Denoising software comparison
Introduction
This report summarizes a comparative test of the denoising performance of four different image processing software programs. At the time of this writing, three of them have a reputation of being the best of the best when it comes to noise reduction (NR), namely Topaz DeNoise AI, DxO PhotoLab (with its DeepPRIME and DeepPRIME XD algorithms), and ON1 PhotoRaw (with its NoNoise AI algorithm).
The fourth contender is the “New kid on the block” when it comes to strong denoising performance. Previously a weak player that could not hold a candle to the other three, Adobe Lightroom’s latest revision adds an AI denoising engine (“Enhanced NR”) that according to the company delivers excellent results.
Since I own all of these software packages, and as I have or used to have cameras from different brands, I thought it would be worthwhile to test each software’s NR performance with high-ISO images taken with each of three different bodies, a Canon R5, a Nikon D850, and an Olympus/OMDS OM-1. In Nikon’s case, I am using a test image that was shot with a DSLR, whereas the R5 and OM-1 are mirrorless bodies. (I also owned Nikon’s Z7 and Z7ii mirrorless cameras at some point but never took high-ISO shots with either, since their AF performance was lacking in low light.)
For comparison purposes, I also included CaptureOne’s noise reduction in the report. This is a ‘classic’, non-AI denoiser.
The Test
Here are the three original test images used for the comparison.
The originals are RAW files (.CR3, .NEF, .ORF). The images shown here were converted to JPEGs in DxO PhotoLab without any edits, then cropped to 1,000×1,000px. The image taken with the Canon R5 has the most acceptable looking noise, in spite of having been shot with the highest ISO setting, at 25,600, which is testimony to that body’s excellent low-noise performance. This particular shot is slightly unsharp, though not by much. The Nikon D850’s noise pattern is less uniform than the Canon’s and the shot looks as noisy as, maybe even more so than, the Canon one, while the Nikon’s ISO setting was quite a bit lower at 10,000. The OM-1 test shot, taken at an ISO of 16,000, shows the most color noise. Together with the high overall noise level caused by that camera’s much smaller sensor and the fact that the shot is underexposed, this presents a particular challenge for any NR engine.
The noise reduction software contestants in this test are:
- Adobe Photoshop Lightroom Classic 12.3
- CaptureOne 23 Pro 16.1.2.44
- DxO PhotoLab 6.4.0
- ON1 PhotoRaw 2023 17.0.2.13102
- Topaz DeNoise AI 3.7.0
Notes: DxO has two different NR engines, DeepPRIME and DeepPRIME XD. Both of them are included in the test. (Truth be told, DxO still also offers the old PRIME engine, but that one performs rather poorly when compared to the others.) DxO’s PureRaw uses these same NR engines and is therefore not part of this comparison. The latest versions of Adobe Camera Raw use the same AI engine as Lightroom does and produce identical results, which is why they are also not included. Furthermore, Topaz recently released Photo AI, which combines their DeNoise, Sharpen and Gigapixel AI apps. It is omitted from this test, as well.
All test shots were processed at each program’s default settings for the respective NR engine. Each program was also blocked from doing anything else with the image, meaning that I loaded the RAW file, selected the noise reduction at the program’s default settings, disabled all other editing functions, and exported the image as a JPEG. In case of Lightroom and because of the way this program works, this meant an interim conversion to the DNG file format. In order to avoid artifacts, each image was exported with a quality setting of 100, which meant large files. No sharpening was applied anywhere. The images shown below represent 100% crops from these exported images, each of them 1,000×1,000 pixels large.
One exception from the above warrants a separate discussion: In its standard setting, Topaz DeNoise AI sometimes achieves little noise improvement and uneven noise removal, certainly nowhere near what the program is capable of. It has several other AI models that each come with their own default settings. I therefore chose “Clear” instead of “Standard”. Using this model with its preferences set to Auto still proved insufficient, though. I had to always keep its “Remove Noise” setting at High to get comparable results, as leaving it at Medium resulted in luminance noise still being high, though the program handled color noise well in that setting. Furthermore, while it CAN be used as a standalone program, DeNoise gives the user no control over the format in which to save. To circumvent this, I imported each test image into Lightroom and used Topaz DeNoise AI as a plug-in, then saved the resulting image as a JPEG with the quality set to 100. This means that the results with Topaz were converted from RAW by Lightroom, and thus show the same overall colors.
On a side note, ON1 crashed my computer twice, each at the same point and with a Blue Screen of Death, while trying to process the OM-1 image. I eventually succeeded by using a different order of unclicking the program’s processing options.
Test Image 1: Canon R5
All six NR engines produced a good result here. After extensive pixel peeping, I picked DxO’s DeepPRIME XD and Topaz DeNoise AI as joint winners, followed by ON1 NoNoise AI, Lightroom, and DxO DeepPRIME. They are within a hair of each other, though, so other viewers may feel differently. CaptureOne leaves considerably more noise in the ‘clean’ areas while also losing a bit more detail in the feathers, but still produces an acceptable result.
Partly because the original shot was slightly unsharp, none of the programs produced any unpleasant artifacts, which are more likely to surface with areas within the shot that have a high microcontrast.
Test Image 2: Nikon D850
I’ll start with the same comment as before: all six NR engines produced a good result here, with CaptureOne again struggling the most. This time, I’d pick DxO’s DeepPRIME XD and ON1 NoNoise AI as joint winners, followed by Topaz DeNoise AI, with Lightroom and DxO DeepPRIME in the second-from-last place, but still close to the others and leaving CaptureOne far behind. If you look at the full images from a normal viewing distance, you’re not going to see much of a difference in noise levels.
What is interesting here is that ON1 and Topaz both sharpened the image quite a bit, but in different ways: in the ON1 image, it is not only the feather detail that is sharp, but the program also sharpened all other parts of the bird, such as the hairs just below the beak. In doing so, it produced some artifacts in that area. Topaz, on the other hand, sharpened even more in this image (look at the eye details, for example), but it apparently did so selectively: the hairs under the beak are left relatively unsharp, which avoided producing any artifacts.
Test Image 3: Olympus/OMDS OM-1
Ok, here comes the result of the hardest test. None of the contestants did a perfect job here, though some produced impressive results nonetheless. Overall, the field is further apart than with the previous two test images. As far as smoothing out the surroundings of the bird goes, DxO DeepPRIME XD did the best job, with ON1 NoNoise AI giving us an equally smooth area that unfortunately shows uneven and unpleasant coloration. On the other hand, DxO leaves the area around the bird’s facial hair with visibly higher levels of noise, with ON1 doing a better job there.
Next come Lightroom and DxO DeepPRIME. In my opinion, the former has a small edge when it comes to smoothing the surroundings. On the other hand, Lightroom produces some rather weird artifacts AND discolorations on the bird’s front just below the beak, where the results from DeepPRIME are more compelling.
Topaz disappoints here: DeNoise AI did not do a nearly as good as the other AI engines in this case, creating lots of artifacts, leaving plenty of noise and also showing some coloration issues, especially in the yellow areas. I tried re-running this one with other settings, but while some improvements were achieved by doing so, none of them produced a fully compelling result. This may have to do with the image structure more than with the amount of noise: it often gets bemoaned in photography forums that Topaz works less well with OM-1 images than with those from other bodies. Personally, I have achieved much more compelling results with high-noise images from other bodies with Topaz in the past.
CaptureOne is again found in last place, even when compared with Topaz. C1 left plenty of noise and a rather mushy area below the beak, where the others all recovered more detail.
General Observations
Predictable, at least for folks who regularly use multiple editors, were the color differences between the different test results. Each program comes with its own rendering engine when converting from RAW, and those never produce identical results. These differences are largely a matter of taste, so I won’t belabor them too much, except for saying that personally, I rather dislike what ON1 did with the Canon and OM-1 shots. In both cases, I would have to do quite a bit of color correction.
ON1’s NoNoiseAI engine, in its default settings, apparently sharpens the image quite a bit, even though the (separate) sharpening function was turned off. While this makes the resulting images look “better” to those who concentrate on sharpness, it can actually be a disadvantage in the context of this test, since sharpening usually increases the visible noise. Considering this, the software did surprisingly well. Others, in particular Topaz DeNoise AI, also apply a small amount of sharpening. DxO DeepPRIME XD produces not only cleaner, but also sharper results than regular DeepPRIME. I’m not sure whether that is because it actually applies dedicated sharpening or because its algorithm simply recovers more of the image’s structure. I suspect it is the latter but have no way of knowing for sure.
The images produced with Lightroom, both with its internal NR engine and with Topaz DeNoise AI as a plug-in, show a small geometrical correction when compared to the original image. The bird’s head was slightly larger than in all other images. I checked and double-checked: all relevant setting were disabled. Lens corrections were set to Manual, with distortion and vignetting both at zero, Transform was off, etc. It seems that Lightroom changes the image geometry some when converting the RAW image, regardless of user input.
Just to see how much better the NR would get, I played around with CaptureOne’s settings. With the Canon and Nikon test shots, upping the luminance setting from the default of 50 to about 85 produced cleaner results in the smooth areas. It did not help much with the feather detail, though. Changing the other settings made almost no difference whatsoever. In case of the OM-1 test image, I could not find any settings that produced a better result than the default settings did: increasing the luminance NR reduction washed out the details too much, while increasing the details setting brought too much of the noise back. The other two sliders again had a negligible impact.
This comparison would be incomplete without mentioning how much time each software needs to process the images. I did not measure the exact times as that was not my primary objective. I use a notebook with a fast SSD, lots of memory and a fairly potent Intel CPU, but it has only the graphics processor built into that CPU. On this machine, ON1 and Topaz ran the fastest, with DxO DeepPRIME needing longer, Lightroom needing MUCH longer, and DxO DeepPRIME XD needing FOREVER, where the latter can mean in excess of 10 minutes per image, slightly less for the smaller OM-1 one. CaptureOne also ran very fast. AI NR engines are never fast, but the difference in speed between the fastest and the slowest ones I tested here was in the neighborhood of 6x-10x. In other words, be prepared to wait. Since my notebook has no dedicated GPU, and since these programs likely benefit from such a GPU to varying degrees, your experience regarding speed may be different from mine.
My Conclusions
To get the absolutely best result, DxO’s DeepPRIME XD is the engine to use. ON1 is very close but presents more of a challenge on the color rendering front than DxO does. Topaz also belongs into this group, but only if you don’t shoot with an OM-1. Overall, Topaz needs more manual intervention to work well. It does offer several adjustment options, however, that come in handy with difficult-to-reduce noise.
Lightroom and DxO DeepPRIME perform at similar levels, and both produce good results that are only a half notch below the top performers. Keep in mind, however, that producing better results usually takes longer time. From a NR-quality-versus-speed perspective, ON1 impressed me most. If only it didn’t mess up my colors!
CaptureOne disappointed with all three test images, at least in comparison to all other NR engines tested. This serves to show that using AI approaches indeed introduced substantial benefits to noise reduction.
So, no, Lightroom isn’t the new king in this jungle. It (finally!) offers good and competent noise reduction, but it runs a little too slowly (on my hardware) and produces good but not outstanding results. Version 12.3 is MUCH better than the previous versions of Lightroom, though.
Oh, and in case you didn’t notice: even that horrible looking OM-1 RAW file can be made into a very nice and clean looking image. Don’t be too scared of high ISOs! 🙂
Lothar Katz
April 2023