Topaz Labs have recently introduced a suite of artificial intelligence (AI) powered post processing software to help photographers get the most from their images. My experience with Topaz Denoise AI has been nothing but positive so I thought I would compare it to the other noise reduction software currently available.
What does noise reduction software do?
If you’re already familiar with noise reduction software you can skip this and the next section and go directly to the test results.
As you increase the ISO values on your camera the signal (light) that is captured is amplified. This amplification makes your image brighter, but also amplifies digital noise. At moderate values the noise is usually not objectionable and will likely not be visible on a standard sized printed image or if the image is downsized for web use. However, as ISO values increase, this noise can become problematic and start to negatively affect the look of a photo. Even some low ISO photos can exhibit high noise levels if they are highly processed in Photoshop (or similar).
Noise reduction software aims to reduce the amount of noise present in the image while retaining the detail in the important parts of your image.
The Noise Reduction Problem
Noise reduction works differently in each application, but basically it has to isolate the noise from the actual image data and reduce it’s appearance by replacing or blending the noise data to match the actual image.
Bad noise reduction software will often have trouble separating actual image detail from the noise and will end up blurring or removing important parts of the image in the process of noise reduction.
Topaz DeNoise AI Details
Topaz Denoise AI uses a machine learning algorithm in order to learn the difference between unwanted noise and important image details. The software has been fed many thousands of images in order to be trained on what a ‘good’ image looks like vs a bad one. It then uses these models to analyse your photo and produce the best noise-free image it can.
The downside to the software that runs off AI is that it requires quite a bit of processing power (or time) to process an image. On my 2018 MacBook Pro with a Radeon Vega 64 graphics card connected via an eGPU, a 24 megapixel image takes around 30-40 seconds to fully process. Time well worth it to clean up noisy images in my opinion.
If you want to learn more you can check out DeNoise AI at the Topaz Labs website.
My experience with this software has been nothing short of amazing and I wanted to test it against other well regarded noise reduction software.
Test Images and Setup
For the test I have chosen an image that shows a significant amount of noise and also has fine detail that needs to be preserved. In order to maximise the results each application is capable of, I worked on RAW images in the applications that supported them and 16-bit converted (with zero noise reduction applied) TIFF images in the other applications. The RAW images were matched in processing as closely as I could to ensure things were as ‘fair’ as possible, but due to the different demosaicing engines there will be subtle differences between the RAW images.
In each application I have adjusted any available settings to get what I think is the best compromise between noise reduction and detail retention. The aim is to produce an image that retains an acceptable level of detail and then compare the noise levels. If the software included a function to sharpen the image I used it in conjunction with the noise reduction to produce the best image possible.
The full image I used in the tests is as follows:
The 100% crop from the above image I used when judging (no sharpening or noise reduction applied):
This image was taken at ISO4500 on a Nikon D750 and the exposure pushed slightly higher in post-processing which exacerbated the noise even further. There are fine details in the bird’s feathers that are important to retain.
I have listed the results below in order from best to worst. I have also included some comments around what I think of each results. If you want you can skip straight to the conclusion.
In order to see the differences in noise reduction the images below should be viewed at 100%. Please click on each thumbnail. You will need to zoom into the image if your monitor is less than 1600px wide.
Topaz DeNoise AI
You can see straight away that the noise has been all but removed from this image, there is hardly a trace left in the background, eye or beak. There has also been almost zero loss of detail in the bird’s feathers.
DxO PhotoLab 3 (PRIME)
Photolab 3 is a RAW processor with PRIME noise reduction (included in the Elite version). Due to the differences in RAW conversion between it and Lightroom there are some subtle colour differences between this and the DeNoise AI image.
It hasn’t done quite as good a job in removing the background noise, but the result is still very impressive and is a huge improvement over the original. Feather details have been retained as well as DeNoise AI. In most display situations the differences between this and the DeNoise AI image would be very hard to find.
Neat Image has done a very good job with reducing the noise and keeping detail. The result has not reduced the amount of detail present in the image, but it hasn’t sharpened the details quite as well as DeNoise AI or PhotoLab. In terms of noise reduction, It is almost even with DXO, but falls slightly short due to a small amount of ‘blotchiness’ present in the background caused by sharpening artefacts that aren’t present on the previous two results.
Capture One Pro
Capture One Pro (version 20) is also a RAW converter and has recently had an update to its noise reduction algorithm that promises better detail retention and smoother noise reduction.
The result is significantly more noisy than the DeNoise AI, PhotoLab or Neat Image efforts and the amount of detail retained isn’t quite a good. The Capture One image is certainly very usable in most situations and a decent result overall, just not as clean as the first two options. For moderate print sizes or web display Capture One has done a fine job.
Lightroom Classic CC
Lightroom was able to retain similar levels of detail as Capture One Pro, but left behind a little more noise in the process. It also introduced some strange artefacts around some edges that are less than desirable (check out the top of the beak in the image above. Still a good result, and very close to Capture One Pro, but no where near as good as first and second place.
The Noiseware image does an OK job at removing noise, but it does so at the expense of image detail. The feathers above the bird’s eye have lost a lot of definition, as have the breast feathers. Not a great result when compared to the results of the better software above.
Nik Define (recently acquired by DXO, the owners of PhotoLab) shows similar results to Noiseware. Too much detail has been lost in this image and I wouldn’t consider using it based on the other options available.
From the tests I have conducted there are certainly three software packages that are head and shoulders above the others, retaining most (if not all) fine detail and doing a stellar job of removing noise:
Of these three, DeNoise AI is the overall winner as it was able to completely remove background noise, whereas PhotoLab left a very fine grain still visible, as did Neat Image. In a print or web sized digital image the difference between the three is negligible and all produce stellar results.
One thing to note is that PhotoLab is a complete (and quite good) RAW converter and only offers PRIME noise reduction when working on original RAW files, whereas DeNoise AI will only work on already converted RAW files (from Lightroom for example).
A distant fourth comes:
Capture One Pro upgraded their noise reduction algorithm in the recently released v20 of their RAW conversion software. It does a very good job of retaining detail (not quite as good as the first two options). It leaves behind much more noise than DeNoise AI and PhotoLab, but is a very large improvement over the original file and would do very well for all but the most picky users or if very large prints are involved.
Lightroom was able to maintain a similar level of detail to Capture One Pro, but in doing so left more noise behind. Lightroom also introduces some edge artefacts that aren’t present on the preceding software. The result is still good, and produces an image that would be usable in most cases – it just is a couple of notches below the others.
The last two placed software are fairly similar in their results. They were both able to reduce a decent amount of noise, but in doing so they obliterated too much detail. I wouldn’t use either of these two options:
- Nik Define 2
Do you have any other favourite software that you’d like me to include in the test? Please let me know in the comments below.