image, but fewer than 1% of them need the further DNAI polish.

This is a test I did almost two years ago, so now somewhat dated:

https://www.flickr.com/photos/recliner/albums/72157708128725115

DxO DeepPrime versus Prime, results and controls

Community Leader

Senior Member

Veteran Member

Forum Pro

Senior Member

Forum Pro

Leading Member

Leading Member

Raist3d

Guest

Raist3d

Guest

Senior Member

Forum Pro

Senior Member

Forum Pro

Senior Member

Forum Pro

Veteran Member

Senior Member

Veteran Member

Forum Pro

Keyboard shortcuts