Lightroom 12.3 - Denoise is there

[John Hollenberg]

Each to their own I suppose but all the images posted here from the DXO product look way too hyped and completely digital. Almost like an in camera JPEG from 10 years ago.

If the only goal is more sharpness, then DXO knocks it out of the park but at the expense of everything else. I personally much prefer the subtlety of the LR images to the digital looking DXO.

I spent more time looking at the images and noticed black halos in some areas at 200% from the DXO product.  I was able to equal the DXO apparent level of detail (without any of the halos) by adding Clarity +30 to the LR image (but not necessarily saying I would want to do this for this image, just wanted to see if I could achieve an equivalent look without any halos).  The LR image looked more natural.  I think I am going to stick with LR for noise reduction.

[mcbroomf]

Or is it created by machine learning ?

How much would that matter if the results look good ?

(Maybe a it would matter in a few niche specialisations)

I (personally) don't care.  My point meant to be about matching output as closely as possible to decide on software to rate.  It seems to me that Adobe have introduced NR software only but included it as part of their Detail panel which allows sharpening.  So if other software (Topaz/DXO/ON1 etc) look better sharpened, or worse, then if I'm trying to see what is "best" for me then I'd tweak all available settings to get the best output, then decide.  A friend of mine did exactly the same comparison with Adobe, DXO and Topaz, though Topaz was the most sharpened in the example I saw.  All 3 were different though.

[Chris Kern]

Or is it created by machine learning ?

It definitely is: Eric Chan's Adobe Blog essay makes that clear.

Denoise uses machine learning to interpolate those patterns and remove noise at the same time. That is, our models are designed and trained to perform both demosaicing and denoising in a single step.

But reading between the lines, it appears to me that Lightroom's Denoise works quite differently than, say, its Super Resolution feature, despite both being offered (currently only as alternatives) as part of the Enhance menu selection.  My guess is that the neural network for Denoise was trained only to identify noise in the raw sensor data—to flag speckles that matched those of the noise speckles in its training set—so that clean replacement data could be interpolated into the image during the demosaicing process.  In other words, if my guess is correct, you're seeing the actual data captured by the camera sensor with only the noise speckles replaced.

[digitaldog]

It seems to me that Adobe have introduced NR software only but included it as part of their Detail panel which allows sharpening.  So if other software (Topaz/DXO/ON1 etc) look better sharpened, or worse, then if I'm trying to see what is "best" for me then I'd tweak all available settings to get the best output, then decide. 

Yes, the design differs, and yes, if you want apples to apples, you need to do more than just Denoise. You have to apply the various sharpening options in Detail as well.

[Redcrown]

I usually end my day watching random YouTube videos. Tonight there were 5 Adobe Denoise AI videos on the first page of YouTube recommendations. I've never seen a single subject do that before. Talk about a big splash!

[JeanMichel]

I just tried the Denoise on a Leica SL2image made at ISO 6400 (47 mpx). Good result but does require additional processing of course. Interesting that the resulting DNG file, before any further adjustments, is quite larger than the original file. The original DNG is 84 MB or so, the Enhanced-NRdng is 218MB or so, almost 3 times larger. It is not an issue, but it is something to keep in mind.

[digitaldog]

The Denoise is Linear DNG hence the larger size.
http://www.barrypearson.co.uk/articles/dng/linear.htm

[fdisilvestro]

A linear DNG will be typically 3 times larger than a mosaicked raw (less if compressed) . No surprises here.

[govindvkumar]

The Denoise AI from Lightroom does a decent job for RAW images. Hope they can bring it for JPEGs too in the future.

Check out the full review over here with Before/ After Images: Lightroom Denoise AI Review + Full Tutorial

[digitaldog]

The Denoise AI from Lightroom does a decent job for RAW images. Hope they can bring it for JPEGs too in the future.

A possibility but it will not be as effective as what you can accomplish from the raw data.

[jrsforums]

A very good review (as Kost normally does) is here https://youtu.be/fkrggaxUzY0

Lots of other reviews and comparisons are available in YouTube.  In my (very limited) testing, I have been really impressed with the results of DenoiseAI and regular LR sharpening (with masking).  In the cases I compared with Topaz PhotoAI (Tiff, not RAW, workflow), LR results appeared better.

In Eric Chan’s blog post, he says “… we built our machine learning models to take full advantage of the latest platform technologies, including NVIDIA’s TensorCores and the Apple Neural Engine…”.  This appears to be quite true.  I have heard many say it takes minutes to view previews or process Denoise.  However, on my Nvidia 3090 Ti, previews are almost immediate and processing is usually 5-10 sec. (24mp sensor)

[digitaldog]

In Eric Chan’s blog post, he says “… we built our machine learning models to take full advantage of the latest platform technologies, including NVIDIA’s TensorCores and the Apple Neural Engine…”.  This appears to be quite true.  I have heard many say it takes minutes to view previews or process Denoise.  However, on my Nvidia 3090 Ti, previews are almost immediate and processing is usually 5-10 sec. (24mp sensor)

See:https://community.adobe.com/t5/lightroom-classic-discussions/denoise-ai-in-12-3/m-p/13749953#M319651
 

Currently, the Apple Neural Engine found in M1/M2 silicon Macs is not used by Denoise. It's expected that performance will be even better when Denoise does use the Apple Neural Engine.

24.4MP (R6MII) raws on my MBP take 12 seconds to process. Saving the new, linear DNG takes about a second, that's not using GPU on this end.

[jrsforums]

See:https://community.adobe.com/t5/lightroom-classic-discussions/denoise-ai-in-12-3/m-p/13749953#M319651
 24.4MP (R6MII) raws on my MBP take 12 seconds to process. Saving the new, linear DNG takes about a second, that's not using GPU on this end.

Wasn’t aware of apple glitch.  I’m sure times will improve when fixed.  The key point I was trying to make is that this is an area where people will need to consider upgrading there hardware if they use on a frequent basis.  BTW…Topaz PhotoAI seems to work slower on my system, but better on older, slower hardware (friends & other poster…hearsay!).

Yes, GPU speed should not have anything to do with file write speed.

[kers]

I spent more time looking at the images and noticed black halos in some areas at 200% from the DXO product.  I was able to equal the DXO apparent level of detail (without any of the halos) by adding Clarity +30 to the LR image (but not necessarily saying I would want to do this for this image, just wanted to see if I could achieve an equivalent look without any halos).  The LR image looked more natural.  I think I am going to stick with LR for noise reduction.

+1 i see they really built a good working noise reduction into LR; well done! Adobe
 it looked much better and cleaner than the dxo files. Also it works better in the workflow. 20 sec a 45MP file in my case is a 3x slower than DXO deepprime2 but also better...
Another reason not to save the noise reducted DNG's :  since they take much space and noise reduction on the original NEF will be better every year.
I only wish that Adobe would let you choose to use Intel's openVINO AI-frameswork so it would be handled by the processor (CPU) instead of the GPU.
Topaz denoise lets you choose , and in my case it is faster on the CPU than on the GPU.
BTW Topaz denoise still can be used in combination with DXO or LR-NR and brings sometimes extra benefit because of the different approach.

What is missing (?) in LR ; a way to sort images by ASA. i would like to use LR-denoise  only on all images above 1600 asa...

[digitaldog]

The Denoise is Linear DNG hence the larger size.
http://www.barrypearson.co.uk/articles/dng/linear.htm

Update: in addition, a lossless DNG “copy” of the original is embedded in the new DNG.
Getting it out? Not yet possible.

[fdisilvestro]

Another reason not to save the noise reducted DNG's :  since they take much space and noise reduction on the original NEF will be better every year.

I'm not sure I follow, The new Denoise in LR does create a new, larger file (as DXO also does)

What is missing (?) in LR ; a way to sort images by ASA. i would like to use LR-denoise  only on all images above 1600 asa...

Maybe not exactly what you desire, but you can filter by ISO (Metadata filter)

[kers]

I'm not sure I follow, The new Denoise in LR does create a new, larger file (as DXO also does)

yes, but you can easely thow it away after you are finished since you can always re-apply it on the original RAW and there will be a better denoise in the future.
for me LR poops out a tiiff that i work on in photoshop ; the DNG is an intermediate that i can waste.

Maybe not exactly what you desire, but you can filter by ISO (Metadata filter)

i will look into it ; i am not a LR expert as you can see...

i looked and yes it is obvious learned something...!

[john beardsworth]

What is missing (?) in LR ; a way to sort images by ASA. i would like to use LR-denoise  only on all images above 1600 asa...

Use the library filter.

[kers]

Unfortunately LR Denoise makes my Apple computer crash after about 20 images and i am not alone....
( i mean a complete computer crash and subsequent restart)

I had the same problem with DXO deepprime and an older videocard...
Now i have a radeon Vega 64 with 8GB of ram in it, so not bad at all...

How are the apple M1 and M2 doing?

update:
 i just installed a windows 10 harddisk and with the same hardware i do not have any crashes, but the speed is about half i had on the mac software.

Would be nice if LR has a way to use less of the GPU or wait some  time from iage to image to prevent crashing the GPU on the mac.. I read more people on  -Windows and Mac's -are experiencing crashes.
Never had that before AI came along.

[StephaneB]

How are the apple M1 and M2 doing?

On my M1 Mac Mini it works well.