DXO Photolab 4 DeepPRIME vs. Topaz DeNoise AI vs. Lightroom Noise Reduction for Wildlife Photography
Unlike working with human subjects or in a studio, wildlife are unpredictable, and rarely are conditions ideal for getting the best quality photo possible. When photographing wildlife, there are a number of challenges that affect the quality of the final photograph:
1) the need to shoot in low-light when many wildlife are most active,
2) the lack of reach due to far-off subjects or “too short” lenses, which necessitates cropping to get a desirable composition
3) sometimes subjects are fast moving and dropping the shutter speed to compensate for low light is not an option
These issues, coupled with limited budgets with which to buy expensive lenses that fare best in low light conditions can only lead to one thing: noisy images. Early in my hobby I was quick to throw away images with noise, but with great improvements in noise reduction software technologies, reasonable improvements can be made to even the most noisy images.
Today I wanted to investigate three different methods of noise reduction for digital photos and answer a question that has been on my mind for a while: how do different post-production noise reduction techniques compare? Do the differences even matter? To do this, I compared the same photo with noise reduction applied using three different software methods: Topaz DeNoise AI, DXO Photolab 4 DeepPRIME, and Lightroom Noise Reduction. I have also posted a link to the full-size jpegs for those who might prefer to make their own comparisons. If you are viewing on mobile, I suggest going straight to those full size jpegs to review the images and using pinch to zoom to inspect closely. For the “too long, didn’t read” crowd, please feel free to jump straight to the results.
What is image noise and why would you want to remove it?
Every camera on the market produces digital noise, and every image taken has some amount of noise. Noise is a digital “grain”, or erroneous colors that show up at the pixel level. These can be caused by random emission of photons from light sources, or during the process of converting the analog signal of light hitting the sensor to digital pixels in-camera.
In simple terms, the weaker or noisier the signal, (i.e., quantity and quality of light photons reaching the sensor), the more noise becomes evident in pictures. Noise is less obvious when the signal is strong such as a brightly lit scene, however, when photos are taken in low light or with high ISO settings, noise becomes more noticeable. It is not that the noise necessarily becomes stronger, it is that the signal isn’t strong enough to overcome the underlying noise present in the image. By lifting the shadows in post or boosting ISO to achieve proper exposure for a dark scene, one is effectively lifting the baseline amount of noise that is present in the image in addition to pixels expressing the true signal. My experience has been that boosting ISO is generally a means to achieve proper exposure at a given shutter and aperture setting in-camera without resorting to post processing, not as a means to lower noise. I’ve found that raising the ISO in camera to achieve proper exposure versus increasing exposure in post achieves more or less the same result in most cases.
When noise becomes more obvious, it starts to become a nuisance. If the goal is to have a realistic looking image, noise can be detrimental by shrouding the image in a veil of multi-colored specks that start to obscure detail. Cropping an image also affects the apparent noise because the size of the noise pixels increases when cropping, making them more prominent.
Post-production noise reduction tools, especially the modern ones with “AI Technologies” promise to reduce the effects of noise while preserving detail. The holy grail of noise reduction would be to reduce or eliminate the apparent effects of digital noise whilst having no detrimental effect on detail. The reality is that too much noise reduction can make an equally unpleasing image by obscuring details or creating “over-baked” looking image.
Notwithstanding, these tools are useful because noise reduction can subjectively improve the final output in many cases. I set out to investigate the effects of three different noise reduction software and to compare the end result of these different approaches to noise reduction.
Please note that I am not paid or sponsored by any of these software companies and purchased licenses with my own funds.
Methods
For the purpose of my investigation, I wanted to choose the same photo for comparison which clearly has some apparent noise when viewed at 1:1, but is also a clean enough image that detail is visible, and it looks reasonably good at full size without any edits applied. For this test, I left all of the default settings in software alone (i.e., no edits) and simply changed the noise reduction settings or toggled them off completely. I then made some subjective comparisons below based on taste. I turned off sharpening on all of the programs to rule out this variable.
Note that for Topaz noise reduction I utilized Lightroom as the raw processor. DXO DeepPRIME requires the use of DXO PhotoLab as the raw processor.
EXIF info for the image used in this comparison is as follows: Canon EOS R5, ISO 2500, F/5.6, 1/250th @ 145mm.
Software versions tested are below.
Denoise AI, ver. 1.3.3, model Denoise AI
DXO ver. 4.3.2
LR ver. 10.4
Results
DXO PhotoLab 4 Deep Prime
In this test I imported the RAW image into DXO, left all settings at default and toggled DeepPrime noise reduction and applied the following settings:
DeepPrime Off
Luminance 40
Luminance 70
Luminance 100
Please make sure to click on the images for a closeup view, otherwise the differences won’t be readily obvious because the noisy pixels are so small. If you are on mobile, you’ll want to view the images directly here: full-size jpegs
Settings from Left to Right (Luminance): Off, 40, 70, 100. You can also download this image here.
Overall, I would say that DXO’s NR algorithm had the least deleterious impact on overall subjective image quality. Out of all the settings I tried, my preference was for Luminance 40 , which was the default setting in the software. It clearly took care of some of the nasty luminance noise but left a pleasing grain and photo that looks natural overall. With a luminance value of 100, I notice a very slight drop-off in fine detail without any appreciable improvement in the overall look of the photo, and to my eye it starts to look ever so slightly “chunky”. I would say that DXO yielded impressive result with this image with minimal effort, all I had to do was import the photo and turn noise reduction on. Only having one slider to toggle was also a bonus.
Topaz Denoise AI
Topaz has three sliders, including “Remove Noise” “Enhance Sharpness”, and “Recover Details”. Remove noise functions as described, removing noise - but it also seems to increase sharpness and boost detail - perhaps part of the AI algorithm. Enhance sharpness seems to function as a simple sharpness slider. Recover details seems to remove some of the chunky/blockiness that is evident when “Remove Noise” is applied, but at the consequence of adding a bit of noise back into the image. I actually prefer it when this setting is added back to the photo because it provides a more uniform grain as opposed to areas with and without apparent noise.
Settings used (Remove Noise - Enhance Sharpness - Recover Details):
NR off
15-15-0
75-15-75
100-15-0
Settings Left to Right (Remove Noise-Enhance Sharpness-Recover Details): 0; 15-15-0; 75-15-75; 100-15-0. You can also download this image here.
Lightroom NR
Lightroom NR has sliders for “Luminance”, “Detail”, and “Contrast”. The luminance slider reduces the amount of luminance noise in the image, but when it’s pushed to the max the resulting image looks like a mushy, blocky mess. The detail slider, which seems to work similarly to Topaz’s recover detail slider, attempts to add detail back into the image at the consequence of adding back a little noise, but unlike Topaz there are still blocky artifacts evident in the image. Maxing these sliders out makes for a rather unrealistic looking image in my opinion, and one would be well advised to use the LR sliders modestly. Contrast is supposed to add back contrast lost in the NR process, but in the case of this image it seems to have very little effect.
Settings used (Luminance-Detail-Contrast):
NR Off
20-50-0
30-50-0
100-100-100
Settings from Left to Right (Luminance-Detail-Contrast): 0; 20-50-0; 30-50-0; 100-100-100. You can also download this image here.
I generally found it was easy to get good noise reduction results with all three programs. That is to say the programs removed luminance noise while generally preserving detail, if used modestly. When I look through the images at full-size on my screen and start pixel-peeping, I personally gravitate towards the DXO images. Even at the strongest setting NR applied, DXO’s algorithm seems to avoid most of the chunky blockiness that is possible with the other two, and the resulting image looks (subjectively) more natural to me. That is not to say that great results aren’t possible with the other two, just that it is far easier to overshoot the mark and either destroy fine detail or produce a blocky/smeared look to the image. That being said, I am inclined to say less is more for all of these programs - it seems the best course of action is to dial in the minimum amount of NR that is necessary to reduce bothersome noise while retaining as much detail as possible. The following shows closeups with all of the images together:
Comparison of all three with respective slider settings. Use the pinch to zoom or CMD+ on your computer, or right click and “Save As” to your desktop to view more closely, if desired. You can also download this image here.
For an image like this where noise isn’t strongly apparent at normal viewing distances, it is perhaps unsurprising that noise reduction doesn’t make an big difference to the look of the image. However, when viewing a close crop, the benefit of noise reduction tools is obvious. To me, I can barely see a difference between any of these images at normal full-screen size (MacBook 16”), but when cropped down to 1:1 the noise is strongly apparent.
Please feel free to take a look at the full-size JPEGs and draw your own conclusions.
How about a Noisier Image?
While the above image shows what can be done with an image with modest amounts of noise, what about an image with more prominent noise? I decided to test this again using another image shown below.
From there I repeated the same test with the same settings that I used above.
In repeating the test with this noisier image, some things became readily apparent to me that weren’t as noticeable in the first image I tested:
Images from left to right with slider values used (generally less to more NR applied from left to right). Use the pinch to zoom or CMD+ on your computer, or right click and “Save As” to your desktop to view more closely, if desired. You can also download this image here.
Topaz
1) Default settings of 15,15,0 in Topaz led to undesirable, blocky patches with and without noise across the image.
2) Bumping up the “Recover Details” slider in Topaz resolved the issue above, at the cost of adding an acceptable amount of noise back into the image.
3) Maxing out the settings made the image start to look a little blocky, “chunky” and unrealistic.
DXO
1) Good results were obtained in the middle ranges between the default 40 and 70 slider values.
2) Maxing out the slider produces a reasonably good image but it begins to loose detail and look a little bit unrealistic.
3) Unlike the other two software its harder to ruin the image by overdoing the NR.
LR
1) Lower settings led to generally more favorable results, but the image remained noisier overall.
2) Higher settings start to destruct details more quickly than the other software
3) Maxing out settings obliterates detail and leads to a mushy image.
After testing this with other images, the above settings seem to be a good starting point for all types of images with noise issues. LR works reasonably but in general I found I could get better results out of both DXO and Topaz. These preferences are of course subjective and may vary from one viewer to the next.
Overall, my favorite results were observed with the following slider values:
Topaz (Remove Noise - Enhance Sharpness - Recover Details) : 75-15-75
DXO (Luminance): 40
LR: (Luminance, Detail, Contrast): 20-50-0
Links to the full-size jpegs can be found here: full-size JPEGs
Part III - Software Updates
A friendly forum-goer sent me a message and let me know that my Topaz version was out of date, so I quickly updated it to the latest version V3.2.0 and did the test over again to see if there was any change when the image was processed with Topaz.
After matching the settings with my original test, I ran it again and checked over the results - which were more or less identical.
Repeat of the test with the latest version of Topaz installed. You can also download this image here.
I then decided to check whether the other available models improved the results at all. Below are four images, top left: no NR applied; top right: 75-15-75, Denoise AI model; bottom left: Clear AI model (auto settings); bottom right: Low Noise model (auto settings).
Clear AI seems to make the grain even finer and less pleasing overall, Severe Noise model seems to have the same issue of patches of noisy and non-noisy areas unless you utilize the “recover details slider”.
In general, I found the best results to my eye so far with the Standard model set to the following slider settings N:75,S:15,R75,C:0 - the result was a uniform grain without weird artifacts (e.g. blobs of noisy areas and non-noisy areas) and a reasonable amount of noise reduction. I should add that the Comparison View provided in the newest Topaz Denoise AI version is really great way to identify the best model and settings for the image at hand.
Topaz utilizes different noise reduction models and the effect can be compared at the same time using comparison view.
Wrapping things up
Is noise reduction worth it? If you are viewing images at full-resolution, cropping minimally, and avoiding extreme ISOs on the camera, probably not. You won’t readily notice the differences. However if you do need to crop or are in a situation where you have to push the ISO to extreme levels, noise reduction can help.
This useful test on a personal level, because it shed light on a lot of questions I had about how the noise reduction compared between these tools, and how images respond to changes in the slider values. For example, maxing out settings in LR led to a great loss of detail and a "mushy and chunky" look, while modest settings looked pretty good. Leaving Topaz at the default settings led to blocky-looking, unfavorable artifacts with patches of noise and no noise, but the Recover Details slider fixes it and the result looks great. The DXO results are the most consistent, and the NR slider is generally the "safer" of the three - it's harder to really mess up the image even if you go overboard on the slider, and you start to see diminishing returns after a value of around 40. Overall, I found that when used carefully, both DXO and Topaz did a great job with noise reduction, retaining lots of detail in the photos even at 1:1. The LR image still looks great at normal viewing distances, but starts to look rough up close especially when heavy NR is applied.
Overall, I still think DXO DeepPRIME provides the most straightforward noise reduction out of all three and the choice I’d go with if I had a troublesome photo with noise issues and not a lot of time or desire to tweak settings. There’s only one slider to adjust and there is minimal potential to add artifacts to the image even if you max out the NR slider. Topaz Denoise AI has the advantage of being a standalone plug-in that works seamlessly with Lightroom workflows, and can achieve similar results with a little bit more effort. Both of these tools clearly offer an improvement over the standard Lightroom NR tool. If noise isn’t apparent at normal viewing distances, which is the case for 99% of the photos I shoot, I can’t see myself spending a lot of time with noise reduction, but there’s no question it helps with those photos where image noise was unavoidable.
