Lightroom Classic is playing catch-up. AI-based masking had been around for a long time in programs like ON1 Photo Raw, but has only recently been released in Lightroom. Now, Adobe has turned its hand to noise reduction. Does it stand up to scrutiny?
There is already an array of AI-based noise reduction programs; Topaz DeNoise AI, ON1 NoNoise AI, and DxO Deep Prime all performed much better than Adobe's appallingly bad muddy noise reduction results. Those applications worked well as plugins for Lightroom and Photoshop. In the case of DXO PhotoLab 6 and ON1 Photo Raw, they have also integrated AI noise reduction into their own asset management and raw development software. Even camera brands’ proprietary software provides its own AI noise reduction; OM Workspace, growing in popularity with OM System users, has AI noise reduction built in.
All those programs work well. In the past, I’ve run numerous tests, running the same raw files through each, and every one of them does a more than adequate job. Some are slightly better than others, but those differences were more to do with speed and usability than noise reduction performance.
I’ve just tested Lightroom’s new AI-driven noise reduction alongside other programs to see how well that would do compared to the other tools.
A Worry About Monopoly
Given that many people use these AI-based programs as plugins for Lightoom and Photoshop, there is concern that Adobe is creating a monopoly, as these third-party plugins will become redundant.
However, there’s more to processing than noise reduction. The other standalone programs, like DxO Photolab 6, ON1 Photo Raw, and Capture One, are considered with high regard by their users, as they produce very different development results that many of their users prefer.
Meanwhile, Topaz Denoise, which is not an integral part of an asset management and raw development suite, is not the only tool in that company’s arsenal. Moreover, it will still be a popular plugin with Capture One owners even if it becomes redundant for Lightroom users; Capture One does not currently have AI-based noise reduction.
Closing the Stable Door After the Horse Has Bolted?
Please excuse the mixing of metaphors, but Adobe’s horse is not only escaping from the stable, but people stopped watching the race it should have been running in.
My first DSLR I could use up to ISO 400 without the photos becoming too noisy to use. With my second, that leapt to ISO 800. I swapped to mirrorless Micro Four Thirds, and that doubled again to ISO 1,600, and then to ISO 3,200. I can use my OM-1 at ISO 6,400 or even 12,800 without thinking of applying noise reduction.
Even increasing shadows by two stops, noise is not an issue for me. Furthermore, with my newest camera, I can shoot long exposures without having noisy images.
Modern cameras perform better than I ever need, and I suspect that is true for many others too. The advancements in sensor technology mean noise reduction is becoming redundant for photographers.
Adobe should have introduced this feature three years ago. Nevertheless, it will be a welcome addition for those who use Lightroom, especially if they use older cameras and shoot moving subjects in low light.
Performance of Lightroom’s Denoise
There were a couple of issues I found. Firstly, I was disappointed that the preview image did not fill the entire screen, something that does happen with ON1 and Topaz, although it is a complaint I've heard with Deep Prime. Worse than that, though, Lightroom's preview was unclear, so it was hard to judge the final outcome after processing. All the other programs I have used give a clear and precise preview of the final image.
Secondly, I have an old computer; I built it about eight years ago. Nevertheless, it has a quad-core Radeon R7 processor that runs at 3.4 GHz, 32 GB of RAM, and solid-state hard drives. Despite its age, it runs pretty quickly for most tasks. ON1 applies noise reduction almost immediately, DxO PhotoLab 6 Deep Prime takes a little longer, Topaz Denoise longer still, but they still work in seconds. Lightroom took 14 minutes to denoise some of the images. Those with newer machines will cut this time down considerably, but it shows that the Adobe noise reduction engine is slower than others I use. Photographers with older machines might think twice before employing it that often.
The Tests
My tests were shot using different ISOs, conditions, and exposure settings, so I would get a variety of images to try. I'm using one image to illustrate my findings. It was shot at ISO 6,400, 1/12,800 second, at f/5. I used aperture priority and dialed in -1 EV.
Before noise reduction, I first zoomed in to 100% in Lightroom and then cropped it to about the same dimensions, I then applied Lightroom's Denoise at the default values.
At first, I had turned off sharpening because I always found Lightroom’s algorithm to be far too aggressive. But I thought the AI's result looked a bit soft. So, I increased the sharpening to 40, yet still, the result looked overly muddy. By turning the sharpening all the way up to maximum and increasing the radius to 1.5, I was able to achieve a more pleasing result, but it was still not perfect.
Here is the same image with noise removed using other programs.
The differences in these results are seen at 100% and are minor. I reiterate that none of the AI programs produced bad results.
Pushing the Noise Reduction to Its Limits
I was eager to push the software further. So, I found the noisiest photo I had in my catalog. It was one I shot for the sole purpose of testing noise reduction software over four years ago, long before I started writing for Fstoppers.
It was an 1,800-second exposure taken at night at the end of an extended long-exposure photoshoot using an Olympus OM-D E-M1 Mark II, which has since had two generational replacements.
Prior to taking this photo, I have been shooting long exposures for over an hour. Consequently, the sensor was warmed up, and there was lots of color noise and plenty of hot pixels too.
The lens I used was an old Four-Thirds lens, which showed more distortion at the edges and was less sharp than the newer Micro Four-Thirds lenses I use now. This would have been a challenging test for any software. Lightroom’s Traditional noise reduction barely touched the color noise and hot pixels on this shot, so I was keen to see how it would perform.
Lightroom Classic AI Denoise did good work with noise reduction. On close inspection, it did remove more hot-pixel artifacts than other programs, but nowhere near all of them. There were many white speckles left on the image that I can see when pixel peeping. It also did a good job of removing the color noise, even if, once again, it left the image looking a bit muddy.
Lightroom wasn't as good as DxO Photolab 6 at correcting lens distortion, and it lost some of the shadow details.
This was an extreme test. None of the programs were able to transform this very poor-quality image into a wall hanger. But it did give an idea of what could be achieved. Similarly, with all these tests, I was pixel-peeping, zooming in to 100%, something that is rarely necessary.
A Quick Word about Super Resolution
The Denoise function also comes with a Super Resolution option that allows you to increase the number of pixels in an image. As yet, I haven't tested that, and that will be a topic for a whole other article.
What I Liked and What Could Be Improved
Lightroom’s Denoise didn’t do a bad job. This new feature be a welcome addition for Adobe's dedicated users, especially if they own older cameras which are more prone to noise than newer models. It will also please those who habitually shoot at high ISO settings.
I particularly liked that the denoised image is a DNG file. However, this does mean that one cannot synchronize the noise reduction across multiple images.
Something to remember is that this is just one element of an entire package. Lightroom is a great catalog-based asset management program, arguably the best available. It also does a reasonably good job of developing raw images, and it has a superb user interface for doing that.
However, although its raw development results aren’t bad, they are also not always the best available. This is, of course, a subjective point of view and you are welcome to disagree with that opinion. The inclusion of AI-driven noise reduction makes the program better than it was before. Nevertheless, I find DxO Photolab 6 development results and noise reduction to be better in most circumstances. But I am unsure whether there is a big enough difference to entice users who are happy with Lightroom away from it.
But for those who have invested in other programs, this new feature is unlikely to entice them back to Lightroom. For example, for the last few months, I’ve adopted DxO Photolab 6 into my workflow, and am unlikely to switch back to Adobe because it has AI Noise reduction.
Lightroom Classic’s noise reduction was very slow compared to other programs, but this may be less of an issue for those with new turbocharged machines.
In short, it will be a welcome addition to those dedicated Adobe users, but the question remains whether it is necessary. On the following cropped image shot at ISO 2,500, can you tell which has noise reduction applied and which hasn't?
It does take a long time to run on my computer. The main issue I have seen is that if I try to run a second image, I get an error after several minutes. To deal with this I have to close and relaunch Lightroom if I want to denoise another image.
The results I have found to be pretty good for some wildlife images I took. I didn't like the way it handled night sky photos.
I have crops of before and after images of a noisy photo of a condor shot at ISO9000. It was a dark and completely overcast day.
May I ask what camera you're shooting? My reason for asking is that I shoot raw files on a Sony a7III (24mp sensor uncompressed raws), my wife shoots raw files on a Sony a6500 (24mp sensor compressed raws), we each have our own PC, but they are identical. Her images take several minutes to denoise, mine take 30 to 45 seconds to denoise. I'm wondering why the difference and maybe, knowing what others are shooting, would help create an understanding.
Wondering if you're updating the graphics drivers (assuming you're on a PC), if you're on a Mac I don't know what to suggest :-)
Both machines were identical, drivers are fairly up to date (I'd be lying if I said we updated every time new drivers were released for the graphics cards, they do it way too often). I did find out today though that her machine is having issues with the hardware, the card is failing to load the graphics driver, so I imagine her times were without graphics acceleration. Thanks,
Nikon Z7ii, so it has 45.7 MP image. I don't know why one would take longer than the other. Are the both 14-bit? Maybe the RAW files are different in some way and they have to be processed differently.
I also have a Z50 with a 20MP sensor, maybe I should perform a speed comparison.
If you do compare, I would love to know.
Slight update, today she got a message that her graphics card wasn't found and I had a look at her machine...the graphics driver has been failing to load for some time now. I think I have the 'root cause' figured out. On the bright side, she'll be getting a new card, on the bad side, my machine will be the "old" one. :)
Sorry that your attention span couldn't cope with the additional information.
Perhaps try taking some photos, it does wonders for cognitive abilities. It would be great to see your gallery populated.
Hi Ivor, you've got a point about the pixelation (minor annoyance) but you need a faster GPU.
PC I tested with: Custom build 10th gen i9 (10 physical cores + hyperthreading, but now "elderly") and an RTX 3060 GPU.
Images: 2 from a Sony A7R III. To understand the relative complexity for the AI, one of an Elegant Bronzeback, the other a dawn image of river in a city, lots of shadow. Both high ISO.
The CPU barely twitches but the GPU maxes out. Total time for each image was 22-24 seconds. For reference, Topaz took 8 seconds for the Bronzeback.
As for muddiness, I wouldn't call the image muddy but I would say soft. Topaz gave me almost identical results if I turned the sharpening to 1. Personally, if I decided to use LightRoom AI denoise, I'd run sharpening as a separate process - Less to go wrong.
I can't see myself using LR AI denoise, but there are times that the work you need to do to get an acceptable result from Topaz could tempt me into trying it.
Thanks Jon.
Adobe from my experience always been a Ram Hog. ugg
If working with large files or many layers (basically anything where your work in progress will result in a .psb file, then photoshop can use a lot of RAM. Though if can be limited at the cost of extra writes on the SSD and slower performance.
Ideally,it is best to just have extra RAM. Photoshop, premiere CC, and aftereffects doesn't get much of a performance difference from slightly lower performance RAM. for example, for DDR4 there is very little performance difference (not noticeable), between DDR4 3400 at 16-16-16-36, and DDR4 3800 at 16-15-15-34, and on DDR5, you are unlikely to get much of a difference outside of margin of error between DDR5 6000, and DDR5 7000.
With all of that in mind, you can go with a cheap 64GB kit. cheap DDR4 3600 64GB kit can be had for around $100-110. and a cheap, and a 128GB kit can be had for around $230-250.
If you do a lot of other work that is memory latency and throughput sensitive, you are looking at close to $280 for a decent 64GB kit where you will get good out of the box latency, but will have ample room to overclock the RAM and tighten the timings more depending on your airflow and temperature.
I used the Adobe AI noise reduction via ACR on a iso 12800 concert photo with a lot a shadow in it. It did reduce the noise better than the noise reduction from Capture one that I normally use. But it took my computer (a 2016 MacBook Pro , the base model 13 inch) four minutes to get a result that made the person in the shot look like a doll. I preferred the image with a bit more noise in it.
--- "a result that made the person in the shot look like a doll. I preferred the image with a bit more noise in it."
There's a slider for that. Did you even try moving it to the left for a lower value?
I tried the sliders but still didn’t get a result that satisfied me. It would look okay on Instagram , but on a monitor I didn’t get a result I liked. Didn’t try other images, didn’t want to waist more time on it. When I get my new computer and the technology has taken a few more steps , I will try again
If you are content with Capture One's NR, how is Adobe's Denoise worst? From my testing, it's way better than Capture One's, and you can dial it down to Capture One's level and still be cleaner.
But, anyways. I'm not trying convince you. If you're not happy with, you're not happy with it. :)
Exactly, for now I’m happy with capture one for my work. And we al use the tools we like
That is one of the issues I noticed with the noise reduction as well, compared to others such as topaz, it struggles more with skin textures. though it has fewer artifacts of mistaking noise for texture and trying detail enhancement on it. Sadly due to how it behaves, by the time you lower the slider enough to make the skin look proper, the rest of the image will have so much noise that you might as well not waste time waiting for the noise reduction to process.
I share your opinion, I lowered the slider to about 20 and then the skin looked okay, but by then the results weren’t any better than what I can get with capture one’s NR. And the colours in the capture one version were much better.
FWIW, the Super Resolution cannot be used with Denoise. It's greyed out. There is a workaround but normally it's unavailable.
Also the speed of operation varies hugely by machine. Current Silicon Macs and PCs with good GPUs are much faster, by a lot.
Lastly, the Denoise function is a 'work in progress' and the various criticisms I see on the forums are all being addressed for future versions.
Hi Nick, yes, I appreciate that a newer machine would give me better performance, and I should be upgrading later this year. However, there are other noise reduction apps that carry out their work almost immediately on my computer. Speaking to other users with faster machines, they are noticing the difference too. Other prograpms, expecially ON1, which gives a full screen preview too, is much faster.
I have managed to get the super resolution working but not in conjunction with denoise. I'm going to do comparison tests with other options to see if there is a difference in quality. Watch this space.
As for it being a work in progress, I always think it is a disappointment when software companies produce half-cooked products that their customers are paying for. If it had remained in beta, until the issues were ironed out, I think that would have been better.
Thanks for taking the time to comment.
IMO the Super Resolution 'can' be excellent. Not on all images but for some it's magic. Sometime it more or less matches the pixel shift high-res images from my Leica SL2 (187MP).
Depending on the specs, you may only need to get a newer GPU to drastically reduce the processing time. Either that or hope that Adobe will be able to make that feature adapt to more widely supported openCL and CUDA functions, where a wider range of GPUs will be able to handle nearly all of the work.
It's not too shabby, actually.
--- "On the following cropped image shot at ISO 2,500, can you tell which has noise reduction applied and which hasn't?"
If an image is clean to begin with, there shouldn't be any discernible difference.
Sorry, wrong thread
Adobe's new Denoise takes about 15 seconds per image on my 13900k / 3080 combo.
It was good enough that I was wishing that there was an option to automatically appy it to any pic over iso 6400 or some set ISO right at import.
I usually import after the shoot and leave it alone until the next day to edit.
50mpx sony DNG
I have not tried this yet but can you set up an Adaptive ISO Preset on a per camera basis and set as an Import Preset? You can do this for normal noise reduction etc, not sure about AI Denoise.
Not as far as I can see, and you cannot sync the denoise across numerous different files. The program creates a DNG copy of the original file,
I appreciate these comments but my early experience with denoise is astounding. I mostly do astrophotography and spent a lot of time talking multiple exposures to blend in starry sky stacker or PS with median blend; you all understand I'm sure. Last week i shot at 12800 and 6400 and achieved stunning results. So now I'm doing single shot and saving myself a ton of time alone in the cold and dark (sigh, so sad when i read that back)
That's cool, David. I am not saying it does a bad job, it actually works well. But it has its functionality limitations.
The noise reduction seems to be almost exclusively designed for GPU compute (with very inefficient operation on CPU), and if you have a video card that is not fully supported, then it will take a long time, even though it will use a wide range of cards to varying degrees, in terms of performance, it seems to behave a lot like stable diffusion, or some of the AI video interpolaters, where an older gen card of similar compute performance to a new one will perform far worse. e.g., a GTX 1080ti and RTX 3060 12GB doing the same AI task, but the 1080ti offering less than half the performance in some applications.
While these tasks can all run directly on a CPU as well, the process is extremely slow, though there are some ports of them that will faster, e.g., one image generation build taking over an hour to process a single image in a 12 core CPU, and a build more optimized for the CPU taking 20 minutes, though a mid range GPU from 2013 will do the task in under 3 minutes, and a more modern GPU doing the task in less than 20 seconds.
Wanted to also add for some GPU combinations, e.g., Nvidia maxwell based GPUs, you may see a CPU top out at 50-60% usage, while the GPU tops out at 20-30% usage with an overall slow result of taking nearly 5 minutes to process a single 24 megapixel raw file (quite a low resolution). (The GPU will also not go above its base clock since the boost clock doesn't kick in at low utilization)
Another issue is it that it is hit or miss with fine detail, though this is an issue with all major AI based noise reduction applications, but Adobe tends to make mistakes where others are not, while doing fairly good results on objects with smoother and more uniform textures that is of a different size and frequency in terms of feature size than the overall noise in the image. Someone needs to get an AI model for noise reduction that has a stronger focus on preserving fine details in textures.
Thanks. I have since upgraded my computer, as I think you saw from a different article of mine. I'm now using a GTX4070 12GB VRAM, and Lightroom's noise reduction is still slower than DxO PL6's Deep Prime, and the preview is not perfect. I hope they will sort that out in due course.
With Adobe, it still seems they are still having issues with hardware acceleration where the amount of GPU offloading is inconsistent.
They need to offer more granular advanced control for simply forcing GPU use rather than dealing with adobe's arbitrary white list where they won't push a card to it's full TDP unless they pretty much tested it. While everyone else will target a CUDA or OpenCL version and every card that supports the version used, will be pushed to their limits regardless of if the company tested it.
Forgot to also mention, that sometimes a major Nvidia or AMD driver change or a beta driver can cause Adobe to not use full acceleration.
It seems to me the biggest advantage is it works on all raw files. I prefer DXO but you may have to wait a couple of months for them to produce the profiles for camera/lens. This works better but with a new camera LR AI denise works straight away.
Thanks Gordon,
Yes and no! I had to wait quite a while before LR released an update for my camera's raw files. There is a longer wait for DxO but that is because there is a more thorough testing of the lens profiles with corrections applied to each focal length for a zoom lens, as opposed to a generic correction applied to all by LR and others.
As yet, the Denoise function does not work on DNG files, nor on other file types.
I think that Photolab 5 (using DeepPrime), on portrait high iso photos, has better colour with less effort than Photoshop + camera raw 15.3 (AI denoise).
(But you need Photolab, don't use Pure raw + Photoshop)
That's interesting, Toni. Thanks for commenting.
Thanks for this review. Since April, 2023. LR CC has upgraded with a few new features included the AI Denoise and a higher end GPU now does play a MASSIVE part in it. First thing I learned is to use AI Denoise first before adding any editing or masks as it would increasing the AI Denoise process time over 400%. In search of speeding up the Denoise from my current Nvidia GTX 1060 6GB DDR5 video card I went on a little research quest on LR CC Denoise speed based on DP Review's A7RV 60MP free download files and with an DIY old Intel i7-6700K 4.0GHz, 32GB DDR4 2400MHz RAM, WD Black 1TB M2 SSD, Win10 @1440p display setup and an ancient 16 years old Antec 190 550W+650W (GPU use only) =1200W case. Here are my average test results after going through five cycles per test:
- GTX1060 6GB GDDR5: 1-pic: 159s 10-pic: 1569s idle: 108W average: 234W Peak: 279W
- RTX3060 OC 12GB GDDR6: 1-pic: 32.08s 10-pic: Not tested Power: Not tested
- RTX3060 Ti 8GB GDDRX6: 1-pic: 26.90s 10-pic: Not tested Power: Not tested
- RTX3070 OC 8GB GDDR6: 1-pic 25.18s 10-pic: 221.73s Power: 117W average: 378W Peak: 585W
- RTX4060 Ti 8GB GDDR6: 1-pic: 26.97s 10-pic: 247.62s Power: 108W average: 288W Peak: 369W
- RTX4070 12GB GDDRX6: 1-pic: 20.08s 10-pic: 180.2s Not tested
- RTX4070 OC 12GB GDDRX6: 1-pic: 19:74s 10-pic: 175.61s Power: 117W average 324W Peak: 414W
Higher end video cards are not test as the increase percentage of performance and energy saving does not correspond to the percentage of price increases for my bouget even though I am sure they are amazingly fast and buttery smooth. As LR CC continue to evolve and uses more GPU functions I am sure the gaps will continue to increase. For example, when using the mask brush in 100% zoom on the 60MP image, the refresh speed on the higher end card is almost instantaneous while the lower speed cards takes a second or even a few seconds to refresh from the pixelated image which makes the editing experience much more fluid and pleasant. Also even though some video cards consumed less wattage they do take much longer time when doing large amount of Denoise so the advantage is no longer there especially when I often do 50~200+ images at a time. Based on the speed, power consumption and GPU generation plus additional features, I ended up choosing the RTX 4070 OC for my own need.
I hope this is helpful to someone.
Thank you, George.
While denoising may not be an issue for current full frame cameras, most of us also use a compact 1-inch or APS-C sensor camera as our second "always with us" device. Those RAW files definitely need denoising, so that horse is clearly not out of that barn.
The AI Denoise tool in LR Classic is an excellent tool for the job. Its advantages compared to DxO Deep Prime, which I used previously and have ditched in favor of the LR tool:
- AI Denoise ONLY deals with noise. It doesn't alter contrast, color, shadow or highlight detail. DxO Deep Prime does, in ways that I can't control and have to adjust for after the fact. It annoys me no end, as I want to control those parameters myself.
- AI Denoise lets me decide precisely what level of denoising I want to apply, through the preview slider. True I could use a larger preview window and one more level of zoom, but that fine control allows allows me to adjust the balance between noise removal and detail preservation for each image. DxO Deep Prime does nothing like that.
- It supports a wide range of RAW formats, allowing me to recover images shot with older cameras that were already in a useful range resolution-wise. That includes the early Panasonic LX cameras, which so many of us used prior to the intro of the Sony RX100 series, and even exotic ones like .MRW, KDC and .SR2 - whereas DxO supports only a very limited range. For those of us who have been shooting RAW files for a long time and have precious images that have languished due to excessive noise, it's a godsend.
Thanks for signing up and commenting.
I actually shoot with Micro Four Thirds, an OM-1, and the noise is very well controlled because of the stacked sensor. There are also shooting techniques that limit the amount of noise.
If you own a camera that is unsupported by a program, then you are right using that program isn't worthwhile. For most people with up-to-date gear, then the camera will be.
When it is supported, then trying different software to see what works for them is an option that people should consider. The free trials from DxO and others are a great way to find out if it works for them.
There is a great deal of subjectivity involved in decisions like this and, whereas you prefer Adobe's denoising efforts, I prefer the results from DxO, Topaz, and ON1 over the Lightroom/ACR results. I also find the Adobe preview next to hopeless at showing the results.
I'm not saying the Adobe output is rubbish, it isn't. I can get reasonable results from it, and if I want to denoise an older photo in a hurry, say for using it to illustrate an article here where the image is reproduced quite small, the drop in quality isn't that critical. But for me other apps, including DxO, do a better job.
I do find that highlights are darkened using DxO, but the difference is minimal. There's a very slight change in the histogram. I don't get annoyed by it, but then I am not easily annoyed! But, the results are sharper with DxO. It's a case of horses for courses, I guess.
Thanks again for signing up and joining in the discussion.