The State of Noise Reduction

I'm a little behind on reporting about software changes that have been occurring, so it's time to do some quick catching up. 

One of the recent trends in processing has been in noise reduction. In particular, the application of AI (artificial intelligence) or ML (machine learning) techniques in trying to reduce both color and luminance noise caused by randomness of photons. I won't get into the details of how noise gets into your files, as that's much more nuanced than I want to get into with this simple article, but suffice it to say that both low light and high ISO values tend to trigger those things. 

If you haven't already figured it out, ever since Google in all their infinite wisdom decreed that all Web sites should be responsive (because Google didn't want to miss out on ad revenue on mobile devices), all us webmasters have had to use some form of image handler in the background. You're not seeing the original pixels of what we post, you're seeing what the image handler shows you.

To get around that, you can right click on an image on this site and ask to Open Image in New Tab. This will open the full JPEG I've put on the site. You may need to click on the magnified cursor to see the original pixels, depending upon your window size.

Unfortunately, Google also decreed that page speed was important to search results, so pretty much all Web sites, including mine, don't post full images, either. Thus, what you'll find are 1600 pixel long edge images when you click through to the "full image" instead of what could be 3644 x 5487 size from the D2x being used here.

All hail Google!

The old Nikon D2x DSLR was a really good test case for noise, as at base ISO in strong lighting, it was an excellent performer. But as you tried to boost ISO in low light, noise got visible fast. And I mean fast. For instance, this ISO 800 image taken just after the sun had gone behind the nearby ridge:

I've used this photo before because it's a really good example of the problems we sometimes run into. Here we have a face in shadow, and to hold the brightness in the water, I've had to underexpose that face some. Here's a closer look at the way the photo looks going through Adobe defaults:

As you can tell, there's noise everywhere, but it's definitely a problem on the face and hand that's a big problem. In particular, note the top edge of the hand, where it's almost like there are JPEG mosquito artifacts.

So, you're asking, what about all the recent noise reduction programs? How will they fare on this image? Have things progressed enough that you should be using them? Short answer: yes, but be careful.

Before we begin, something you should be aware of is that the AI/ML approaches take time. They work best with a lot of GPU horsepower. For instance, my iMac Retina 5K has what was at the time a pretty good GPU (AMD Radeon Pro 5700 XT). But it's as little as one-sixth as fast as my M1 Max laptop with the same amount of RAM when the noise reduction process is being done locally.

If you're processing only one image at a time, waiting a minute or more for the results probably isn't going to kill you on older machines, but if you're doing a lot of consecutive processing, this all adds up fast. My first recommendation is this: if you think you want to use one of the AI/ML approaches to noise reduction and you're using an older computer, you're going to find you want to upgrade that computer. This is becoming particularly true of Macs, as the right M1/M2 chip is clearly faster than any of the older Intel-based Macs with modern noise reduction software.

A word of warning: I'm not going to apply any color, tonal, or other processing to this image as I step through some options. Only noise reduction (and sometimes sharpening). I don't want this to turn into a "let's see what's the best image I can get from each option" type of test. Many controls will tend to interact with one another, particularly if done during the demosaic. You're going to see one tool make such a change, and it's not for the good, which means you have to go in and try to fix that unintended consequence when all you wanted to do was get rid of the noise. Also, I'm going to stick to what the software thought was the best answer for noise, for the most part.

Let's start with Adobe's new enhanced noise reduction technique, which creates a DNG file from the original:

In general, Adobe's default approach is straightforward and not heavy-handed. It leaves a bit of noise and it tends to find edges well (note the top edge of the hand has been much improved). The default for Enhanced NR is better than most of you will achieve trying to use the Detail sliders in the older Adobe approach. I consider this a good step forward. Anyone in the Adobe Creative Cloud Photography program should be using this approach on noisy images. 

Note, however, that Adobe creates a new DNG file from your original when you use Enhanced NR. I'd caution you to keep the older NEF file (or other raw format you're using) around, because as we can see from this old D2x image, things have changed in software over the years, and you'll want the original data around to work on as it does.

Next up, DxO DeepPrime XD. 

I created this image via DxO PureRAW 3. Again, I'm not trying to adjust for best possible image, but rather show you what the noise reduction is doing. DxO PhotoLab 6 has the same noise reduction capabilities in it, but also starts applying other corrections by default. Overall, I believe the DeepPrime XD approach removes a bit more of the noise than Adobe does, and I also think it does a slightly better job on holding onto face detail. (Disclosure: I was a beta tester for PureRaw, and had early access.)

Next we have two Topaz approaches. First, let's try the older Topaz Denoise AI:

In this case, I had to select a "method" (standard, clear, low light, etc.), and I simply picked the one that I felt did the best job on removing noise from this image (clear). However, while noise has been removed, note the artifacts on the whitewater paddler's right chin (just to the left of his mouth) and on the clip on the shoulder strap towards the top of the image. These artifacts were worse in the other Topaz approaches, which is one reason why I picked clear. 

This is a common Topaz complaint: it does an excellent job except where it doesn't ;~). One thing I've noted is that the Topaz tendency towards artifacts is higher with their default selections than it is with some of the sub-selections you can make. I'll make another comment about that later in this article.

Topazlabs these days would likely just say "you should be using Photo AI, our newer program." Well, maybe not, because here's the result:

Yes, the noise handling is better, but note the color shift in the face and some blurry edges and a couple of color artifacts. Like Adobe, Topaz Photo AI creates a DNG file, so you'll have the ability to do some HSL processing on this to take out all that orange. Still, I don't like it when one tool I use now makes it so I have to use another tool to fix problems that the first tool added.

Speaking of which, let's go back to the Adobe Enhanced NR result and apply one additional thing to it: sharpening.

Which illustrates my point about multi-tool use. We've lost texture and granular detail. Everything is getting that all too familiar "plastic-look" result from heavy-handed processing of noise and detail. 

So how do I deal with the noise issues these days? Well, I know that many of you wonder why I still use Photoshop CC instead of Lightroom Classic. This is the answer: layers. In essence, I use layers to isolate the impacts of each tool. 

Yes, you can mostly do something similar now through the use of masking in Lightroom Classic (and ACR itself), however some tools can't be used in masking (HSL, for example), and Adobe's own sharpening leaves a lot to be desired.  

Here's what I quickly ended up with in my usual layer approach (and I mean quickly. Three layers, three tools applied, correct the layer masks. Done.):

By controlling which tool is applied to each layer (other than Adobe's noise reduction) I was able to deal with the issues that come up with each additional tool, particularly the sharpening ones. 

There's still a lot that could be done with this image, and I'd be doing that with layers, as well. 

Which brings me to my further recommendations:

  • Both Adobe's Enhanced NR and DxO's various Prime noise reductions are good starting points, and you'll apply them globally. I'm less enthused with Topaz's noise reduction choices, or those I've tried from others (e.g. onOne or CaptureOne). So pick one: Adobe or DxO. If you're a Creative Cloud Photography plan user, it'll probably be Adobe. If not, you have two choices from DxO: (1) use PureRAW to pre-process your files; or (2) move to their raw converter with PhotoLab.
  • Sharpening interacts with noise reduction, so finish/creative sharpening should be applied locally, as needed. Adobe's built-in sharpening leaves a lot to be desired, frankly. I wish Adobe had acquired the old Piccure deconvolution sharpening instead of Frazer's. Moreover, applying Adobe's choices locally via masking further limits what you can do. Topaz Sharpen AI is a reasonable choice with a huge caveat: you must verify the pixels. Sharpen AI has a tendency, particularly at the defaults, to get overenthusiastic when it finds something that it can sharpen. In an image like the one I've been using, for instance, Sharpen AI will find some out of focus water or rocks and whack them up so that they don't look like adjacent water/rocks. Very artificial, and it calls attention to areas you don't want attention paid to. I'm quick to mask them out.
  • Find an image in your files that clearly needs noise reduction, as I did here, and then use the trials available for each of the products to see how they fare on it using your workflows. Do your own evaluation. Moreover, do as I did at the end, and see what's the best possible image you can create from your file. See how the noise reduction interacts/interferes with the other things you'd do with that image.
 Looking for gear-specific information? Check out our other Web sites:
DSLRS: | mirrorless: | Z System: | film SLR: all text and original images © 2023 Thom Hogan
portions Copyright 1999-2022 Thom Hogan
All Rights Reserved — the contents of this site, including but not limited to its text, illustrations, and concepts,
may not be utilized, directly or indirectly, to inform, train, or improve any artificial intelligence program or system.