lokatz
Veteran Member
Noise Reduction (NR) frequently pops up as a subject here and in other forums. Quite a few claims popped up in relation to this subject, such as
None of these statements line up well with my previous experience, but such discussions spiked my interest. Noise comparisons are found all over the internet, but there don’t seem to be folks who are as interested in assessing the effects of NR, which to me is critical. As a bird shooter, I often have to accept high ISOs, sometimes VERY high ones.
To learn something myself, and to test the above theories (if they can be called that), I spent a few hours conducting a real-world test and analyzing the results. I shot the same subject, at a distance of about 25 ft / 8 meters, with my Z8 and Z 800 f/6.3 on a sturdy tripod, varying shutter speeds and ISOs such that the exposure as metered by the camera remained the same. The light wasn’t great but also not horrible.
Next, I processed these images, using DxO PhotoLab 8.2.1, in three different ways:
The following gives you lots of details and plenty of comparison shots. If you are the impatient type and only want my bottom line, feel free to jump straight to the Conclusions section, but you will quite a bit.
.
Target
While the image of the target, a tin bird shot through a window, may not look like much, I find it well suited for the purposes of my testing: painted on the “bird” are finely detailed white feather-like structures that make denoising errors and artifacts stand out, plus a relatively dark background lets image noise stand out more clearly. (Example shots of the full target are found near the bottom of this post.)
How bad is the noise at higher ISOs? Well, all of the following images are small crops of 1000x1000 pixels from the “bird’s” head. The below series shows the image noise for ISO settings ranging from 100 to 51200, in full-step increments.
Image noise becomes visible pretty soon. While the differences between IS0100 and ISO200 are minimal, we already start seeing a little noise at ISO400. Not a real issue, neither here nor at ISO800, when looking at the full image, but for me, ISO800 is the highest I would go without applying some NR. As ISO levels keep rising, say somewhere between ISO12800 and ISO25600, increasing color noise starts to cause visible discolorations, too. None of this is new or should come as a surprise. I am only including these images for reference.
.
How much of a difference does NR make? And are DxO’s standard NR settings fine, or should I modify them?
In order to answer both of these questions, I put together a series of images, below, where I show the unedited image on the left, a noise-reduced version in the middle where I applied DxO’s default settings (Luminance set to 40, no modification of advanced settings), and an optimized noise-reduced version on the right. The latter received custom settings that depend on the ISO level, which I explain further below.
I am skipping ISO levels 100, 200, 800 and 3200 here, since little can be learned from looking at them: the lower two show no differences between unedited and edited versions, the higher two simply follow the pattern the remaining images demonstrate anyway.
At ISO400, what little noise was contained in the original shot is already gone after applying NR with standard settings. I cannot find any impact on image details whatsoever, so NR did a fine job here.
At ISO1600, we start seeing differences. Image noise, now clearly visible on the unedited shot, is mostly gone even with standard NR settings, but using custom settings further reduces the noise floor while not really affecting image details. (Settings: The image goes from a bit too noisy to just about perfect. DeepPRIME XD/XD2s seems to apply some light sharpening, as well, which helps ‘polish’ the appearance of the image.
ISO6400 is where things get really interesting: the original shot now has plenty of noise, including stronger color noise. With its standard setting, DxO reduces that noise quite a bit and largely eliminates color noise. However, noise is still clearly visible in the denoised image. With custom settings applied, the NR engine shows remarkable improvements. The price for that are minute losses in detail, but those are hard to see even in this 100% view, and completely gone when comparing full images. If you’re unsure which details I am talking about, compare the optimized ISO400 and ISO6400 shots and look at the minute spider web at the top of the beak, for instance, or the lint standing out from the back of the bird, close to the right edge of the image.
The trend continues as ISOs go up further. At ISO12800, even the optimized NR now leaves a little visible noise, though in my opinion, what remains is fully tolerable. The loss of fine details, such as the minute spider webs, also the color detail on the wing, is now a bit more obvious, but the white facial streaks that mimic feathers still maintain good detail and look almost identical, with optimized NR, as they do in the ISO400 shot. In fact, I was quite surprised myself at how good the resulting image still looks, given the poor quality of the unedited image. Not something I’d use often, but I would use it without hesitation if the light dictates it.
ISO25600 is where it becomes obvious, at least in this 100% view, that too much is too much. The “feather detail” (the white streaks) now looks fuzzy, and further artifacts appear in the small details. If there is no other option, I might still try to salvage an ISO25600 shot, say of a rare bird, this way, but this is something I wouldn’t want to showcase anywhere. At ISO51200, this becomes even more pronounced: the whole image looks washed out and now suffers not only from poor details but also from poor coloration.
Speaking of coloration, by the way, compare the ISO400 and ISO25600 shots: color differences, quite pronounced in the noise of the unedited shots, are small, usually barely perceptible, after NR optimization. This used to be far worse with older non-AI-powered denoising engines, where image colors tended to suffer quite a bit, so AI-powered NR obviously offers significant improvements in this field.
Bottom line, optimizing the NR engine’s settings, at least with shots at, say, ISO3200 or higher, is worthwhile. With such images, DxO’s default settings look less-than-ideal in my example shots. I cannot think of any reason why this would be different when shooting other subjects.
So what are those ominous “custom settings” I used to optimize the effects of NR? In a nutshell, this is what I did:
How many stops of improvement do you get with optimized denoising?
That’s the Million Dollar Question, isn’t it? Rather than merely sharing my opinion, I would like to invite you to participate in a little test. Please look at the four pairs of images below and decide which one of them you find to be the most similar. Yes, I know, there are differences in noise levels, details, artifacts etc., but try to judge them on their overall merit: if forced to decide in favor of one or the other, where would that decision be the hardest, with the pair in the first, second, third or fourth row?
Made your choice? Well, this might tell you something about what NR can do for you. The choice is subjective, as it involves trade-offs: some of us find noise more off-putting, while others may prefer to maintain the greatest level of detail. If deciding between two similar images gets hard, that means their overall image quality is about equal.
Time to fill you in: the image on the left is always the same, namely the unedited shot taken at ISO400. The shots on the right, from top to bottom, were taken at 1600, 3200, 6400 and 12800, respectively. These all received optimized NR.
The row I personally struggled with the most when trying to make up my mind is the third one. Yes, the image on the right misses small details, though I’d argue that none of them are visible in full image view anyway, but image noise to me looks quite a bit better than in the image on the left. (To be fair, some of you might argue that none of that noise in the left image will be visible in full image view, either ;-) ). If you also found Row 3 to be the hardest choice, you and I agree that the benefit of optimized NR, as defined here, equals about four stops (ISO 6400 versus ISO400). Others may find Row 2 closer to parity, but I doubt many will pick Rows 1 or 4. This means that most of us find optimized denoising to yield about a three-to-four-stop gain in overall image quality. Quite an improvement, and proof that the old two-stop rule is no longer valid when using state-of-the-art NR.
.
Full view versus pixel peeping
I know, I know: most of us look at images in 100% view when judging image sharpness, noise, and other aspects. That’s why I did the same here. Keep in mind, however, that doing so counts as pixel peeping: we use electronic means to look at what we would not see, at least not nearly as clearly, when looking at the full image. Whether you print your images and proudly display them in the hallway, or show them to your friends on a large monitor, they want to look at the whole image, not 100% details. Some image details simply won’t be visible then unless you print VERY large, so they can be safely ignored if that is not the case.
What I used in the above discussion are crop views of much-larger images. I therefore also want to discuss how these images look when you compare full images.
At full image view, I find it completely impossible to determine differences in image details between the optimized NR results on my 4K monitor unless ISOs are 25600 or more. Expectably, some dynamic range and coloration differences might show up at lower ISOs, say, at ISO6400 and above, but those are small and can mostly be equalized in post. In fact, even 25600 still looks fine to me when looking at the full image. Here, I said it: ISO25600 still produces decent images when not pixel peeping.
Judge for yourself, but please DON’T pixel peep = refrain from looking at the 100% view: what differences do you see that indicate at what respective ISOs these four images were recorded? (They range from ISO100 to ISO25600, all with optimized NR applied. There are small lighting differences, but those are not indicative of ISO levels.) Can you clearly identify the best and the worst one? Which ones do you find usable or unusable, and why? Interested in your opinions.
Image 1
Image 2
Image 3
Image 4
(Later in this thread, I will reveal which is which.)
.
Conclusions
These are the conclusions I drew from my analysis. Yours may be different, as some of this boils down to personal preferences.
I plan to conduct this same test with my OM-1ii MFT body, which might also offer a good foundation for a meaningful real-world FF vs MFT comparison.
.
One final request
PLEASE refrain from responding to this post by using “Reply with Quote”. NOBODY wants to see or read all of this again. If you refer to specific parts of my post, please cut out everything else from the quote. Thank you.
- “NR improves an image by 2 stops”
- “All NR introduces artifacts that are easy to see”
- “I can always detect if an image has been subjected to NR, even at full view and normal viewing distance”
None of these statements line up well with my previous experience, but such discussions spiked my interest. Noise comparisons are found all over the internet, but there don’t seem to be folks who are as interested in assessing the effects of NR, which to me is critical. As a bird shooter, I often have to accept high ISOs, sometimes VERY high ones.
To learn something myself, and to test the above theories (if they can be called that), I spent a few hours conducting a real-world test and analyzing the results. I shot the same subject, at a distance of about 25 ft / 8 meters, with my Z8 and Z 800 f/6.3 on a sturdy tripod, varying shutter speeds and ISOs such that the exposure as metered by the camera remained the same. The light wasn’t great but also not horrible.
Next, I processed these images, using DxO PhotoLab 8.2.1, in three different ways:
- Applying no editing whatsoever, I merely converted them to JPGs.
- Using the default setting, I applied NR, then converted them to JPGs without any other edits.
- I applied NR while optimizing the settings for each image, then converted them to JPGs without any other edits.
The following gives you lots of details and plenty of comparison shots. If you are the impatient type and only want my bottom line, feel free to jump straight to the Conclusions section, but you will quite a bit.
.
Target
While the image of the target, a tin bird shot through a window, may not look like much, I find it well suited for the purposes of my testing: painted on the “bird” are finely detailed white feather-like structures that make denoising errors and artifacts stand out, plus a relatively dark background lets image noise stand out more clearly. (Example shots of the full target are found near the bottom of this post.)
How bad is the noise at higher ISOs? Well, all of the following images are small crops of 1000x1000 pixels from the “bird’s” head. The below series shows the image noise for ISO settings ranging from 100 to 51200, in full-step increments.
Image noise becomes visible pretty soon. While the differences between IS0100 and ISO200 are minimal, we already start seeing a little noise at ISO400. Not a real issue, neither here nor at ISO800, when looking at the full image, but for me, ISO800 is the highest I would go without applying some NR. As ISO levels keep rising, say somewhere between ISO12800 and ISO25600, increasing color noise starts to cause visible discolorations, too. None of this is new or should come as a surprise. I am only including these images for reference.
.
How much of a difference does NR make? And are DxO’s standard NR settings fine, or should I modify them?
In order to answer both of these questions, I put together a series of images, below, where I show the unedited image on the left, a noise-reduced version in the middle where I applied DxO’s default settings (Luminance set to 40, no modification of advanced settings), and an optimized noise-reduced version on the right. The latter received custom settings that depend on the ISO level, which I explain further below.
I am skipping ISO levels 100, 200, 800 and 3200 here, since little can be learned from looking at them: the lower two show no differences between unedited and edited versions, the higher two simply follow the pattern the remaining images demonstrate anyway.
At ISO400, what little noise was contained in the original shot is already gone after applying NR with standard settings. I cannot find any impact on image details whatsoever, so NR did a fine job here.
At ISO1600, we start seeing differences. Image noise, now clearly visible on the unedited shot, is mostly gone even with standard NR settings, but using custom settings further reduces the noise floor while not really affecting image details. (Settings: The image goes from a bit too noisy to just about perfect. DeepPRIME XD/XD2s seems to apply some light sharpening, as well, which helps ‘polish’ the appearance of the image.
ISO6400 is where things get really interesting: the original shot now has plenty of noise, including stronger color noise. With its standard setting, DxO reduces that noise quite a bit and largely eliminates color noise. However, noise is still clearly visible in the denoised image. With custom settings applied, the NR engine shows remarkable improvements. The price for that are minute losses in detail, but those are hard to see even in this 100% view, and completely gone when comparing full images. If you’re unsure which details I am talking about, compare the optimized ISO400 and ISO6400 shots and look at the minute spider web at the top of the beak, for instance, or the lint standing out from the back of the bird, close to the right edge of the image.
The trend continues as ISOs go up further. At ISO12800, even the optimized NR now leaves a little visible noise, though in my opinion, what remains is fully tolerable. The loss of fine details, such as the minute spider webs, also the color detail on the wing, is now a bit more obvious, but the white facial streaks that mimic feathers still maintain good detail and look almost identical, with optimized NR, as they do in the ISO400 shot. In fact, I was quite surprised myself at how good the resulting image still looks, given the poor quality of the unedited image. Not something I’d use often, but I would use it without hesitation if the light dictates it.
ISO25600 is where it becomes obvious, at least in this 100% view, that too much is too much. The “feather detail” (the white streaks) now looks fuzzy, and further artifacts appear in the small details. If there is no other option, I might still try to salvage an ISO25600 shot, say of a rare bird, this way, but this is something I wouldn’t want to showcase anywhere. At ISO51200, this becomes even more pronounced: the whole image looks washed out and now suffers not only from poor details but also from poor coloration.
Speaking of coloration, by the way, compare the ISO400 and ISO25600 shots: color differences, quite pronounced in the noise of the unedited shots, are small, usually barely perceptible, after NR optimization. This used to be far worse with older non-AI-powered denoising engines, where image colors tended to suffer quite a bit, so AI-powered NR obviously offers significant improvements in this field.
Bottom line, optimizing the NR engine’s settings, at least with shots at, say, ISO3200 or higher, is worthwhile. With such images, DxO’s default settings look less-than-ideal in my example shots. I cannot think of any reason why this would be different when shooting other subjects.
So what are those ominous “custom settings” I used to optimize the effects of NR? In a nutshell, this is what I did:
- Increased the ‘Luminance’ setting, keeping it at the default of 40 for low-ISO images but gradually raising it for higher ones. For instance, I used values of 53 at ISO6400, 61 at ISO25600 and 73 at ISO51200. This had the biggest impact.
- At higher ISO levels, I increased the ‘Dead pixels’ and ‘Force details’ sliders to their maximum of 100 each. In general, this seemed to have little impact, but it did introduce slight improvements at ISO12800 and above.
How many stops of improvement do you get with optimized denoising?
That’s the Million Dollar Question, isn’t it? Rather than merely sharing my opinion, I would like to invite you to participate in a little test. Please look at the four pairs of images below and decide which one of them you find to be the most similar. Yes, I know, there are differences in noise levels, details, artifacts etc., but try to judge them on their overall merit: if forced to decide in favor of one or the other, where would that decision be the hardest, with the pair in the first, second, third or fourth row?
Made your choice? Well, this might tell you something about what NR can do for you. The choice is subjective, as it involves trade-offs: some of us find noise more off-putting, while others may prefer to maintain the greatest level of detail. If deciding between two similar images gets hard, that means their overall image quality is about equal.
Time to fill you in: the image on the left is always the same, namely the unedited shot taken at ISO400. The shots on the right, from top to bottom, were taken at 1600, 3200, 6400 and 12800, respectively. These all received optimized NR.
The row I personally struggled with the most when trying to make up my mind is the third one. Yes, the image on the right misses small details, though I’d argue that none of them are visible in full image view anyway, but image noise to me looks quite a bit better than in the image on the left. (To be fair, some of you might argue that none of that noise in the left image will be visible in full image view, either ;-) ). If you also found Row 3 to be the hardest choice, you and I agree that the benefit of optimized NR, as defined here, equals about four stops (ISO 6400 versus ISO400). Others may find Row 2 closer to parity, but I doubt many will pick Rows 1 or 4. This means that most of us find optimized denoising to yield about a three-to-four-stop gain in overall image quality. Quite an improvement, and proof that the old two-stop rule is no longer valid when using state-of-the-art NR.
.
Full view versus pixel peeping
I know, I know: most of us look at images in 100% view when judging image sharpness, noise, and other aspects. That’s why I did the same here. Keep in mind, however, that doing so counts as pixel peeping: we use electronic means to look at what we would not see, at least not nearly as clearly, when looking at the full image. Whether you print your images and proudly display them in the hallway, or show them to your friends on a large monitor, they want to look at the whole image, not 100% details. Some image details simply won’t be visible then unless you print VERY large, so they can be safely ignored if that is not the case.
What I used in the above discussion are crop views of much-larger images. I therefore also want to discuss how these images look when you compare full images.
At full image view, I find it completely impossible to determine differences in image details between the optimized NR results on my 4K monitor unless ISOs are 25600 or more. Expectably, some dynamic range and coloration differences might show up at lower ISOs, say, at ISO6400 and above, but those are small and can mostly be equalized in post. In fact, even 25600 still looks fine to me when looking at the full image. Here, I said it: ISO25600 still produces decent images when not pixel peeping.
Judge for yourself, but please DON’T pixel peep = refrain from looking at the 100% view: what differences do you see that indicate at what respective ISOs these four images were recorded? (They range from ISO100 to ISO25600, all with optimized NR applied. There are small lighting differences, but those are not indicative of ISO levels.) Can you clearly identify the best and the worst one? Which ones do you find usable or unusable, and why? Interested in your opinions.
Image 1
Image 2
Image 3
Image 4
(Later in this thread, I will reveal which is which.)
.
Conclusions
These are the conclusions I drew from my analysis. Yours may be different, as some of this boils down to personal preferences.
- Applying noise reduction is worthwhile at most ISO levels. It is almost a must-do when shooting at levels higher than ISO800.
- My testing strongly hints at that old rule that “NR improves an image by 2 stops” being, well, old. In my view and when weighing all aspects (remaining noise, artifacts, discoloration, …), DxO buys you 4 stops or more when comparing full images, slightly less than that when pixel peeping. If you shoot at ISO3200 or below, your images will look just as good, after proper NR, as if they were shot at ISO200. That’s good news for action and low-light shooters.
- When using DxO, there is an incentive for customizing the NR engine’s settings, at least if your shots have ISO values of 3200 or higher.
I plan to conduct this same test with my OM-1ii MFT body, which might also offer a good foundation for a meaningful real-world FF vs MFT comparison.
.
One final request
PLEASE refrain from responding to this post by using “Reply with Quote”. NOBODY wants to see or read all of this again. If you refer to specific parts of my post, please cut out everything else from the quote. Thank you.
Last edited:
