Topaz DeNoise AI is a hot and popular product for cleaning up noisy images. But does it work on astrophotos?
All images by Alan Dyer (amazingsky.com)
Topaz DeNoise AI
Who Is It For? Deep sky and nightscape photographers contending with noise at the final stage of processing.
“AI” is the current hot feature in image processing software. A filter trained with artificial intelligence calculates and applies its effect automatically, based on what it “knows” has worked well on thousands of other images similar to yours. AI promises to take much of the guesswork out of photo editing.
Leading the trend to AI is Topaz Labs, with a suite of popular programs: Gigapixel, Mask, Sharpen and, tested here, DeNoise. Even users of specialized programs such as PixInsight have found DeNoise AI a useful tool late in the processing workflow of deep sky astrophotos. I’ve found it works especially well cleaning up noise in high-ISO nightscape images.
But, as the results below show, if you are not careful DeNoise AI can ruin your images. The artificial intelligence isn’t yet smart enough to fully understand astrophotos. The issue was not so much wiping out stars, but adding unsightly artifacts in place of the random noise it was smoothing.
Stand Alone or Plug-In
DeNoise AI can run as a stand-alone program capable of opening JPGs, TIFs, and many formats of raw files. As such, it can be fed a folder of images for batch processing, perhaps for a time-lapse set. After it processes images, it can export files back out as raw DNGs, or as JPGs, TIFs or PNGs.
However, when you install DeNoise it also installs as a plug-in for Adobe Lightroom and Photoshop. I ran all my tests on it out of Photoshop, applying DeNoise as a non-destructive “smart filter” to image layers that were smart objects. I prefer that workflow as it allows DeNoise AI to be reopened at any time to change its settings.
In testing the plug-in version of DeNoise v3.0.3, I discovered several deficiencies:
— If the image layer has a mask, DeNoise AI refuses to apply, giving an error message, even if the mask is blank with no content.
— Unlike most smart filters, it does not save the unique settings applied to that smart object. It reopens with the last settings you used anywhere, not with the settings you applied to that image you are re-editing.
— There is no way to take settings you’ve found that work well and save them as “user favorites.” You always have to start from scratch when editing an image.
While the Auto button might work, to avoid the issues I show below I found I had to fiddle with settings trial-and-error. This partly defeats AI’s promise of one-click processing.
Deep Sky Example
For this example I used a stack of eight 8-minute exposures of Orion shot with an Optolong L-Enhance filter, using a red-sensitive Canon EOS Ra camera at ISO 3200. Despite the stacking, a fair degree of noise remained in the dark sky background. Could DeNoise AI clean it up?
Yes, but at the risk of introducing odd-looking image artifacts.
DeNoise AI offers several modes of operation: the namesake DeNoise AI is the main mode, but there’s a legacy AI Clear mode (which I found rarely worked well), and a Low Light mode which ostensibly should be best for most astrophotos. It wasn’t.
When applying the DeNoise AI mode, it’s certainly worth trying Auto. However, you might need to increase the “Remove Noise” slider and decrease “Enhance Sharpness,” as the latter can add dark rings around stars. This can be the case even when Sharpness is slid over to 1, the lowest setting.
“Color Noise Reduction” can help eliminate the colored halos around stars imparted by some narrowband nebula filters. But turning it up too high wipes out star colors.
“Recover Original Detail” can bring back the fine structures and small stars smoothed out by the noise reduction, but at the cost of less noise reduction, always the trade-off with such programs, AI not withstanding.
The latter two sliders are not affected by the Auto button, so they have to be dialled in manually for each image. They’ll stay where you last set them.
I found DeNoise settings that at first glance looked good could introduce strange pixel-level speckling in the darkest areas of sky. It was smoothing some noise, but adding another form of unwanted pixels.
At first glance, the Low Light mode seems to work better, as it does allow for 0 Sharpening. But again, inspect carefully. I found Low Light often added a bizarre banding across most parts of the image. Random noise was replaced by a uniform texture that got worse with higher levels of noise reduction.
This was also at the pixel level, to be sure, but that’s what we’re concerned with when finessing noise reduction, because stars, like noise, live down at the level of individual pixels. Increasing “Recover Original Detail” helped eliminate this pattern noise, but again at the cost of lower noise reduction.
I found similar results with typical nightscape images. Low Light mode was problematic. The standard DeNoise AI mode worked best, avoiding the textured sky artifacts introduced by Low Light.
But Auto often wasn’t very effective. As with deep sky images, “Remove Noise” had to increase, and “Enhance Sharpness” had to be set to 1. While this applied some useful sharpening to ground details, it often added edge artifacts characteristic of over-sharpening.
Problem areas can be masked out in DeNoise, so its effect does not apply to that part of the image. But that involves more manual work.
The examples here show that, yes, while Topaz DeNoise did smooth random noise, it can introduce other problems photographers working with “normal” images might never notice.
But we astrophotographers deal with tiny stars, planets and crescent Moons, among other demanding subjects.
While I always apply noise reduction at the raw stage early in my processing workflows, the routines I use from Adobe, while good, are often not sufficient. I’ve found DeNoise AI can be a fine additional tool when applying finishing touches at the final stage of image editing.
But pixel peep. Don’t trust the Auto AI. If you are not careful you might find Topaz DeNoise adds more problems than it eliminates, trading random noise for pattern noise or over-sharpening artifacts.
Topaz has some new competitors. As I write this, two other noise reduction options have just been introduced or will soon be available:
— DxO’s PureRAW (https://www.dxo.com) for pre-processing raw files, and
— ON1 PhotoRAW’s NoNoiseAI (https://www.on1.com/products/nonoise-ai/), a direct competitor to DeNoiseAI.
Download the trial versions of each to see which one works best for you.
UPDATE: Since I prepared the review, Topaz updated DeNoise to version 3.1, which has added a new “Severe Noise” mode, and renamed the “DeNoise” mode as “Standard.” While some of the pattern artifacts described seem to have been reduced, they have not been eliminated. Inspect your images closely.
Plus: When applied judiciously, it smooths noise in high ISO images very well.
Minus: When applied injudiciously, it can add pattern noise and sharpening artifacts.
Summary: Topaz DeNoise AI can smooth noise better than most standard noise reduction filters. But inspect the results carefully. Be prepared to fiddle with the settings for optimum results.
Retail price: $79.95 for MacOS or Windows (free trial copy available)