Topaz Denoise Vs Denoise Ai

Topaz Labs released version 2.4 of DeNoise AI with several new improvements (see the full change log here). Topaz Labs also introduced two new saving offers (valid from January 27 – February 12): Topaz Labs DeNoise AI is now $20 off Topaz Labs Utility Bundle is now $50 off You can use coupon code RUMORS15. Dec 29, 2019 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.

Topaz Denoise Vs Denoise Air Fryer
Topaz Denoise Vs Denoise Ai Machine Learning
Topaz Denoise Ai Upgrade
Topaz Denoise Vs Denoise Ai Adobe

Topaz does recommend using Denoise as early on in the process as possible, so other edits don't throw off the algorithms. I usually run my raws through it and export as tiffs for editing, but I may move it to after running through Raw Therapee.
High Noise Workflow (Using Lightroom, Topaz Denoise and On1 Effects)In this video, I demonstrate my workflow for an image that was shot at a higher ISO exhib.

Noise Reduction (NR, for short, since that term’s gonna come up a lot in this post) is an oft-misunderstood topic. It’s not as easy as it sounds since proper NR involves careful manual adjustments based on the level of noise present in the image, not just dragging a slider all the way to the right (and no, that does not constitute manual labor lol).

What is Noise?

Before we get to the meat of this post, it’s important to understand what noise is and how NR works. Noise is an artifact caused by digital sensors and is present in all digital images, no matter how low an ISO is used. Noise does not exist on its own though. The ‘Signal-to-Noise ratio’ (SNR) is a good indicator of how much detail (Signal) is captured relative to the amount of noise in an image. For example, at a camera’s native ISO (say ISO 100), the SNR is very high because the level of detail captured is significantly higher than the level of noise present. As the ISO is increased, the SNR falls and the level of noise catches up with the level of detail. Why is this important? Because that’s the basic premise of NR.

How does Noise Reduction work?

NR seeks to minimize the level of noise by analyzing a detail-less sample area within an image. This allows the software to distinguish between the pattern of noise and areas of detail. At the moment, the only software that allows precise manual control over the selection of the sample, calibration, and level of noise reduction is Neat Image. Over the past few months, though, I’ve seen and received a ton of ads about a new, shiny, AI-based, Rolls-Royce-esque software on the market – Topaz Labs’ DeNoise AI.

Keep it simple. What’s DeNoise AI?

Basically, Topaz DeNoise applies NR using algorithms developed by “feeding it millions of noisy/clear images until it actually learned what noise is and how best to remove it.” (from Topaz Labs’ website).

Sounds neat (pun totally intended), but does it work?

Well I’m glad you asked! The short answer is….it’s complicated. It does work and it performs better than I expected, but I have a few reservations. For the purposes of this post, I’m only going to talk about High-ISO results; Low-ISO results are more or less the same with any generic NR software.

My NR Process

Before I show you the samples, I’ll briefly explain my NR workflow. I always apply NR selectively using layer masks. I apply a pass of very strong NR on the background, and a lighter pass on the subject. This ensures that there is minimal loss of detail on the subject while rendering the background smooth and pleasing.

Purple Sandpiper – ISO 4000

This is a very noisy image shot at ISO 4000 on my Nikon D850. As you can (probably) see, the background’s quite rough. The good thing, though, is that the noise is grainy and NR should be an easy task. The bad news? There’s a lot of noise on the out-of-focus area on the bird’s body. NR on this area will require a compromise between noise suppression and detail retention. Anyways, let’s see how the software fared.

Compared to the original capture, both Neat Image and Topaz DeNoise AI have performed very well. Looking at the results critically, Neat Image has rendered the background way smoother than DeNoise AI. This is because the latter has attempted to preserve detail in the out-of-focus areas, but it has caused parts of the background to turn blotchy and ‘nervous’. The bird’s head seems a hair (hehehe…) sharper in the DeNoise AI image, but there’s a reason for the difference.

Understanding the Algorithms

From my experience, what DeNoise AI does is that it suppresses noise, but it also adds sharpening and contrast (even when the Sharpening and Preserve Details sliders are set to the lowest setting of 0). While this may seem appealing in promotions and even in real-life use, it’s actually misleading and, ultimately, reducing the quality of your files. The claim that the software, according to Topaz Labs, “recovers crisp detail” may sound true on paper, but sharpening is NOT detail recovery. It’s the addition of contrast at a pixel level that gives the illusion of a higher level of detail, when it’s actually destructive. Sharpening is supposed to be the final step in processing an image, and should never be applied to the master image during NR.

In contrast, Neat Image has rendered the file a lot smoother. There are no hints of nervousness in the out-of-focus regions. The bird’s head is rendered slightly softer, but that’s because Neat Image does not apply sharpening or contrast during the NR process; the fine feather detail’s still present. The main downside is that the learning curve for the software is steep and takes some experimentation and getting used to.

A pattern I noticed throughout my tests was that practically every image that DeNoise AI was applied on had a few areas that were either completely smudged due to excessive NR or had too much noise due to no/low NR. My theory is that DeNoise AI sometimes fails to separate areas of detail and noise in the image. The problem with this is that there is no opportunity to manually reduce/apply NR in those areas; it can only be done by the software.

Peregrine Falcon – ISO 4000

Here’s another ISO 4000 example of a Peregrine Falcon. Again, the background and the subject are very noisy, but the problem here is that the subject is also full of detail, and isolating the noise from the detail is a massive ask for either software. Let’s see the results from this test.

Again, the results are similar, with Neat Image having a slightly grainier background, but a more even NR on the subject. Topaz DeNoise, even at very low NR settings, is quite heavy-handed. However, I’m (mostly) impressed with it’s ability to separate detail from noise. Layer Masking with Neat Image was quite a painful experience in this case even though, in my opinion, it gave the better end result. Topaz DeNoise was comparatively much quicker which may be appealing to most people.

I applied the same level of output sharpening for both the Neat Image as well as the DeNoise AI results. Since DeNoise AI already applies a layer of sharpening in its NR process, the final image looks sharper when down-sized for web use, but halo artifacts can be found in the master TIFF file. Also, when you look at certain parts of the image, such as the underwing shadows, you can see that DeNoise has applied very strong NR, which has eliminated detail. The ‘Blacks’ are also deeper due to the added contrast mentioned earlier.

TL;DR

Topaz DeNoise AI

Pros:

Fast NR
Can be used without layer masks with a good degree of accuracy
Good accuracy, although it can sometimes cause artifacts

CONS:

Less accurate NR than Neat Image; irregular NR in parts of an image with no ability to manually override AI
No manual control over sample area or other noise level parameters
Heavy-handed NR even at low settings
Applies sharpening and contrast – reduces quality but gives the illusion of a sharper file
Preserves detail in the background (even when ‘Preserve Details’ is set to 0) – causes blotchy backgrounds. Artifact suppression isn’t provided.

Neat Image V8

Pros:

Very accurate NR and gives precise manual control over sample area, noise level and NR parameters
Fast processing
Does not usually cause artifacts – offers artifact suppression tools to reduce them if they occur

CONS:

Very steep learning curve; difficult for beginners
Best used with Layer Masks – can be time-consuming and strenuous

Conclusion

When I first heard of Topaz DeNoise AI, I was quick to dismiss it as just another gimmick. After using though, my opinion has changed somewhat. I still maintain that it’s not the ideal NR solution – Neat Image is still the king of NR software – but it’s certainly a big step forward compared to the other products on the market, and even LR and C1Pro’s own inbuilt NR sliders.

My main gripe with the product, though, isn’t the quality of the NR – it’s more than decent – but with the algorithms themselves. I don’t get why Topaz would want to implement sharpening and contrast within the NR algorithms for any reason other than to give an illusion to the user of, as they put it, “a pixel-perfect photo in any situation.” That’s simply not true. Sharpening anything but the down-sized/up-sized file AFTER NR is not ideal. The sharpening/detail sliders seem redundant to me considering they’re applied to the image regardless of the setting you choose. DeNoise AI is a genuinely good NR software (even if it has its quirks), but it’s been over-cooked, in my opinion, by some of the decisions at Topaz Labs – it’s not “groundbreaking technology”, it’s not going to make your images “noise-free”, but it gets the job done relatively easily. As with most things in photography, a ‘one-click solution’ is never going to give you the best results. However, for the photographer who needs to process images relatively quickly and isn’t too concerned about critical image quality, DeNoise AI might be a good fit.

Disclaimer

I’m not sponsored by either Neat Image or Topaz DeNoise. I used the 30-day free trial of DeNoise AI that’s available on Topaz Labs’ website, and the Pro version of Neat Image V8 that I purchased for my personal use.

Link to the 30-day free trial of Topaz DeNoise AI: https://topazlabs.com/denoise-ai/

Most photographers I know would rather spend time creating new images than painstakingly enhancing recent captures. Topaz Labs has introduced four new programs that use artificial intelligence to take some of the drudgery out of post-production.

From improving image quality in smartphones to creating software presets that emulate prominent painters’ styles, the branch of AI called machine learning is changing the imaging world. Firms including Adobe, Skylum, CyberLink, and Topaz are using AI to extend the capabilities of their products.

Topaz has released Sharpen AI, DeNoise AI, Gigapixel AI, and JPEG to Raw AI. Each one functions as a standalone application, while Sharpen and DeNoise are also plug-ins for Adobe Photoshop, Adobe Photoshop Lightroom Classic, and the free Topaz Studio.

The product names make their primary uses self-explanatory, but each goes beyond its obvious purpose. Sharpen AI has learned the difference between image information and image noise. It sharpens the image, actually seeming to slightly smooth out-of-focus areas, and ignores noise. But there are two additional tabs to the Sharpen mode: Stabilize and Focus.

© Stan Sholik
Original image before Sharpen/Stabilize© Stan Sholik
Final image after Sharpen/Stabilize with camera movement eliminated and a much sharper flower

Stabilize reverses and often eliminates image blur caused by handheld camera shake. Similar to but better than Photoshop’s Shake Reduction, it eliminates the artifacts sometimes found in Shake Reduction. The Focus module actually improves focus—within limitations. Topaz claims it can correct up to 10 pixels of defocus. With high-res sensors, 10 pixels isn’t much, but I’ve found it helps with macro photos. Applying Focus to an entire image can create artifacts. It’s often best applied by compositing the newly sharpened area back into the original image.

Topaz Denoise Vs Denoise Air Fryer

© Stan Sholik
Final image after Sharpen/Focus. The eyes are much sharper with sharpness decreasing naturally with no change in out-of-focus background.

Image noise reduction hasn’t advanced much since it was first introduced, with the exception of DxO Prime noise reduction in the Elite edition of DxO PhotoLab 2, which to me remains the gold standard. Using AI machine learning, Topaz reportedly fed its noise algorithm millions of images to teach it the difference between noise and image detail. The result is DeNoise AI, which incorporates AI Clear, the Topaz application that formerly provided noise reduction and image sharpening. I found it superior to any noise reduction software other than DxO Prime in its ability to eliminate noise while retaining and even enhancing image detail. DeNoise eliminated noise in my high-ISO Milky Way images without eliminating the stars.

Topaz Denoise Vs Denoise Ai Machine Learning

Final image (cropped) after noise reduction

Topaz Denoise Ai Upgrade

Topaz Denoise Vs Denoise Ai Adobe

Denoise AI interface showing split-screen preview and settings

Image enlargement algorithms have improved over the years, but even the best from ON1 Photo Raw 2019 and Alien Skin Exposure X4 are based on pixel interpolation. Using machine learning on millions of images, Topaz has trained its Gigapixel AI algorithms how to add detail into areas to increase image detail and resolution. In testing I found impressive sharpness and detail in regions where there are natural, random textures such as grass, foliage, stone, water, and even skin. Enlarging images that predominately feature these textures shows amazing results. With images containing straight lines, buildings, and the like it does less well under close inspection. Gigapixel AI performs enlargements up to 600% and has a batch processing mode for multiple images or processing an entire folder of images. The ability to turn a 1,260x720-pixel smartphone image into a 24x14-inch inkjet print with improved highlight and shadow detail is now in the realm of possibility.

If you’ve captured images in JPEG format but you wish you had the image enhancement controls available for raw files, you need JPEG to Raw AI. Not only does the program output a DNG file that makes all raw file enhancement options available, JPEG to Raw AI improves highlight and shadow detail, removes JPEG artifacts, and enhances dynamic range along the way.

The four Topaz AI programs have a nearly identical look. The controls in each are a minimal set of sliders, and there is no learning curve to their operation. You do need to refrain from pushing the sliders to their maximum setting in either direction so that you don’t create artifacts somewhere in the image.

A movable vertical partition in the preview window allows you to see the effect of each slider adjustment. I wish there were a way to see the full image in the preview so I could choose the area to magnify—all of the programs default to at least a 50% magnification. This would allow you to preview artifacts in areas of the image that are not magnified.

But it would likely slow the process further as speed is the biggest issue with these programs. There is some heavy GPU processing going on behind the scenes in each. While my home computers aren’t the fastest, even my latest iMac at the studio kept me waiting for a few minutes while processing a focus correction. Each pan to a different area of the image and each slider adjustment is met with a corresponding pause while the preview processes. Fortunately, a button is available to disable the automatic preview updating.

Both Topaz Sharpen AI and DeNoise AI are major improvements over most competitive offerings and deserve to be in every photographer’s toolbox. Sharpen AI and DeNoise AI are available from Topaz for $79.99 each, Gigapixel AI and JPEG to Raw AI for $99.99 each, and the bundle of four is $249.99. A 30-day trial version of each program is available. If you decide on the bundle, the Topaz recommended order of use is: JPEG to Raw AI (if your original image is a JPEG), Sharpen AI, DeNoise AI, and last, Gigapixel AI.

Stan Sholik is a writer and photographer in San Clemente, California.

View Gallery

Tags:post capture