Luminous Landscape Forum
Raw & Post Processing, Printing => Adobe Lightroom Q&A => Topic started by: StephaneB on April 18, 2023, 07:50:55 am
-
Well, as DigitalDog implied in another thread, Adobe added an AI-based demoise feature in the Enhance photo. It generates a DNG with both Denoise and Raw details applied. You can tune the intensity of Denoise.
Inevitably I had to compare with DXO PureRaw 3, for Which I still have 8 days in the trial period.
Even without touching DeepPrimeXD, just with deepPrime, the results I see after just a few pictures are that with some photos it can be very close after adding sharpening to the Denoise, but in most cases PureRaw DeepPrime is clearly better. I mean, much better. Better details, better contrast, good sharpness. In those cases, sharpening in LR does not uncover the extra details DXO brings.
DeepPrimeXD leaves no chance to LightRoom.
PureRaw is much faster but the workflow is not as well integrated, with PureRaw creating a DXO subfolder instead of putting the DNG in the same folder as the original and stacking them together, which I would prefer.
DXO really did something extraordinary in my opinion.
Regarding the difference in details it is only in higher ISO, pretty noisy photos, like ISO 3200 with a D800. With less noisy images, like ISO 400 on a D850, there is by default more detail in the DXO conversion, but tuning the sharpening in LR brings the same levels of details as DXO.
Is the difference worth 129€? Well, when I see what people (me included) sometimes spend extra on lenses for much less visible differences, I'd say yes.
Edit: I have since tried yet other pictures and it is not as clear cut.
-
As someone who's tested this for months and someone who purchased a license to DxO PureRAW2, I will not be upgrading to version 3.
Also necessary at this point to put out: high ISO doesn't cause noise (sometimes it does the opposite). Underexposure does.
-
As someone who's tested this for months and someone who purchased a license to DxO PureRAW2, I will not be upgrading to version 3.
If I had version 2 I probably wouldn't either. It is just that on some pictures XD increases the difference between Adobe's Denoise and PureRaw. I would be considering PureRaw whether it had XD or not.
I expected Adobe to be closer to DXO.
-
I agree. DXO has done something really extraordinary w/ Pure Raw, and V3 Deep Prime XD is truly remarkable. Thanks for the comparison. I’ll be doing my own testing later today, but your example shows a pretty dramatic advantage for DXO. One of the things that’s too rarely commented upon re DXO Pure Raw is that it’s way more than an excellent noise reduction tool. Its ability to extract better detail from the raw file than does LrC is a significant thing. And their sharpening and lens correction modules are superb as well. The price for this is time. With files from my GFX 100, DXO can be painfully slow to render and unsuitable to use on “everything.” It is reserved for “keeper” images that will be printed, and usually printed large.
I’m more excited about the curves panel for masks in LrC 12.3. This is a powerful enhancement. As I learn more combinatorial uses of “intersect with” in masks, my round tripping to Photoshop is diminishing. Curves for masks will further this trend. 😎
Adobe has really lived up to their promise to keep developing LrC and rolling out updates as improvements are made. Even though I wasn’t a big fan of the subscription model, I have to say that as a “value proposition” the Adobe Photographer’s bundle is really a bargain. Couldn’t be happier.
Rand
-
Thanks @StephaneB for the example.
Not updated to LR 12.3 yet, but in your above example is this result typical of DeepPrimeXD?
Has done a better job of NR for sure but at the expense of all they wierdness in the background and hyped sharpening and over contrast edges almost like a PS paint/sketch filter.
I much prefer the more realistic LR result here.
-
Has done a better job of NR for sure but at the expense of all they wierdness in the background and hyped sharpening and over contrast edges almost like a PS paint/sketch filter.
I much prefer the more realistic LR result here.
Firstly, keep in mind this is a 100% view. Full screen is better with DXO in my opinion.
Secondly, this is after further sharpening in LR, just to see how it behaved.
Default sharpening in DXO is pretty well juged. Furthermore, you can apply negative clarity in LR to tone it down globally and put it back locally with masks.
I find the Adobe Denoise version does not give you as many options because it does not react well to sharpening., contrarily to the DXO result.
-
Firstly, keep in mind this is a 100% view. Full screen is better with DXO in my opinion.
The only accurate way to view the processing is at 1:1 or greater (100% +) and in Develop module. All other zoom ratios (zooming out) are subsampled and not an appropriate way to view NR or sharpening.
-
I tested with an image I took last week of the Victoria Crested Pigeon at the San Diego Wildlife Park (AKA part of San Diego Zoo). Taken with Canon 5DSr and Canon 100-400mm L IS II at 400 mm and ISO 3200. Here is a view comparing Adobe Enhanced NR with DXO Pure Raw 2 at 100% pixels. I prefer the Adobe version which looks much more natural to me; the DXO version is oversharpened in this area.
-
I processed other pictures; with better original sharpness and less noise (D850 at 1600) and there I too preferred the Adobe result. In comparison the DXO processing indeed appears to oversharpen.
On the attached example, I could not find a difference between DeepPrime and DeepPrimeXD.
-
https://jkost.com/blog/2023/04/denoise-masking-and-edit-in-photoshop-updates-in-lightroom-classic-v12-3.html
-
Here is a comparison of a portion of the pigeon's crest at 100% showing the image before noise reduction and after Adobe Noise Reduction. The background is very clean with minimal if any loss of sharpness in the subject. In my opinion Adobe hit it out of the park with this feature.
-
https://jkost.com/blog/2023/04/denoise-masking-and-edit-in-photoshop-updates-in-lightroom-classic-v12-3.html
One recommendation she makes that is rather important and was expressed by Adobe engineering to testers: "I would suggest that you apply Denoise early in your workflow"
-
I've spent quite a few hours now trying the new NR feature and it's really good.
Comparisons to the DXO conversions are complicated by the addition of their optical corrections, which are sometimes better than the ones in LR. I think DXO PL still has something to offer for some cameras/lenses to get the maximum quality output.
For LR users the convenience of keeping NR 'in house' is huge. On my W10 system processing a file is twice as quick as passing it through DXO PL5.
It's not perfect though;
First LR crash for three years. Maybe a one off and might not be repeated but...
Not sure that the 200% preview is as helpful as a 100% one.
A shame there doesn't seem to be any way of finding out what settings were used for NR in the final DNG or in the export history of the master file.
Annoying there can't be a full size preview option, even if it did take a while.
Curious that the DNG file sizes can be a lot bigger than DXO's DNGs.
-
One recommendation she makes that is rather important and was expressed by Adobe engineering to testers: "I would suggest that you apply Denoise early in your workflow"
That was also mentioned in Eric Chan's Adobe Blog post (https://blog.adobe.com/en/publish/2023/04/18/denoise-demystified), which includes numerous examples and some descriptive information about the technology.
The new Lightroom Denoise feature only works on raw files from Bayer and X-Trans sensors. And based on what Eric writes, I infer that the neural network was trained specifically to reduce sensor noise, and not deal with other types of speckles. I'm hoping future enhancements will include rendered files and different defects, such as the dust and scratches that afflict scanned negatives. The weblog post seems to suggest that work in those directions is taking place.
Denoise is our third Enhance feature. We’re proud of what it can do today, but we’re already looking ahead to make it even better. For instance, we have some ideas on how to use additional training data to improve resolution. We’d like to support additional file formats and combine Denoise with Super Resolution. We’re even looking into ways to speed up the workflow by not needing to make a new DNG file.
I also license, use, and am favorably impressed with DxO PureRaw, but my preference is to perform as much post-processing as possible in Lightroom without invoking additional, external programs unless doing so results in a distinct improvement in image quality that is visible in a normal viewing environment.
-
I processed other pictures; with better original sharpness and less noise (D850 at 1600) and there I too preferred the Adobe result. In comparison the DXO processing indeed appears to oversharpen.
On the attached example, I could not find a difference between DeepPrime and DeepPrimeXD.
On the images shown, I agree. The feather detail on the DxO version, looks like wires. However, I guess, a lot has to do with settings. We need to see more. Of course, having it all in the one package is a workflow advantage that can't be dismissed.
-
Have just tried Denoise and am impressed. Julianne Kost’s tutorial is good. Updates to LR these days make it worth the subscription IMHO.
Jonathan
-
I tried another photo of Ranunculus from The Flower Fields in Carlsbad taken with Canon 5DSr and 100-400mm lens. This one at 150mm and ISO 400. In this case Pure Raw 2 extracts more detail than Adobe LR with no significant artifacts. I prefer the Pure Raw 2 image here. Can't post the 100% comparison view due to restrictions on total uploads to a thread, but you can see the image here:
https://www.dropbox.com/s/qdqhz5zt1mmg3fi/Flower%20Denoise%20Comparison.jpg?dl=0
-
I tried another photo of Ranunculus from The Flower Fields in Carlsbad taken with Canon 5DSr and 100-400mm lens. This one at 150mm and ISO 400. In this case Pure Raw 2 extracts more detail than Adobe LR with no significant artifacts. I prefer the Pure Raw 2 image here. Can't post the 100% comparison view due to restrictions on total uploads to a thread, but you can see the image here:
https://www.dropbox.com/s/qdqhz5zt1mmg3fi/Flower%20Denoise%20Comparison.jpg?dl=0
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 tried another photo of Ranunculus from The Flower Fields in Carlsbad taken with Canon 5DSr and 100-400mm lens. This one at 150mm and ISO 400. In this case Pure Raw 2 extracts more detail than Adobe LR with no significant artifacts. I prefer the Pure Raw 2 image here. Can't post the 100% comparison view due to restrictions on total uploads to a thread, but you can see the image here:
https://www.dropbox.com/s/qdqhz5zt1mmg3fi/Flower%20Denoise%20Comparison.jpg?dl=0
Is it really extracting more detail or is it sharpening more than Ai Denoise? None of us know what's under the hood but I'd compare them at equivalent NR and sharpening (with the Detail panel).
-
Is it really extracting more detail or is it sharpening more than Ai Denoise?
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)
-
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.
-
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.
-
Or is it created by machine learning ?
It definitely is: Eric Chan's Adobe Blog essay (https://blog.adobe.com/en/publish/2023/04/18/denoise-demystified) 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.
-
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.
-
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!
-
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.
-
The Denoise is Linear DNG hence the larger size.
http://www.barrypearson.co.uk/articles/dng/linear.htm
-
A linear DNG will be typically 3 times larger than a mosaicked raw (less if compressed) . No surprises here.
-
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 (https://www.photographyaxis.com/post-processing/lightroom-denoise-ai/)
-
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.
-
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)
-
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.
-
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.
-
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...
-
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.
-
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)
-
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...!
-
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.
-
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.
-
How are the apple M1 and M2 doing?
On my M1 Mac Mini it works well.
-
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.
Hi Andrew,
Any further info both 1) known, and 2) can be shared, on the issue with the neural engine not being used during denoise? I've checked the Adobe thread you posted. No update nor anywhere else so far as I can see. Particularly interested in;
How much faster might Denoise be when using NE as expected
Are we waiting for an Apple fix or a temporary Adobe workaround? (I'm reminded of the access to a Lightroom catalog on external drive that accompanied Ventura)
Thanks
Mike
-
Have just tried Denoise and am impressed. Julianne Kost’s tutorial is good. Updates to LR these days make it worth the subscription IMHO.
Jonathan
I agree.
In light of the offerings from Topaz and DxO, this is long overdue. For better or worse, many of us are effectively tied to Adobe, so it's gratifying that we are not forced into cumbersome external editing and more cost. We must remember that Denoise is still straight out of the box. There has been not one tweak as yet. I expect there to be many in the weeks to come.
I'm running LrC on a 12th Gen Intel i7-12700K with just the integrated GPU. It runs LrC perfectly. However, Denoise has thrown a spanner in the works. I am addicted. Denoise runs but the iGPU takes 4 minutes at 100% to process a 20MB file. That's too long. I'm resigned to having to instal some thumping big discrete GPU but I'll wait a bit to see what Denoise tweaks come along and for Adobe to update advice on optimum hardware requirement in view of Denoise. In the meantime, I'll be taking a few more 4 minute breaks and put a bit of money aside.
-
For better or worse, many of us are effectively tied to Adobe, so it's gratifying that we are not forced into cumbersome external editing . . .
The ideal situation, for this and the other Lightroom Enhance features, would be to apply the adjustments to the raw file dynamically ("parametrically," in Jeff Schewe parlance) rather than producing a linear DNG—albeit one that is scene-referred.
Apparently, Adobe is working on this (https://blog.adobe.com/en/publish/2023/04/18/denoise-demystified):
We’re even looking into ways to speed up the workflow by not needing to make a new DNG file.
One of Lightroom's most valuable attributes—at least, from my perspective—is that it allows the photographer to maintain a raw workflow. Demosaicing should be relegated whenever possible to producing an output file format for printing or static display.
-
I was thrilled when this update was issued.
I own most of the available denoise software. DXO v2 and now 3, topaz and topaz AI and on1 denoise. I'm a bit of a software junky. Early on I stopped using topaz and ON1 mostly because of the DNG files they returned.
I have run some tests on about 50 files that I previously processed in DXO. IN DXO2 I found my self frequently turning off global lens sharpening because the output looked over sharpened (crunchy! ;-) ). particularly bird feathers. In DXO3 I have found myself using "soft" selection in the lens softness selection. This in an attempt to avoid over sharpening particularly in the XD method.
I have found that adobe denoise does as good a job removing noise. It also looks as good after it has been properly sharpened. The new XD selection in DXO3 runs slower than adobe. I have had no problems with excessive run times with adobe. They come in at about the projected times. In the case of both pieces of the software the complexity of the image seems to matter.
I will be using adobe denoise from now on. It returns simplicity to my workflow and really does the job. I will keep DXO around in case I get images that adobe doesn't handle well but will probably not upgrade. It may not work better but I will be able to experiment. (software junky) No more renewals for Topaz. I keep ON1 around for its effects but probably will not use their denoise.
Thank you Adobe.
-
Hi Andrew,
Any further info both 1) known, and 2) can be shared, on the issue with the neural engine not being used during denoise? I've checked the Adobe thread you posted. No update nor anywhere else so far as I can see. Particularly interested in;
Are we waiting for an Apple fix or a temporary Adobe workaround? (I'm reminded of the access to a Lightroom catalog on external drive that accompanied Ventura)
DXO is not using the NE either, they state it creates colour shifts and they are working with Apple on the issue. It could be the same issue Adobe is facing.
-
I'm running LrC on a 12th Gen Intel i7-12700K with just the integrated GPU. It runs LrC perfectly. However, Denoise has thrown a spanner in the works. I am addicted. Denoise runs but the iGPU takes 4 minutes at 100% to process a 20MB file. That's too long. I'm resigned to having to instal some thumping big discrete GPU but I'll wait a bit to see what Denoise tweaks come along and for Adobe to update advice on optimum hardware requirement in view of Denoise. In the meantime, I'll be taking a few more 4 minute breaks and put a bit of money aside.
I am using an older i7-9700 (32gb ram and 1tb SSD) with a Quadro P1000 4gb, the denoise AI would take 4.5 min for a Sony A7R3 42mp file (estimate showed 8min). Also LR CC was very slow in general. A friend with a new Mac Mini M2 pro took 38sec for denoise AI with the identical 42mp RAW file. I really didn't want to upgrade my whole computer so I took a chance and installed a GFX 3060 12GB OC video card. LR CC is now very quick and denoise AI processes an A7R3 file in 20sec and 24mp RX1 files in 10sec. I was very surprised the jump with just the video card, I would have been happy with sub 90sec.
-
I am using an older i7-9700 (32gb ram and 1tb SSD) with a Quadro P1000 4gb, the denoise AI would take 4.5 min for a Sony A7R3 42mp file (estimate showed 8min). Also LR CC was very slow in general. A friend with a new Mac Mini M2 pro took 38sec for denoise AI with the identical 42mp RAW file. I really didn't want to upgrade my whole computer so I took a chance and installed a GFX 3060 12GB OC video card. LR CC is now very quick and denoise AI processes an A7R3 file in 20sec and 24mp RX1 files in 10sec. I was very surprised the jump with just the video card, I would have been happy with sub 90sec.
I've just ordered the same GPU to upgrade my system (an older desktop with an i7 7800 and a Quadro P2000 / 5GB). I've seen several posts that suggest similar Denoise times for 3060, 3050 and 2060 GPUs while there's not a huge step up in performance for 40 series. Given the amount of high ISO captures I make this should give my system enough legs to last through 1-2 more rounds of software development and maybe Apple resolve the NE issue to see if a newer Studio or updated PC might be my next upgrade.
-
In terms of Denose (or any NR) and high ISO, can someone explain why high ISO instead of underexposure is always in the mix when people speak of noise?
An example made this morning:
(http://digitaldog.net/files/ISO200Normal_minus3stops.png)
Details:
Canon R6 Mark II.
ISO 200 (yes, I can set it lower with a custom setting, no need).
Meter: Minolta Flash Meter III incident reading.
IncidentMeterNormal.dng:
Normal exposure is recommended exposure of the Minolta meter: 1/15th@F8
In-camera meter tells me that would be 1 stop overexposed. It recommends 1/30th@F8.
-3stops.dng
125th@F8 or three stops below incident meter recommendations: UNDER-EXPOSED.
All settings in Lightroom Classic’s Detail OFF.
Does all this talk of 'high ISO' take the new (or older) NR off the radar for some users?
URL of screen captures if that makes it easier to view: http://digitaldog.net/files/ISO200Normal_minus3stops.png
-
One more test.
1. ISO 800 but using above-recommended exposure 15th@F8.
2. ISO 50 but 125th@F8.
Detail settings in Lightroom Classic OFF.
Low ISO produces less noise: No.
High ISO produces more noise: No.
-
In terms of Denose (or any NR) and high ISO, can someone explain why high ISO instead of underexposure is always in the mix when people speak of noise**?
Those are the people who believe that ISO is part of the exposure.
-
Those are the people who believe that ISO is part of the exposure.
+1. It isn't.
-
One more test.
1. ISO 800 but using above-recommended exposure 15th@F8.
2. ISO 50 but 125th@F8.
Detail settings in Lightroom Classic OFF.
Low ISO produces less noise: No.
High ISO produces more noise: No.
In this case, to prove your point, wouldn’t you want the exposure in both cases to be the same? That is case #2, ISO 50 with 1/15@f8. Then, assuming a reasonably linear sensor, both images would show same noise….that is, no change due to ISO change
-
In this case, to prove your point, wouldn’t you want the exposure in both cases to be the same? That is case #2, ISO 50 with 1/15@f8. Then, assuming a reasonably linear sensor, both images would show same noise….that is, no change due to ISO change
This is a test to show ISO alone isn't the cause of more noise. It is a test to look at high ISO with automatic exposure (see below, sorry, another copy and paste):
ISO does not affect exposure. ISO affects an exposure recommendation. You (can) control exposure if you wish.
Setting ISO speed does not change the sensor's sensitivity to incoming light like volume control does not change the sensitivity of a radio. In both cases the setting (ISO or volume) controls only the signal processing, while the input stage (sensor, antenna) provides the same input signal and in this case, exposure to the sensor.
That's why when ISO setting is cranked up, automatic exposure results in more noise - automatic exposure, in this case, decreases the exposure (that is, the combination of aperture and shutter speed is set to allow less light captured by the sensor): less exposure, less light, more noise.
-
In this case, to prove your point, wouldn’t you want the exposure in both cases to be the same? That is case #2, ISO 50 with 1/15@f8. Then, assuming a reasonably linear sensor, both images would show same noise….that is, no change due to ISO change
Actually, I did this years ago and it IS worth considering:
(http://digitaldog.net/files/100vs800iso.jpg)
The higher ISO with the same exposure has less noise in this camera.
I should also test this with my newer Canon.
-
I suggest that everyone do their own tests, but what I have found is that given a specific aperture / shutter speed combination, you will get the best results by increasing ISO as much as possible without clipping the highlights.
-
I suggest that everyone do their own tests, but what I have found is that given a specific aperture / shutter speed combination, you will get the best results by increasing ISO as much as possible without clipping the highlights.
From what I have read, that is generally the case, but, for most current sensors, negligible unless you are utilizing dual ISO sensor.
-
I've just ordered the same GPU to upgrade my system (an older desktop with an i7 7800 and a Quadro P2000 / 5GB). I've seen several posts that suggest similar Denoise times for 3060, 3050 and 2060 GPUs while there's not a huge step up in performance for 40 series. Given the amount of high ISO captures I make this should give my system enough legs to last through 1-2 more rounds of software development and maybe Apple resolve the NE issue to see if a newer Studio or updated PC might be my next upgrade.
Hope you see a significant improvement as well.
-
Hope you see a significant improvement as well.
30 seconds for a 50mp Sony A1 raw, vs 3-4 minutes before. Pretty happy with that until my whole system needs to be replaced.
-
The higher ISO with the same exposure has less noise in this camera.
Welcome to the Canon club! This high read noise is a perk of Canon sensors (among others).
With a perfect (ISOless) sensor, noise should be the same with the same exposure (counted in photons).
-
Welcome to the Canon club! This high read noise is a perk of Canon sensors (among others).
With a perfect (ISOless) sensor, noise should be the same with the same exposure (counted in photons).
Yes, indeed (although I haven't tried the same test with my new R6 Mark II. It Sounds like you're saying don't waste the time). At least with my old Canon, it isn't ISO invariant.
I've been working on a (short?) paragraph or two on this; comments are welcome to get it smoothed out:
The term "ISO invariance" refers to a camera's ability to maintain consistent image quality while capturing images at different ISO settings. Suppose a camera is (perfectly) ISO invariant at all ISO settings. In that case, it will produce the same level of image quality whether the ISO setting value that was set at capture or image brightness is adjusted in a raw converter. This ISO setting “doesn’t matter” in that scenario.
A non-ISO invariant camera may introduce more noise or less noise depending on the exposure due to this ISO setting; there is a variant to the outcome of setting the ISO.
In both kinds of systems, ISO will affect the recommendation of exposure (Aperture + Shutter speed), which affects the result of noise. In a non-ISO invariant camera, a higher ISO at optimal exposure can produce less noise than a lower ISO setting using the same exposure! This should not be the case with an ISO-invariant camera system. Hence the name and behavior. This behavior destroys the myth that “higher ISO produces more noise.” Underexposure is the result of (increased) noise, not ISO.
Altering ISO and, thus, the exposure recommendation can come in handy in low-light situations where you may need to increase the ISO setting and this exposure recommendation to avoid camera shake or stop motion with a higher shutter speed. Or if you wish to alter DOF, that wouldn’t be recommended* at a lower ISO setting. This results in less exposure and more noise. Again, it is worth stressing: underexposure is the cause of (increased) noise, not ISO.
* Lastly, the ISO provides a ‘recommended’ exposure (Aperture + Shutter speed), but the savvy photographer does not need to accept this! Even without the inclusion of ISO, a savvy photographer will often ignore a reflective meter recommendation, for example, a black cat on coal or a white dog on snow. Such a recommendation would produce an overexposed or underexposed capture, despite the ISO setting in those examples.
-
The term "ISO invariance" refers to a camera's ability to maintain consistent image quality while capturing images at different ISO settings.
Just to nitpick a bit and add to my former message, this does only concern the shadows.
In the highlights, higher ISO may mean more clipped highlight (and therefore a loss of information).
The term that I'd prefer for cameras with good ISO behaviour is ISOless, because you don't need to set a compromise between highlight clipping and read noise in the shadows, and the ISO setting just stay put at the base (optimum) value.
Guillermo Luijk spoke of that much better than I could, see eg https://forum.luminous-landscape.com/index.php?topic=56906.0
-
Just to nitpick a bit and add to my former message, this does only concern the shadows.
In the highlights, higher ISO may mean more clipped highlight.
Due to that expose yes?
-
Due to that expose yes?
Clipping due to ISO I'd rather say : with constant exposure, a higher ISO will cut more of the highlights due to the gain applied.
-
Clipping due to ISO I'd rather say : with constant exposure, a higher ISO will cut more of the highlights due to the gain applied.
Got it; thanks.
-
Clipping due to ISO I'd rather say : with constant exposure, a higher ISO will cut more of the highlights due to the gain applied.
I don't think this is correct.
I've just made a test; two sets of exposures of an X-Rite Colorchecker, one set at ISO 200 and one set at ISO 6400, successively increasing the exposure to +5 stops. I can see no visible difference between the ISO 200 clipping and ISO 6400 clipping at the same levels of over exposure.
This makes sense since at higher ISOs the light meter is recommending successively reduced exposures (less light on the sensor) therefore resulting in a reduced signal which is then being amplified before going to the ADC (Analogue to Digital Converter). So, I would expect more noise but I would not expect highlight clipping at higher ISOs.
As regards to increasing ISO and noise, I believe this is more complicated with more modern cameras. Initial testing of my new DSLR suggests that there may be some kind of noise reduction being applied; but, as a general rule, the higher the ISO the more noise I am getting, particularly in the shadows – although I have to say the high ISO results I am achieving are astonishingly good compared to my previous camera.
I am not sure I subscribe to the "ISO-invariance" idea. I don't think that any digital sensor/camera can actually be ISO-invariant although it might be possible to describe them as 'noise-invariant'; or perhaps, more accurately, as 'less noise-variant'.
-
What camera Simon?
-
I don't think this is correct.
I've just made a test; two sets of exposures of an X-Rite Colorchecker, one set at ISO 200 and one set at ISO 6400, successively increasing the exposure to +5 stops. I can see no visible difference between the ISO 200 clipping and ISO 6400 clipping at the same levels of over exposure.
This makes sense since at higher ISOs the light meter is recommending successively reduced exposures (less light on the sensor) therefore resulting in a reduced signal which is then being amplified before going to the ADC (Analogue to Digital Converter). So, I would expect more noise but I would not expect highlight clipping at higher ISOs.
You are changing exposure. The previous comment was about keeping the same exposure (shutter speed / aperture) while changing ISO.
-
You are changing exposure. The previous comment was about keeping the same exposure (shutter speed / aperture) while changing ISO.
Indeed.
Now this is an old Canon DSLR, and as mentioned, it isn't ISO Invariant.
Same Exposure, only difference is the ISO (100 vs. 800).
I've placed the eyedropper over the same area in Lightroom Classic and set the readout for Lab. Unfortunately, it's nearly impossible in that product to control the exact position of the eyedropper, unlike Photoshop or the sampling. But you can see the differences are tiny and insignificant. It appears in this case, the ISO does not affect the highlights. YMMV.
(http://digitaldog.net/files/ISO800_100Clipping.png)
-
You are changing exposure. The previous comment was about keeping the same exposure (shutter speed / aperture) while changing ISO.
Yes, a good point.
But if you set the ISO to a higher figure and keep the exposure the same then the resulting file will have higher tonal values thus clipping the highlights. By way of example, if you expose a grey card at 200 ISO to produce a mid-grey (nominally, the correct exposure) and then increase the ISO to 800, say, but keep the exposure the same, then you have effectively pushed the grey-card exposure into the highlights at +3 stops. So, I would expect to start to see some clipping.
However, I think we may be missing something in our understanding.
THE SENSOR
In analogue terms, before the signal is sent to the ADC, setting a higher ISO would in theory give the sensor (before the signal is amplified) more headroom since the exposure, as advised by the meter, is reduced. So, in theory, increasing the ISO by 5 stops, as in my test, would give an extra 5 stops more headroom above mid-grey (assuming that the ‘base’ ISO is 200). But my results show identical tonal compression and loss of colour fidelity (clipping) at the same amounts of over-exposure for both ISO 200 and 6400.
At the moment I don’t really know why this should be. It could be down to the ADC or some other digital processing. I need to think about this some more and perhaps do some more tests.
As I understand it, at the moment, on the analogue side of things the sensor is limited by the maximum amount of volts it can produce. Exposing it to more light will not increase its output voltage (full bucket). This varying voltage or signal is produced when the sensor is ‘read’ and all those buckets full of electrons are emptied ;D.
SETTING ISO
I believe the ISO setting on the camera, or more properly the Exposure Index (E.I.), does at least three things. It changes:
• the sensitivity of the camera’s exposure meter
• the analogue amplification the signal sent to the ADC
• the ISO flag in the EXIF data to the ISO selected.
I suspect that it may do more than this. Possibly it may introduce analogue noise reduction at higher ISOs or introduce some processing of the resulting digital RAW file. Current evidence from some modern DSLRs and higher-end mirrorless cameras suggest this may be the case.
-
What camera Simon?
I was being deliberately coy about the camera I tested as I didn’t want to set-off a fan-boy type exchange with other contributors.
The camera I am testing is a Canon EOS 5D MkIV. Previously, I have run tests on the 5D and the 5D MkIII. I do not currently have access to other cameras but the Nikon D8XX series reportedly has outstanding noise performance. I do not know much about the Sonys, although I understand they share the same sensor as the Nikons (and are spoken of very highly by their owners). I am very interested in the results from your Canon R6.
My own interest is in trying to understand how sensors work, how ISO works, and the dynamic range of the sensor of my camera, so that I can better determine my exposures for the best results. This stems from a lifelong interest in photography à la the Ansel Adams ‘school’ of photography. If you’ve read Ansel Adams book ‘The Negative’ you will realise my tests replicate those he recommended for understanding how film works.
-
THE SENSOR
In analogue terms, before the signal is sent to the ADC, setting a higher ISO would in theory give the sensor (before the signal is amplified) more headroom since the exposure, as advised by the meter, is reduced. So, in theory, increasing the ISO by 5 stops, as in my test, would give an extra 5 stops more headroom above mid-grey (assuming that the ‘base’ ISO is 200). But my results show identical tonal compression and loss of colour fidelity (clipping) at the same amounts of over-exposure for both ISO 200 and 6400.
You cannot use that "extra" headroom. The analogue gain is applied to the signal before the Analog to digital conversion (ADC), so the ADC reaches saturation at a lower signal than compared to a lower gain. There is nothing mysterious here.
-
You cannot use that "extra" headroom. The analogue gain is applied to the signal before the Analog to digital conversion (ADC), so the ADC reaches saturation at a lower signal than compared to a lower gain. There is nothing mysterious here.
So that would suggest that the clipping (tonal compression and loss of colour fidelity) is a function of the ADC rather than the sensor itself ?
The voltages from the sensor, prior to being amplified, are directly proportional to the amount of light falling on it. So, if the sensor saturates or clips at +3 stops and then you under expose by 5 stops (prior to amplification) the sensor now has 8 stops before it saturates or clips. No?
I'm not, as yet, convinced since the ADC is simply creating binary encoding of the voltages from the output of the amplifier. I can understand how the voltages, being analogue, might vary according to the output of the sensor (then amplified), but not the resulting binary; unless something else is happening during conversion ? Possibly?
More thought needed…
-
I was being deliberately coy about the camera I tested as I didn’t want to set-off a fan-boy type exchange with other contributors.
But the kind/make can make a difference here.
Out Canon's are not ISO Invariant.
-
So that would suggest that the clipping (tonal compression and loss of colour fidelity) is a function of the ADC rather than the sensor itself ?
The voltages from the sensor, prior to being amplified, are directly proportional to the amount of light falling on it. So, if the sensor saturates or clips at +3 stops and then you under expose by 5 stops (prior to amplification) the sensor now has 8 stops before it saturates or clips. No?
I'm not, as yet, convinced since the ADC is simply creating binary encoding of the voltages from the output of the amplifier. I can understand how the voltages, being analogue, might vary according to the output of the sensor (then amplified), but not the resulting binary; unless something else is happening during conversion ? Possibly?
More thought needed…
Suppose you have an 8 bit ADC to convert a signal from 0 to 1V, then an input of 1V will result in the maximum number out of the encoder, which is 255.
The ADC does not know if that 1V is because the signal out of the sensor was 1V or if it was 0.2V with a gain of 5x. It is the ADC that reaches the maximum encoding value when the "Clipping" occurs.
As far as the ADC is concerned, the signal out of the sensor multiplied by the gain is a black box
-
Suppose you have an 8 bit ADC to convert a signal from 0 to 1V, then an input of 1V will result in the maximum number out of the encoder, which is 255.
The ADC does not know if that 1V is because the signal out of the sensor was 1V or if it was 0.2V with a gain of 5x. It is the ADC that reaches the maximum encoding value when the "Clipping" occurs.
As far as the ADC is concerned, the signal out of the sensor multiplied by the gain is a black box
Yes. I agree.
But the clipping which my tests show is that with increased exposure first there is 'soft' clipping where tonal separation is compressed and there is a gradual loss of colour fidelity as exposure is increased. This can also be observed in the graphs from RawDigger.
ADC clipping will be 'hard'; a complete loss of variation in the signal; i.e. highlights with no discernible nor retrievable detail.
This makes me wonder whether the 'soft' clipping is the result of the signal coming from the analogue sensor. Or is this being introduced by some digital processing ? I don’t think it would occur in the situation you postulate as this would result in completely blown highlights, nothing there at all.
Puzzling…
And worth thinking about some more.
-
My example above was an oversimplification. Many things may be at play there. The signal may be "compressed" to avoid hard clipping, but this is only a possibility.
In any case, my point is that the "extra" headroom in the sensor is not available to use, at least not keeping a linear response of the whole system (sensor + ADC)
-
Hard Clipping?
Soft Clipping?
OK, I'm game what's what here?
-
Hard clipping -> the output increases linearly until it reached the maximum value of the encoder and there is an abrupt transition. Typical response of basic, uncorrected digital systems.
In soft clipping, the output "shoulder-off" and there is a gradual transition until saturation or maximum value. This response is more common in analogue systems and more pleasant to the observer.
The use of compression of the analogue signal, both in audio and images, allow for soft clipping at the expense of lower fidelity to the original signal.
-
Hard clipping -> the output increases linearly until it reached the maximum value of the encoder and there is an abrupt transition. Typical response of basic, uncorrected digital systems.
In soft clipping, the output "shoulder-off" and there is a gradual transition until saturation or maximum value. This response is more common in analogue systems and more pleasant to the observer.
The use of compression of the analogue signal, both in audio and images, allow for soft clipping at the expense of lower fidelity to the original signal.
I'd like to see a photo since I'm still a bit confused.
In each case, there's a point of complete clipping right?
Maybe the term should be smooth clipping and abrupt clipping?
-
Maybe the term should be smooth clipping and abrupt clipping?
Yes, it could be that!
-
I'd like to see a photo since I'm still a bit confused.
In each case, there's a point of complete clipping right?
Maybe the term should be smooth clipping and abrupt clipping?
May be ‘clipping’ should be reserved for describing when the highlights are irretrievably lost and referring to what happens prior to this as ‘tonal compression’.
Andrew, I'm thinking about how best to show this with some pictures…
-
My example above was an oversimplification. Many things may be at play there. The signal may be "compressed" to avoid hard clipping, but this is only a possibility.
In any case, my point is that the "extra" headroom in the sensor is not available to use, at least not keeping a linear response of the whole system (sensor + ADC)
I think you’re right, that compression is being used prior to the signal being sent to the ADC. This is how it works in the audio world, compress the high level signals prior to the ADC so that they don’t abruptly become clipped in conversion. That’s the thing with digital, it’s either there or it isn’t ;D. I note that my little Sony RX100 has the tendency to abruptly clip the highlights. I have to make sure I under-expose a little with high contrast subjects.
-
May be ‘clipping’ should be reserved for describing when the highlights are irretrievably lost and referring to what happens prior to this as ‘tonal compression’.
IMHO, not maybe.
Now one of three (or four) channels may clip while the others do not but clipping is clipping.
It is outlined here: http://www.digitaldog.net/files/Langs_Render_to_Print.pdf but also referred to (and this may be a better term for full clipping and more) as sensor saturation: page 5 with a good illustration.
-
IMHO, not maybe.
Now one of three (or four) channels may clip while the others do not but clipping is clipping.
It is outlined here: http://www.digitaldog.net/files/Langs_Render_to_Print.pdf but also referred to (and this may be a better term for full clipping and more) as sensor saturation: page 5 with a good illustration.
Agreed.