A Comparison of Noise Reduction Programs for Astrophotography

On this detailed technical article, I examine six AI-based noise discount applications for the calls for of astrophotography. Some can work wonders. Others can destroy your picture.

Over the past two years, now we have seen a spate of specialised applications launched for eradicating digital noise from photographs. The brand new technology of applications makes use of synthetic intelligence (AI), AKA machine studying, educated on hundreds of photos to higher distinguish undesirable noise from fascinating picture content material.

At the very least that’s the promise – and for noisy however regular daytime photos they do work very properly.

However in astrophotography, our principal topics – stars – can look rather a lot like specks of pixel-level noise. How properly can every program scale back noise with out eliminating stars or needed particulars, or introducing odd artifacts, making photos worse?

To search out out, I examined six of the brand new AI-based applications on real-world – or quite “real-sky” – astrophotos. Does one program stand out from the remaining for astrophotography?

Notice: All the photographs are full-resolution JPGs you’ll be able to faucet or click on on to obtain for detailed inspection.

TL;DR Abstract

The brand new AI-trained noise discount applications can certainly get rid of noise higher than older non-AI applications whereas leaving tremendous particulars untouched and even sharpening them.

  • Of the group examined, the winner to be used on simply star-filled photos is a specialised program for astrophotography, NoiseXTerminator from RC-Astro.
  • For nightscapes and different photos, Topaz DeNoise AI carried out properly, higher than it did in earlier variations that left plenty of patchy artifacts, one thing AI applications may be susceptible to.
  • Whereas ON1’s new NoNoise AI 2023 carried out tremendous, it proved barely worse in some circumstances than its earlier 2022 model. Its new sharpening routine wants work.
  • Different new applications, notably Topaz Photograph AI and Luminar’s Noisemuch less AI, additionally want enchancment earlier than they’re prepared for use for the pains of astrophotography.
  • For causes defined beneath, I might not suggest DxO’s PureRAW2.
The three take a look at photos in Adobe Digicam Uncooked displaying the Primary settings utilized.

Methodology

As described beneath, whereas among the applications can be utilized as stand-alone purposes, I examined all of them as plug-ins for Photoshop, making use of every as a sensible filter utilized to a developed uncooked file introduced into Photoshop as a Digicam Uncooked good object.

Most of those applications state that higher outcomes is likely to be obtainable by utilizing the stand-alone app on unique uncooked recordsdata. However for my private workflow I choose to develop the uncooked recordsdata with Adobe Digicam Uncooked, then open these into Photoshop for stacking and layering, making use of any additional noise discount or sharpening as non-destructive good filters.

Many astrophotographers additionally select to stack unedited unique photos with specialised stacking software program, then apply additional noise discount and modifying later within the workflow. So my workflow and take a look at procedures replicate that.

Nevertheless, the exception is DxO’s PureRAW2. It might work solely on uncooked recordsdata as a stand-alone app, or as a plug-in from Adobe Lightroom. It doesn’t work as a Photoshop plug-in. I examined PureRAW2 by dropping uncooked Canon .CR3 recordsdata onto the app, then exporting the outcomes as uncooked DNG recordsdata, however with the identical settings utilized as with the opposite uncooked recordsdata. For the nightscape and wide-field photos taken with lenses in DxO’s intensive database, I used PureRAW’s lens corrections, not Adobe’s.

As proven above, I selected three consultant photos:

  • A nightscape with star trails and an in depth foreground, at ISO 1600.
  • A large-field deep-sky picture at ISO 1600 with an 85mm lens, with very tiny stars.
  • A detailed-up deep-sky picture taken with a telescope and at a excessive ISO of 3200, displaying thermal noise sizzling pixels.

Every is a single picture, not a stack of a number of photos.

Earlier than making use of the noise discount, the uncooked recordsdata obtained simply fundamental colour corrections and a distinction increase to emphasise noise all of the extra.

The Contenders

Within the take a look at outcomes for the three photos, I present the unique uncooked picture, plus a model with noise discount and sharpening utilized utilizing Adobe Digicam Uncooked’s personal sliders, with luminance noise at 40, colour noise at 25, and sharpening at 25.

I exploit this as a base comparability, because it has been the noise discount I’ve lengthy utilized to pictures. Nevertheless, ACR’s routine (additionally present in Adobe Lightroom) has not modified in years. It’s good, however it’s not AI.

The brand new good AI applications ought to enhance upon this. However do they?

Please notice:

  • I’ve kept away from offering costs and explaining shopping for choices, as frankly, some may be complicated!
  • For these particulars and for trial copies, go to the software program’s web site by clicking on the hyperlink within the header product names beneath.
  • All applications can be found for Home windows and macOS. I examined the latter variations.
  • I’ve not offered tutorials on the best way to use the software program; I’ve simply reported on their outcomes. For troubleshooting their use, please seek the advice of the software program firm in query.
ON1 NoNoise 2023’s management interface.

ON1’s principal product is the Lightroom/Photoshop various program referred to as ON1 Photograph RAW, which is up to date yearly to main new variations. It has full cataloging choices like Lightroom and picture layering like Photoshop. Its Edit module comprises the NoNoise AI routine. However NoNoise AI may be bought as a stand-alone app that additionally installs as a plug-in for Lightroom and Photoshop. It’s what I examined right here. The newest 2023 model of NoNoise AI added ON1’s new Tack Sharp AI sharpening routine.

Model examined: 17.0.1

Topaz DeNoise AI’s four-pane view to pick out the most effective AI mannequin.

This program has confirmed extremely popular and has been adopted by many photographers – and astrophotographers – as a vital a part of an modifying workflow. It performs noise discount solely, providing a alternative of 5 AI fashions. Auto modes can select the fashions and settings for you primarily based on the picture content material, however you’ll be able to override these by adjusting the energy, sharpness, and restoration of unique element as desired.

A separate program, Topaz Sharpen AI, is particularly for picture sharpening, however I didn’t take a look at it right here. Topaz Gigapixel AI is for picture resizing.

Model examined: 3.7.0

Topaz Photograph AI’s management interface for its three principal features: noise, sharpening and upscaling.

In 2022 Topaz launched this new program which contains the trio of noise discount, sharpening, and picture resizing in a single bundle. Like DeNoise, Sharpen, and Gigapixel, Photograph AI works as a stand-alone app or as a plug-in for Lightroom and Photoshop. Photograph AI’s Autopilot routinely detects and applies what it thinks the picture wants. Whereas it’s doable to regulate settings, Photograph AI affords a lot much less management than DeNoise AI and Topaz’s different single-purpose applications.

As of this writing in November 2022, Photograph AI is having fun with nearly weekly updates and appears to be the place Topaz is focusing its improvement and advertising efforts.

Model examined: 1.0.9

Luminar Neo’s Edit interface with selections of many filters and results, together with Noisemuch less AI.

Not like the opposite noise discount applications examined right here, Luminar Neo from the software program firm Skylum is a full-featured picture modifying program, with an emphasis on one-click AI results. A type of is the brand new Noisemuch less AI, accessible as an extra-cost extension to the principle Neo program, both as a one-time buy or by annual subscription. Noisemuch less AI can’t be bought by itself. Nevertheless, Neo with most of its extensions does work as a plug-in for Lightroom and Photoshop.

Being new, Luminar Neo can also be up to date often, with extra extensions coming within the subsequent few months.

Model examined: 1.5.0

DxO PureRAW’s easy interface with few selections for Noise Reduction settings.

Like ON1, DxO makes a full-featured various to Adobe’s Lightroom for cataloging and uncooked improvement referred to as DxO PhotoLab, in model 6 as of late 2022. It comprises DxO’s Prime and DeepPrime noise discount routines. Nevertheless, as with ON1, DxO has spun off simply the noise discount and lens correction elements of PhotoLab right into a separate program, PureRAW2, which runs both as a stand-alone app or as a plug-in for Lightroom – however not Photoshop, as PureRAW works solely on unique uncooked recordsdata.

Not like all the opposite applications, PureRAW2 affords primarily no choices to regulate settings, simply the choice to use, or not, lens corrections, and to decide on the output format. For this testing, I utilized DeepPrime and exported out to DNG recordsdata.

Model examined: 2.2

Noise Terminator’s controls permit adjusting energy and element.

Not like the opposite applications examined, NoiseXTerminator from astrophotographer Russell Croman is designed particularly for deep-sky astrophotography. It installs as a plug-in for Photoshop or Affinity Photograph, however not Lightroom. It’s also accessible beneath the identical bought license as a “course of” for PixInsight, a complicated program standard with astrophotographers, as it’s designed only for modifying deep-sky photos.

I examined the Photoshop plug-in model of Noise XTerminator. It receives occasional updates to each the precise plug-in and separate updates to the AI module.

Model examined: 1.1.2, AI mannequin 2

Nightscape Check

As with the opposite take a look at photos, the panels present a extremely magnified part of the picture, indicated within the inset. I shot the picture of Lake Louise in Banff, Alberta with a Canon RF15-35mm lens on a 45-megapixel Canon R5 digicam at ISO 1600.

The take a look at outcomes on a pattern nightscape.
  • Adobe Digicam Uncooked’s fundamental noise discount did job, however like all common routines it does soften the picture as a by-product of smoothing out high-ISO noise.
  • ON1 NoNoise 2023 retained panorama element higher than ACR however softened the star trails, regardless of my including sharpening. It additionally produced a considerably patchy noise smoothing within the sky. This was with Luminosity backed off to 75 from the auto setting (which all the time cranks up the extent to 100 whatever the picture) and with the Tack Sharp routine set to 40 with Micro Distinction at 0. It left a uniform pixel-level mosaic impact within the shadow areas. Regardless of the brand new Tack Sharp choice, the picture was softer than with final 12 months’s NoNoise 2022 model (not proven right here as it’s not accessible) which produced higher shadow outcomes.
  • Topaz DeNoise AI did a greater job than NoNoise retaining the sharp floor element whereas smoothing noise, all the time extra apparent within the sky in such photos. Even so, it additionally produced some patchiness, with some areas displaying extra noise than others. This was with the Commonplace mannequin set to 40 for Noise and Sharpness, and Recuperate Particulars at 75. I present the opposite mannequin variations beneath.
  • Topaz Photograph AI did a poor job, producing plenty of noisy artifacts within the sky and an over-sharpened foreground riddled with colourful speckling. It added noise. This was with the Regular setting and the default Autopilot settings.
  • Noisemuch less AI in Luminar Neo did a good job smoothing noise whereas retaining, certainly sharpening floor element with out introducing ringing or colourful edge artifacts. The sky was left with some patchiness and uneven noise smoothing. This was with the urged Center setting (vs Low and Excessive) and default ranges for Noise, Element, and Sharpness. Nevertheless, I do like Neo (and Skylum’s earlier Luminar AI) for including different ending results to pictures comparable to Orton glows.
  • DxO PureRAW2 did easy noise very properly whereas enhancing sharpness quite a bit, nearly an excessive amount of, although it didn’t introduce apparent edge artifacts. Be mindful it affords no probability to regulate settings, aside from the mode – I used DeepPrime vs the conventional Prime. Its principal downside is that in making the conversion again to a uncooked DNG picture it altered the looks of the picture, on this case darkening the picture barely. It additionally made some faint star trails look wiggly!
  • Noise XTerminator actually smoothed out the sky and did so very uniformly with out doing a lot hurt to the star trails. Nevertheless, it smoothed out floor element unacceptably, not shocking given its specialised coaching on stars, not terrestrial content material.

Conclusion: For this picture, I’d say Topaz DeNoise AI did the most effective, although not excellent, job.

This was shocking, as assessments I did with earlier variations of DeNoise AI confirmed it leaving many patchy artifacts and coloured edges in locations. Frankly, I used to be delay utilizing it. Nevertheless, Topaz has improved DeNoise AI rather a lot.

Why it really works so properly, when Topaz’s newer program Photograph AI works so poorly, is tough to grasp. Absolutely they use the identical AI code? Apparently not. Photograph AI’s noise discount is just not the identical as DeNoise AI.

Equally, ON1’s NoNoise 2023 did a worse job than their older 2022 model. One can assume its efficiency will enhance with updates. The difficulty appears to be with the brand new Tack Sharp addition.

NoiseXTerminator is likely to be a good selection for decreasing noise in simply the sky of nightscape photos. It’s not appropriate for foregrounds.

Large-Subject Picture Check

I shot this picture of Andromeda and Triangulum with an 85mm Rokinon RF lens on the 45-megapixel Canon R5 on a star tracker. Stars at the moment are factors, with small ones simply mistaken for noise. Let’s see how the applications deal with such a picture, zooming right into a tiny part displaying the galaxy Messier 33.

The take a look at outcomes on a pattern wide-field deep-sky picture.

Conclusion: The clear winner was NoiseXTerminator.

Topaz DeNoise was a good second place, performing higher than it had performed on such photos in earlier variations. Even so, it did alter the looks of faint stars which could not be fascinating.

ON1 NoNoise 2023 additionally carried out fairly properly, with its softening of brighter stars but sharpening of fainter ones maybe acceptable, even fascinating for an impact.

Telescopic Deep-Sky Check

I shot this picture of the NGC 7822 complicated of nebulosity with a SharpStar 61mm refractor, utilizing the red-sensitive 30-megapixel Canon Ra and with a narrowband filter to isolate the crimson and inexperienced mild of the nebulas.

Once more, the take a look at picture is a single uncooked picture developed solely to re-balance the colour and increase the distinction. No darkish frames have been utilized, so the 8-minute publicity at ISO 3200 taken on a heat night time reveals thermal noise as single “sizzling pixel” white specks.

The take a look at outcomes on a pattern deep-sky close-up.
  • Adobe Digicam Uncooked did job smoothing the worst of the noise, suppressing the recent pixels however solely by advantage of it softening all the picture barely on the pixel stage. Nevertheless, it leaves most stars intact.
  • ON1 NoNoise 2023 additionally did job smoothing noise whereas additionally seeming to spice up distinction and construction barely. However as within the wide-field picture, it did easy out star photos slightly, although considerably photogenically, whereas nonetheless emphasizing the faintest stars. This was with no sharpening utilized and Luminosity at 60, down from the default 100 NoNoise applies with out fail. One wonders if it truly is analyzing photos to provide optimum settings. With no Tack Sharp sharpening utilized, the outcomes on this picture with NoNoise 2023 regarded similar to NoNoise 2022.
  • Topaz DeNoise AI did one other good job smoothing noise whereas leaving most stars unaffected. Nevertheless, the faintest stars and sizzling pixels have been sharpened to be extra seen tiny specks, maybe an excessive amount of, even with Sharpening at its lowest stage of 1 in Commonplace mode. Low Mild and Extreme modes produced worse outcomes, with plenty of mottling and unevenness within the background. Not like NoNoise, no less than its Auto settings do differ from picture to picture, providing you with some assurance it truly is responding to the picture content material.
  • Topaz Photograph AI once more produced unusable outcomes. Its Regular modes produced plenty of mottled texture and haloed stars. Its Sturdy mode proven right here did easy noise higher, however nonetheless left plenty of uneven artifacts, as DeNoise AI did in its early days. It actually looks like Photograph AI is utilizing outdated hand-me-down code from DeNoise AI.
  • Noisemuch less AI in Luminar Neo did easy noise however inconsistently, leaving plenty of textured patches. Stars had grainy halos and this system elevated distinction and saturation, changes normally greatest left for particular adjustment layers devoted to the duty.
  • DxO PureRAW2 did easy noise very properly, together with wiping out the faintest specks from sizzling pixels, but it surely additionally worn out the faintest stars, I believe unacceptably and greater than different applications like DeNoise AI. For this picture, it did depart fundamental brightness alone, possible as a result of it couldn’t apply lens corrections to a picture taken with unknown optics. Nevertheless, it added an odd pixel-level mosaic-like impact on the sky background, once more unacceptable.
  • Noise XTerminator did an amazing job smoothing random noise with out affecting any stars or the nebulosity. The Element stage of 20 I used really emphasised the faintest stars, but additionally the recent pixel specks. NoiseXTerminator can’t be counted on to get rid of thermal noise; that calls for the applying of darkish frames and/or utilizing dithering routines to shift every sub-frame picture by a number of pixels when autoguiding the telescope mount. Even so, Noise XTerminator is so good customers may not have to take and stack as many photos.

Conclusion: Once more, the winner was NoiseXTerminator.

Deep-sky photographers have praised “NoiseX” for its effectiveness, both when utilized early on in a PixInsight workflow or, as I do in Photoshop, as a sensible filter to the bottom stacked picture underlying different adjustment layers.

Topaz DeNoise can also be a good selection as it might work properly on many different varieties of photos. However once more, play with its varied fashions and settings. Pixel peep!

ON1 NoNoise 2023 did put in a good efficiency right here, and it’ll little question enhance – it had been out lower than a month once I ran these assessments.

Based mostly on its odd conduct and leads to all three take a look at photos I might not suggest DxO’s PureRAW2. Sure, it reduces noise fairly properly, however it might alter tone and colour within the course of, and add unusual pixel-level mosaic artifacts.

Evaluating DxO and Topaz Choices

DxO and Topaz DeNoise AI provide probably the most selections of AI fashions and energy of noise discount. Right here I examine:

  • Topaz DeNoise AI on the nightscape picture utilizing three of its fashions: Commonplace (which I used within the comparisons above), plus Low Mild, and Extreme. These present how the opposite fashions didn’t do nearly as good a job.
  • The set beneath additionally compares DeNoise AI to Topaz’s different program, Photograph AI, to point out how poor a job it’s doing in its early kind. Its Sturdy mode does easy noise however over-sharpens and leaves edge artifacts. Sure, Photograph AI is one-click and simple to make use of, however produces dangerous outcomes – no less than on astrophotos.
Evaluating DeNoise’s and Photograph AI’s completely different mannequin settings.

As of this writing DxO’s PureRAW2 affords the Prime and newer DeepPrime AI fashions – I used DeepPrime for my assessments.

Nevertheless, DxO’s dearer and full picture processing program, PhotoLab 6, additionally affords the even newer DeepPrimeXD mannequin, which guarantees to protect or recuperate much more “Xtra Element” over the DeepPrime mannequin. As of this writing, the XD mode is just not supplied in PureRAW2. Maybe that may watch for PureRAW3, little question a paid improve.

Evaluating DxO’s varied Prime mannequin settings. DeepPrimeXD is barely in PhotoLab 6.
  • The set above compares the three noise discount fashions of DxO’s PhotoLab 6. DeepPrime does do a greater job than Prime. DeepPrimeXD does certainly sharpen element extra, however on this instance it’s too sharp, displaying artifacts, particularly within the sky the place it’s including buildings and textures that aren’t actual.
  • Nevertheless, when used from inside PhotoLab 6, the DeepPrime noise discount turns into extra usable. PhotoLab is then getting used to carry out all of the uncooked picture processing, so PureRAW’s alteration of colour and tone is just not a priority. Conversely, it might additionally output uncooked DNGs with solely noise discount and lens corrections utilized, primarily performing the identical duties as PureRAW. In case you have PhotoLab, you don’t want PureRAW.

Evaluating AI to Older Non-AI Programs

The brand new technology of AI-based applications has garnered all the eye, leaving older stalwart noise discount applications trying slightly forlorn and forgotten.

Right here I examine Digicam Uncooked and two of the most effective AI applications, Topaz DeNoise AI and NoiseXTerminator, with two of probably the most revered of the “old-school” non-AI applications:

Nik Dfine2’s management interface.
  • Dfine2, included with the Nik Assortment of plug-ins bought by DxO (proven above), and
  • Cut back Noise v9 bought by Neat Picture (proven beneath).
Neat Picture’s Cut back Noise management interface – the easy panel.

I examined each by utilizing them of their computerized modes, the place they analyze a bit or sections of the picture and regulate the noise discount accordingly, however then apply that setting uniformly throughout your entire picture. Nevertheless, each permit handbook changes, with Neat Picture’s Cut back Noise providing a bewildering array of technical changes.

How do these older applications stack as much as the brand new AI technology? Listed here are comparisons utilizing the identical three take a look at photos.

Evaluating outcomes with Neat Picture and Nik Dfine2 on the nightscape take a look at picture.

Within the nightscape picture, Nik Dfine2 and Neat Picture’s Cut back Noise did properly, producing uniform noise discount with no patchiness. However the outcomes weren’t considerably higher than with Adobe Digicam Uncooked’s built-in routine. Like ACR, each non-AI applications did easy element within the floor, in comparison with DeNoise AI which sharpened the mountain particulars.

Evaluating outcomes with Neat Picture and Nik Dfine2 on the wide-field take a look at picture.

Within the tracked wide-field picture, the variations have been tougher to tell apart. None carried out as much as the usual of Noise XTerminator, with each Nik Dfine2 and Neat Picture softening stars slightly in comparison with DeNoise AI.

Evaluating outcomes with Neat Picture and Nik Dfine2 on the deep-sky take a look at picture.

Within the telescopic deep-sky picture, all applications did properly, although none matched NoiseXTerminator. None eradicated the recent pixels. However Nik Dfine2 and Neat Picture did depart needed particulars alone and didn’t alter or get rid of desired content material. Nevertheless, additionally they didn’t get rid of noise in addition to did Topaz DeNoise AI or NoiseXTerminator.

The AI know-how does work!

Your Outcomes Could Differ

I ought to add that the character of AI implies that the outcomes will definitely differ from picture to picture.

As well as, with many of those applications providing a number of fashions and settings for energy and sharpening, outcomes even from the identical program may be fairly completely different. On this testing, I used both this system’s auto defaults or backed off these defaults the place I assumed the impact was too robust and detrimental to the picture.

Software program can also be a consistently shifting goal. Updates will alter how these applications carry out, and we hope for the higher. For instance, two days after I revealed this take a look at, ON1 up to date NoNoise AI to v17.0.2 with minor fixes and enhancements.

And do bear in mind I’m testing on astrophotos and pixel peeping to the acute. Rave critiques claiming how properly even the poor performers right here work on “regular” photos would possibly properly be legitimate.

That is all by means of saying, your mileage might differ!

So don’t take my phrase for it. Most applications (Luminar Neo is an exception) can be found as free trial copies to check out in your astro-images and in your most well-liked workflow. Check for your self. However do pixel peep. That’s the place you’ll see the issues.

What About Adobe?

Within the race for AI supremacy, one wonders the place Adobe is within the discipline.

Within the final couple of years, Adobe has launched a number of superb and highly effective “Neural Filters” into Photoshop, which work wonders with one click on. And Lightroom and Digicam Uncooked have obtained highly effective AI-based choice and masking instruments far forward of a lot of the competitors, with solely Luminar Neo and ON1 Photograph RAW coming shut with related auto-select capabilities.

Neural community Noise Reduction is coming to Photoshop. At some point!

However AI noise discount? You’d suppose it could be a excessive precedence.

A neural filter for Noise Reduction is on Adobe’s Wait Record for improvement, so maybe we are going to see one thing within the subsequent few months from Adobe to compete with the AI choices of Topaz, ON1, and Luminar/Skylum.

Till then now we have plenty of selections for third-party applications that each one enhance with each replace. I hope this assessment has helped you make a alternative.


Concerning the writer: Alan Dyer is an astronomy images and writer. You could find extra of his work and writing at his web site, The Wonderful Sky. This text was additionally revealed right here.