Adobe Lightroom 12.3 has AI denoise function

Not sure if processing software news should be in this section or on the site at all - please feel free to remove!

I have this morning updated my Adobe Lightroom Classic to version 12.3 and found that it has an AI DeNoise facility now in the Photo>Enhance menu. Having found that the OM Workspace version of denoise requires a higher grade of graphics card than installed on my PC, I was pleasantly surprised to find that the LR version does work, and to my eye makes a visible improvement, albeit to a photo with not too much original noise.

Anyone else tried it yet?

Sorry, was not able to upload full sized version

I updated mine, tried the AI NR on an ISO 400 photo that had noise in the deep shadows. It seemed to do a fair job, looked a little mushier than I’m used to with Deep Prime. Will compare to the other NR software I have when I get time.

I tried it, too, this morning. Awfully slow on my PC (i5, 16GB). Results are much better and faster with Topaz DeNoiseAI or Topaz PhotoAI. It seems that Adobe don’t want to lag behind the other software producers.

I use PureRAW 2 on a Ryzen 7 with integrated Radeon GPU and it has acceptable performance - about 15s on a noisy 20MP file using DeepPrime. It does a good job but annoyingly doesn’t save to the original directory so Lightroom can’t stack with the original unless I manually move it. Annoying.

I tried the new Lr denoise and man is it slow - about 4m (yes, MINUTES) - for the same file that PureRAW did in a fraction of that time. On a plus note it puts the new file in the right place and optionally stacks it. Perfect.

I don’t use PureRAW often so I could live with the (lack of) speed of Adobe’s attempt… except despite its glacial performance it doesn’t do as good a job (much better than nothing though) and the file size is a bit bigger too, although that may be down to it having much more residual noise than the PureRAW version.

Bottom line: the Adobe denoise is free (if you already use Lightroom Classic) and I’d certainly use it if that was all I had available, but I don’t think it’s going to worry DxO any time soon.

The first time I tried LR failed due to time taken, I tried again with the same photo, and it did it relatively quickly. I have noticed before with Adobe PS that the first image processed seems to taken longer than subsequent ones, particularly when resizing or ‘saving as’. I have used ON1 as a plug-in to LR for noise reduction in the past - not particularly impressed, and my ON1 is now not updated as is an older version. Cannot really justify spending any more on a stand-alone NR program, so will probably stick with the LR version.

Interested to see how it compares though.

1 Like

I tried it a second time on the same image and indeed there was a difference in speed - but it still took over two minutes.

The good part is that I can more or less match the NR part of DeepPrime by upping the amount to 70, and in some respects the handling of blurred detail is more realistic in Lr, but DeepPrime is generally better at recovering real detail if we ignore the fake interpolated stuff it also produces.

I think a bit more experimentation is in order but I’m beginning to like the Adobe rendering overall. Speed (or lack of) is the big fly in the ointment; probably not an issue if you have a souped-up PC but a bit of an annoyance for many of us.

1 Like

I’ve been playing with a photo that was shot at ISO 25,600 with my OM-1. I think Deep Prime might have a slight advantage in detail retention at about the same amount of NR, but it’s really tricky to equate. My workflow for high noise includes Deep Prime XD so I’ll probably keep on using it, but the LR tool is pretty impressive. Both are slow on my PC, probably not appreciably different.

1 Like

Topics about software are very much welcome here! Thank you for posting.

Hi folks,

This is my personal opinion based on testing on my favorite object, a photo taken in bright sunlight with the Pen-F camera set to ISO 25600 by mistake. Thus, it contains a lot of noise. Over the years I have used this for testing noise reduction in LrC as well as in Topaz Denoise and later Photo AI. The attached screenshot contains from left to right jpg exports of the new AI denoise function of LrC, then Photo AI version and finally the original photo without adjustments.
Photo AI seemingly does a better job than LrC AI, however I am not convinced. As usual, Photo AI has a tendency of making skin tones look plasticcy and removes fine details visible on the shirt and the neck strap. I tried both settings, Standard and Strong. Standard ails to do a proper job on this photo, and Strong overdoes even when the handles are pulle all the way to the left.
LrC AI on the other hand still looks somewhat grainy, but retains fine details and skin tones to a much better degree, in my opinion.
Therefore, my preference for this particular photo is the results fron LrC’s new AI denoise function.
For other pictures, the results may be different.
I don’t have access to e.g. DxO.


I tried XD in a version 3 trial of PueRAW and found it very slow on my PC, about the same as the Adobe denoise so that matches your experience. I prefer a bit of monochrome noise in images - it tends to enhance the appearance of detail - and Adobe’s default setting does a good job in this respect.

The overall “look” of the image is different compared with DeepPrime, better in some respects and worse in others, and very close inspection shows more accurate real detail reconstruction with fewer artefacts in the Adobe software which is probably why it can look a little softer by comparison. The only imaging flaw I found is related to colour - it can sometimes fill small white highlight areas with colour from any slight fringing, something the DxO software never does.

It’s a good start from Adobe and I’m sure it will only get better.

1 Like

Interesting comparison, thanks. I agree, in this instance LR comes out best.

1 Like

Definitely a good start! After a lot of pixel peeping, the only difference I was really sure of was that Deep Prime XD removed much more noise and apparently obliterated more detail than LR AI when the sliders were each at 50%. :slight_smile:

The more I succeeded in getting a similar amount of NR, the more I appreciated the LR AI tool. However, as you know, it can take a long time on a slow machine to render a denoised image with both LR and Deep Prime/XD and I spent far too much time trying to match the NR results in order to see how much detail was left. I also tried to determine if one or the other tool gave a better final result after sharpening and other edits in a typical workflow but in the end I was getting bleary eyed and decided to call it a draw and have a cocktail.

I’ll probably keep sending noisy images to PhotoLab 6 Deep Prime XD simply because I’m more familiar with how to tweak it.

1 Like

I have an Intel i7-12700K CPU also with integrated graphics and no discrete GPU. Mine also takes about 4m 15sec to do Denoise. I can live with that… for the moment. I’m amazed that it actually works at all with an integrated CPU. OM Workspace wouldn’t even let me download the plugin. So glad I now have AI NR on the one app. Makes workflow a snack… apart from the time… which can easily be fixed with a moderate GPU… if so inclined.

1 Like

Adobe software tends to be agnostic about GPUs but disables them by default if it suspects potential issues. Sometimes forcing Lr to use the GPU for some or all purposes (in “Preferences”) sometimes does cause issues, sometimes not, so it’s always worth trying and there is some granularity to the setting

Given the almost infinite permutations of hardware that Windows software has to cope with it’s not surprising that some developers choose the lazy option of not even attempting to run on hardware that doesn’t match whatever they tested against.

I’m no fan of Adobe but credit where it’s due - they are pretty good at what they do and rarely introduce major features that break existing usage; the new stuff sits there waiting to be used but all the old stuff continues to work as though nothing had changed. That’s why I still use Lightroom: evolution rather than revolution is baked into the design.


I have tested LR AI on an older photo which is both very soft and quite noisy (taken with E-M1 Mark II at ISO 1250 with 100-400). A few weeks ago I have started to use Topaz Sharpen / Denoise.

I am impressed at what LR did compared to the Topaz version. Both have their merits, in this case the photo in question in the LR version comes out a tack sharper and more natural than when treated with Topaz. This of course may be down to my lack of experience with Topaz :innocent: as I have not used it very often yet.

And yes, LR denoise is quite slow, but bearable, less than 1 minute (Mac Mini M1). As photography is my hobby, I am not bothered by its lack of speed in this function - apart from it LR works fast. And I admit I am a LR enthusiast from the start, also for the reasons JohnB explains in the above post. It took me a long time to accept the subscription policy, but now that I have it (it’s LR, Bridge and PS in one parcel for a very reasonable fee of CHF 12 / month), I am perfectly happy. As LR does what I need it to do, every new function is a perk rather than fix of a fault.

1 Like

I am running a Windows 11 i7 with 64 GB RAM and a NVIDIA RTX 2060 graphics card and LR Denoise takes only about 15 seconds. At the moment I still use Topaz Photo Ai because it also provides AI sharpening.


That’s useful information. Sounds like denoise is leaning heavily on the GPU much more than other Lr tools, so a “proper” GPU is needed if speed is important. Pretty much the same with DeepPRIME XD but not the original DeepPRIME which works fast on any GPU, even integrated laptop types.

1 Like

Adobe have not updated their system requirements for LrC since December 2022. It surprises me they didn’t do it with the release of Denoise with its heavy reliance of GPU power. Perhaps that’s still to come but I wish it comes soon. I could be addicted to Denoise and itching to install a discreet GPU. My iGPU takes 4+ minutes to denoise a 20MB image running at 100%. That’s just a bit too long.