DxO vs LR noise reduction

[DanK]

I did the first comparison of DxO prime NR with Lightroom's.

The shot was taken at ISO 25,600 with a Canon 5D Mark III. The image was well-exposed.

I imported it into LR with the default capture sharpening and the default color noise reduction. I corrected white balance. I then applied luminance NR, value 76, moving the NR contrast slider up to 50.

I opened the same image in DxO. Chrominance NR was set to the maximum by default, so I didn't change that. I set luminance NR at 90. I exported to LR as DNG and left the default settings alone. I corrected WB.


Flipping back and forth at 200%, these seemed about as similar as I could get. I examined the lines and numbers on the ruler and the smoothness of the ruler and the background. It was resting on glossy-surface mat board.

First, LR without cropping:



Now, the uncropped DxO:



Now an extreme crop of the LR, roughly similar to 2:1 on the screen:



And finally, a crop from the DxO image:



These seem very similar to me. The two images aren't identical--looking closely at the crops, you can see somewhat different flaws in the two corrected images. However, in practical terms, they look fairly similar to me.

What do you think?

[proseak]

You set luminance NR at 90 in DxO, I see. That's a very high figure; the default auto is 40, and I tend to dial it back to 20 for very high ISO.

Was that standard, or PRIME NR?

[DanK]

You set luminance NR at 90 in DxO, I see. That's a very high figure; the default auto is 40, and I tend to dial it back to 20 for very high ISO.

Was that standard, or PRIME NR?

Prime.

I deliberately set the level in both programs to nearly eliminate all signs of noise at 200%, rather than using a more modest amount that I might choose in practice. I had two reason. One, I wanted to push both to their extremes, to see how they can handle detail when severe noise reduction is needed to get rid of noise. Second, I thought that pushing both algorithms to the extreme would more likely show differences between them.

You may notice the color differs. The Lightroom colors are accurate--it was an aluminum ruler placed on white matboard (turned gray by exposure). I suspect that is because the initial rendering differs, and I didn't set white balance separately for the DxO edit in either package. I had set WB in LR for the LR edit, and I simply copied that temperature to the DxO DNG after importing it to LR.

[Black Pearl]

The WB in the DxO shots seems a fair bit off showing a pink/magenta cast?

All the testing aside isn't it fantastic that we can shoot at these ISO's and get a usable image - anyone remember Konica 3200

[Dave Humphries]

Hi Dan,

Looking to compare for noise and detail, I have ignored the uncropped images.

They are pretty similar, but on balance, I see more detail surviving in the LR version, particularly on the majority of the horizontal ruler markings. When the figures are studied closely while switching between, then some of the numbers are better on DxO and others on LR.

When I look at the granular surface behind the mat board, there's more sharp detail in LR and no mistake.

There's no dark area in the crops to compare noise reduction effectively - and it cannot be judged from the full frame versions (for these purposes*) - so perhaps if we saw 100 or 200% crop of the black book spines, then a more valid assessment could be made of NR - and it doesn't matter that they are not sharply focused, indeed that's a benefit.

Going by the quite bright area of mat card, DxO seems to have less structure/contrast to the noise, but given what I said about the granular area, that's to be expected.

The more noise is filtered out, the softer an image is likely to be and vice versa.

Overall, my view is that with a slight adjustment to one or other, you could probably make them even more indistinguishable.

Therefore it probably comes down to personal preference as to which people get on with best.

Cheers, Dave

* normally of course, the full DISPLAY size image is bearing the results of NR, so that is what is relevant, not what we see pixel peeping during PP.

[NorthernFocus]

Judging by the cropped sample the DxO image seems to have higher contrast edges but has not eliminated the noise as well as the LR version. With a bit of sharpening on the LR version they may be tit for tat. But I'm not sure that I see enough of a difference to warrant adding yet another piece of software to the kit.

I don't use the noise reduction in LR at all but in fairness haven't tried it in LR6. Not sure if the standalone version is the same as cloud version or not. Now I'll have to give it a look I suppose

[DanK]

I have edited this to add a new test that changes my conclusions.

Robin: re the color cast, see the post immediately before yours. I addressed this in the new test below.

Re the uncropped image: Dave, we're on the same page about this. I just included them to show that at the resolutions we normally use, there isn't any appreciable difference between these two software packages, at least in this one test. You have to dig to see the differences.

It was very hard to make the two comparable, and others of you might have stopped at different values. I tried looking at three variables: edge clarity, high-frequency noise, and low-frequency noise (the blobs that NR can leave behind). They don't change at the same rates in the two packages. If I used one variable as the criterion, the two would be different on the others. So, I tried to balance all three, but that is a subjective judgment.

Dan wrote

But I'm not sure that I see enough of a difference to warrant adding yet another piece of software to the kit.

That's exactly what I am trying to evaluate for myself. There is a very large time cost involved in changing software. I happen to be using LR, and I will stay with it unless I see a real advantage to adding something else. Basically, this is the same as your calculation, just inverted. The second test below may persuade me, but I am not sure yet.

Dave, you wrote:

There's no dark area in the crops to compare noise reduction effectively - and it cannot be judged from the full frame versions (for these purposes*) - so perhaps if we saw 100 or 200% crop of the black book spines, then a more valid assessment could be made of NR - and it doesn't matter that they are not sharply focused, indeed that's a benefit.

I mostly disagree, although I do plan one other test to address the part of this with which I do agree. Noise is randomly distributed across the frame. There isn't more noise in shadow areas; there is just less signal to mask it. Out-of-focus areas aren't very useful for testing NR. Most NR, pushed hard enough, will smooth out noise well. The issue is how much detail remains after you push it that far, and to ascertain that, you need an in-focus image. That's the reason I used the ruler and made it the plane of focus--it provides a good test of preserving detail. However, I do agree that dark areas are useful because it increasing brightness makes noise more apparent.

To address the shadow issue, I did much the same thing this morning with a similar image that was roughly 2.4 stops underexposed. I brightened it and then did my best with both packages. I did it quite quickly and probably could have done a bit better with both. Judge for yourself. it seems to me that DxO is clearly superior, at least when pixel-peeping. However, this is a VERY extreme case, more NR than I would almost ever use.

It's clear that the extreme NR affected WB on both. In the case of LR, it created a magenta cast, and doing a WB adjustment on the post-NR crop did not correct it properly. In the case of the DxO image, I applied WB in LR on import. That image appears to have a tan cast.

The settings:

DxO prime: luminance 90, chrominance 100 (which it set automatically). LR default color NR (25) applied at import, but I don't think this mattered.

LR: Luminance set to 100 (maximum), detail 50, contrast 58; color 50, detail 12, smoothness 50.

Here are the shots, both roughly 200%:

Lightroom:



DxO:

[Dave Humphries]

I don't think we do disagree really Dan,

The desire to see the darker, oof areas, was for exactly and only, the reasons you state they might be helpful for. Everything else had already been revealed by the crops you provided, well I suppose an area of mostly black with sharp white/grey/coloured text might be helpful.

As you say, these days, any NR system worth its salt, should detect a lack of HF image detail and consequently wind up the NR applied in those areas.

Like Dan, and yourself, I don't see anything here to make me change my ways.

Currently; I use Neat Image inside PS, because I don't use LR.
Previous tests with ACR NR wasn't great, but that was a few generations ago, so may be I need to revisit.

However, with the arrival of the new computer, I intend to start using LR (for better DAM mainly), so who knows, I might yet change my ways.
I'll obviously be getting used to a revised RAW workflow anyway, so might as well tack on NR - after I have learnt how it works.

Cheers,
Dave

PS, oh lookie, your post got bigger and added pictures while I was replying, no I'm not so sure about LR NR.

[DanK]

Dave,

I forgot to point out that my second test showed that one of my comments was wrong:

Most NR, pushed hard enough, will smooth out noise well.

The second case seemed to exceed LR's capability to do this.

I find LR to be an extremely comfortable starting point, and I don't intend to give it up. I think these two trials suggest that when moderate NR is needed, the two packages will be similar. The question for me is whether I will have enough extreme cases to make it worth getting another software package to use in those cases. Like Manfred, I try very hard to shoot in ways that make NR unnecessary, but there are cases when one can't avoid it.

I have occasionally used Nik Define, and perhaps I will try it with these images. The problem with Nik is that once you call up the filter, you are forced to work on TIFF files. I like the idea of doing NR at the beginning while retaining the entire raw file.

[DanK]

And for completeness, here is a very quick try with NIK, setting contrast correction almost to the max, 160. Not comparable to DxO, it seems to me, but again, your results may differ.

If anyone would like to play with this, I have placed the original raw file (the one underexposed by 2.4 stops) in my public dropbox folder here. Dropbox is discontinuing this service soon, but for now, you should be able to grab the file.

[proseak]

Version 9 is currently on a free offer here -

http://www.dxo.com/us/digitalcamerauk

This does have the PRIME NR, but it is much slower than the current version

I can't get your Dropbox link to work btw

[DanK]

Version 9 is currently on a free offer here -

http://www.dxo.com/us/digitalcamerauk

This does have the PRIME NR, but it is much slower than the current version

I can't get your Dropbox link to work btw

Sorry, it looks like the dropbox public sharing function doesn't work, even though it was supposed to until mid-March. Unfortunately, I don't have another ready way to share a 32 MB file.

You can get a 31-day free trial of the current version of DxO from their website. That's what I did to test it. But if the NR is the same--even if slower--the old free version would be fine for occasional use, I think. Thanks for the link.

[proseak]

Sometimes I upload files to DB, and share by saving the link; this gives anyone with the link the ability to download but not alter the file itself. I last used this las t week, so I think it still works...

[DanK]

Sometimes I upload files to DB, and share by saving the link; this gives anyone with the link the ability to download but not alter the file itself. I last used this las t week, so I think it still works...

I think I used the wrong link, created by windows rather than by dropbox. Try this. Sorry about the inconvenience