-
Welcome to the new forums! Please read this first. For known issues we are working to resolve, click here.
For those with two systems....
- Thread starter Captive18
- Start date
Messier Object
Forum Pro
- Messages
- 13,909
- Solutions
- 15
- Reaction score
- 14,947
I don't have a simple answer
I've started typing an answer to this question seveal times and each time my reply gets too long and complicated and I've ended up deleting it.
Same again this time around
so I'll just say that each of the camera formats I'm still using:
Fourthirds DSLR
Canon FF DSLR
Canon APS-C DSLR
Canon EOS R system APS-C
Olympus Micro Fourthirds
. . . has merits and deficiencies and some are more fun to use than others
Peter
MEDISN
Senior Member
I would be interested in this too Ken. Specifically how the built in stacking methods (LiveND, HR, HHHR) impact this type of noise as well as the impact of the fancy AI noise reduction/sharpening programs (if any). Thanks!Could you please share more info about mid tones noise level, or of you have any real comparison FF with M43? I am wondering if I would able to see it as wellthank you
kenw
Veteran Member
Artistic CaveatI would be interested in this too Ken. Specifically how the built in stacking methods (LiveND, HR, HHHR) impact this type of noise as well as the impact of the fancy AI noise reduction/sharpening programs (if any). Thanks!Could you please share more info about mid tones noise level, or of you have any real comparison FF with M43? I am wondering if I would able to see it as well
So of course the most important preface to any of this is:
Content matters more than technical details of an image
I have a few images from my old iPhone 4S that I like better than 90% of my Z7 images. I also always remember walking around a NatGeo contest gallery at the Smithsonian where there were some huge enlargements (like five feet to a side) from 1/2.7" sensors up to MF. The 1/2.7" images still had huge impact, even with noise and some lack of sharpness compared to the MF stuff. The content mattered way more than the details.
And the effect I'm addressing here (mid-tone base ISO noise) is really not glaring at all either. As will be discussed below it depends on lots of things and as a result undoubtedly a very large fraction of photographers are going to consider the difference very marginal compared to the cost, size and weight impacts of FF sensors.
I expect both of you know that already, but so many threads derail if we don't state this clearly at the start!
Technical Context
Next, to be clear, when I say "base ISO mid-tone noise" I'm talking about optimally exposed shots meant for RAW processing. So exposure chosen so "important" highlights are not clipping but otherwise we've "exposed to the right" as much as practical.
The follow up to that is that in these conditions fancy new sensors do essentially nothing. In this region we are entirely bound by the physics of light (i.e. photon shot noise). So it is basically down to sensor size and full well capacity (i.e. "base ISO" as far as a RAW file is concerned which is often a bit different from specified ISO for JPEGs).
Samples below are from a Nikon Z7, Olympus E-M5II and Panasonic GM1. For these sensors the difference in base ISO RAW noise in midtones is about 2.6 stops. The newer 20MP m43 sensor improves on this a bit getting to about 2.1 stops. Remember, just the size of the sensor is responsible for 2 of the stops, so the difference in the sensor technologies is modest to negligible (0.6 to 0.1 stops). But overall because of the sensor size this is a pretty big difference of over 2 stops. The question of course is at what threshold can you notice it?
Processing Effects
It is critical to understand that mid-tone noise is directly impacted by anything you do in post processing that amplifies high spatial frequencies, either in luminance or chrominance. So that includes (using LR slider terminology):
- Contrast
- Clarity (aka Local Contrast Enhancement)
- Texture
- Saturation
- Vibrance
- Dehaze
- Positive Highlights or Whites
- Negative Shadows or Blacks
- Tone Curve controls that steepen the curve
- B&W Mix
- Detail Amount and Radius
The point is there is no one "threshold" at which mid-tone noise becomes a "problem". If I'm going for a softer high-key look and I've got low Contrast, negative Clarity and negative Saturation then I might have "acceptable" mid-tone noise from a 1/2.7" sensor at base ISO. If I crank Contrast and Clarity in a B&W conversion combined with a steep Tone Curve and wild settings on the B&W Mix panel then I might see visible noise from a FF or even MF sensor.
Also, remember that as a starting point almost any profile has an S-shaped tone curve embedded in it which means that it is middle gray in the output that often has the most noise amplification. Hence blue skies and cloudy grey skies are often where we see it first. But any uniform, deep, saturated mid-tone can really make it show up - so for example colorful sunsets.
Now, a lot of my shots are desert landscapes, which in general can require a fair bit of contrast and saturation boost, while applying dehaze to the horizon is very common as well. So I'm the poster child for seeing this problem. Taking brilliant colored urban shots? You might never notice because you might not crank things as hard in post processing. Taking boudoir portraits? Well everything you are doing to fix the model's skin is reducing the mid-tone noise as well - you'll be fine.
Try it yourself!
I always recommend people grab DPR studio image RAW files and just play with them. You can learn so much without ever touching another camera. There are some nice uniform patches in there to examine. Just remember the studio scene is pretty brilliant color wise and so you might need to crank up processing to "clown makeup" levels to mimic what a landscape shot might have happen to it. You can process the RAWs identically and resize or print to whatever output is relevant to you and see if there is a difference you care about.
My Examples
Since you can reach your own conclusions about controlled tests from the DPR studio shots I'm not going to explore that here. Instead I'll provide some real world examples from both systems. They aren't identical shots. They are of course each processed appropriately for the scene. They are all base ISO and ETTR as much as was practical in the field. The point is these are not "controlled" but rather "real world" and so it is a more pragmatic test (to my mind) as to whether there are any practical differences in real world use.
With all that blah-blah out of the way here are my examples as two groups. These are all resized to 4K sizes (because they come from sensors with different aspect ratios and then many of them are slightly cropped to different aspect ratios the scaling is not perfectly normalized). Of course you are going to need to open the originals in new windows/tabs to see anything useful. By all means resize to whatever you think is of interest to you and view from whatever distance is useful to you.
THESE DO HAVE EXIF, so if you are looking to do your first look "blind" then avoid accidentally seeing their EXIF info!
Group A
Group B
What to look for
Again, if you are trying to pick out the two groups blind I'm not going to disclose which group is which just yet - so read on safely!
The most obvious differences are in the skies. Again, those uniform mid-tone regions are what get the most amplification in post-processing. If you look at this 4K images at 100% the differences are usually quite clear. Downsize more or view from a larger distance and that difference will become less noticeable or vanish entirely.
But you can also often see a more subtle effect in detailed regions. Remember that random noise actually creates the perception of "acuity" or to bend the word a bit, "sharpness". So process the very same RAW file and then make a second copy with just a bit of added noise and your brain will perceive increased sharpness that isn't there. Thing is this "noise acuity" also has a "crunchy" aspect to it that most people don't like. So you'll sometimes hear people say smaller sensor images look "crunchy" and then blame it on sharpening artifacts due to using a lower resolution sensor. I honestly don't think that is really the issue much of the time. I think often times it is once again down to mid-tone noise.
One more special example
Which group do you think this one belongs to?
Take a look at this one, especially compared to the other B&W examples and decide if you think it is a Group A or a Group B shot.
ANSWERS BELOW - BEWARE!!!
.
.
.
.
.
.
.
.
.
.
.
.
.
Which was which?
I suspect if you eventually viewed these 4K samples at 100% you concluded that Group A was the m43 cameras and Group B was the FF camera. The rainbow shot in Group A makes it clear for a color image as the noise in the clouds is very evident, and the B&W Group A shot with the truck in it has a very noisy sky as well. But if we displayed just a single pair of images and took the "best m43" and the "worst FF" one might be hard pressed to pick for sure unless you took the time to really analyze what post processing you thought had happened to them. Again, what you do in post really can alter mid-tone noise to a degree greater than sensor differences.
And so what about that last example? I suspect if you looked at it closely you would put it in Group B with the other FF images. It obviously has a lot of "cooking" in post that would amplify noise and give an ugly sky and crunchy detail for a m43 shot. Thing is, it is an m43 shot. Or really, it is a bunch of m43 shots averaged. So this is the point that averaging base ISO m43 exposures can definitely give you the "FF look" for landscape even with just a 16MP sensor. Again - it often isn't about resolution, it is about base ISO mid-tone noise.
Exposure Averaging
For my tenth legally allowed image in a DPR post another averaging example, or rather 100% crops from a single GM1 exposure and an average of multiple exposures (ten I think, but I didn't look it up):
Again, open the original to see the difference but here even in the little DPR box you can see the difference. This processing is a little bit more than just averaging as I was hand-holding for the exposures. Thus I had to align them, and I did that in Photoshop after first upsampling them by 2. Once the stack was merged I downsized again to the original resolution. You can see a lot of enhanced detail here with greatly reduced noise which also allows for more sharpening to recover even more detail. Basically this is doing HHHR externally using Photoshop.
Pixel Shift, Hi Res, HHHR
As the example above illustrates any technique that effectively averages multiple exposures will reduce mid-tone noise. So yes, all the various multi-exposure resolution increasing modes in the most recent cameras basically do this kind of averaging in camera. When I compare a downsized Hi Res shot from my E-M5II to my Z7 they are very similar as far as base ISO noise goes.
What you have to do to combat mid-tone noise is collect as many photons as possible. That's it really. Collect them however you can. Start with base ISO. Do ETTR as much as the scene allows. Then either average multiple exposures or get a bigger sensor.
Comments?
Hopefully that was vaguely useful. Again, there is no "right" answer here. These days a FF sensor definitely will have two stops less mid-tone noise. If your process enough or zoom enough you will see the difference. Will there be a noticeable or "important" difference in the final image? That's a huge "depends" for sure!
Feel free to point out anything I said wrong, or share observations about the images and how you perceive the differences (or lack thereof).
--
Ken W
See profile for equipment list
Last edited:
MEDISN
Senior Member
Well damn Ken, hell of an explanation! Even I can understand that. I’ll admit I was fooled by the grouping, I had them flipped. You have answered my question regarding the in camera stacking impact. A lot of good stuff in there, maybe we should have started a separate thread 
Thanks for doing this. I really appreciate it.
Thanks for doing this. I really appreciate it.
kenw
Veteran Member
Ah, interesting! Do you mean fooled by the overall groups or fooled by which group the last example fit into?
I do realize now I didn't answer your question about noise reduction...You have answered my question regarding the in camera stacking impact. A lot of good stuff in there, maybe we should have started a separate thread
The answer is I haven't really played with the more recent AI stuff enough to have an up to date opinion. What I will say, admitting it is out-dated, is:
- The pre-AI fancy NR stuff never really seemed to work well on base ISO mid-tone noise. I think this is because those algorithms were working best on structured pattern noise in high ISO shadows. Photon shot noise of course has no pattern or structure to it, it is truly random, so there is only so much you can do.
- The early AI NR seemed to work best on more obvious structures it was trained on such as feathers, fur, hair, material textures and larger scale structures like buildings. Seemed to do well on larger natural things like leaves when they were large enough to be identifiable. But a lot of landscapes down at the pixel level are almost fractal in nature (especially deserts without trees) and that seemed to not work so well.
No worries, the least I can do in return for all your excellent posts on HR modes!
Eric Nepean
Veteran Member
- Messages
- 6,418
- Solutions
- 8
- Reaction score
- 4,069
Ken, that is one of finest expositions on art and techniques of diigital photography I have ever read.
lokatz
Veteran Member
Ken, Thanks for sharing your excellent write-up and comparison. I'm sure it will help photographers think more comprehensively about why they use one format or the other, and what impact this has.
What puzzles me in many of the endless "MFT vs FF" discussions is that the aspect of noise reduction (NR) is often omitted, as it is in your post. In my view and experience having used both systems extensively, it is an important factor to take into account.
Up until a few years ago, NR engines used to be rather mediocre and, when used aggressively, came at a price most of us weren't prepared to pay, namely a reduction in image sharpness and/or the introduction of artifacts. That is probably why many shied away from using more than a hint of NR, which still left substantial and clearly visible noise differences between MFT and FF images.
Fast-forward to today: the introduction and further development of AI-based NR engines has been a game changer, as it brought giant improvements in this field. If you've never used DxO's DeepPRIME XD for noise reduction, you may not know what I am talking about: used carefully, this technology almost levels the field between systems as the noise is largely gone without affecting sharpness or detail of the image. Lightroom is not quite there but also received substantial improvements in its latest versions.
Only when using extreme pixel peeping, going to 100 percent or more, will you be able to detect any differences once such NR has been applied, and those differences are small and difficult to see. Looking at the whole image, which I hope we can all agree is the most important aspect anyway, good NR leaves no differences between FF and MFT shots. Similarly, ISO levels that used to be unacceptably high have become usable with this technology. (Examples at https://thisbeautifulplanet.de/denoising-software-comparison.)
As I see it, this whole discussion needs a different direction. Noise no longer is the main distinguishing factor between systems if good NR is used, which makes it no more than an inconvenience.
Where a true difference remains is in dynamic range (DR). The two-stop difference remains unchanged in spite of NR. A Z7 has about 11.6 stops of DR at base ISO, whereas an OM-1 has about 9.6. There is nothing any piece of software could do about that. When shooting MFT, I find every so often that I get to choose whether to blow my highlights or get pitch black dark areas, but I won't be able to avoid both within a single shot.
When shooting landscapes, I can simply bracket my exposure and blend the shots later, which again (usually) makes DR a mere inconvenience. Shooting moving subjects, however, I have to live with the issue. Shooting an animal in thick underbrush, this is not much of a problem since such scenes usually do not have high DR anyway. Shooting a bird in flight against a bright sky can present a dilemma, however, that I may not be able to avoid. If all else were equal (which it isn't, but that's a different discussion), that would make FF the better choice.
In short, I am of the opinion that giving the state of technology we are at, DR warrants special attention in this debate, whereas sensor noise does not/no more.
Lothar
What puzzles me in many of the endless "MFT vs FF" discussions is that the aspect of noise reduction (NR) is often omitted, as it is in your post. In my view and experience having used both systems extensively, it is an important factor to take into account.
Up until a few years ago, NR engines used to be rather mediocre and, when used aggressively, came at a price most of us weren't prepared to pay, namely a reduction in image sharpness and/or the introduction of artifacts. That is probably why many shied away from using more than a hint of NR, which still left substantial and clearly visible noise differences between MFT and FF images.
Fast-forward to today: the introduction and further development of AI-based NR engines has been a game changer, as it brought giant improvements in this field. If you've never used DxO's DeepPRIME XD for noise reduction, you may not know what I am talking about: used carefully, this technology almost levels the field between systems as the noise is largely gone without affecting sharpness or detail of the image. Lightroom is not quite there but also received substantial improvements in its latest versions.
Only when using extreme pixel peeping, going to 100 percent or more, will you be able to detect any differences once such NR has been applied, and those differences are small and difficult to see. Looking at the whole image, which I hope we can all agree is the most important aspect anyway, good NR leaves no differences between FF and MFT shots. Similarly, ISO levels that used to be unacceptably high have become usable with this technology. (Examples at https://thisbeautifulplanet.de/denoising-software-comparison.)
As I see it, this whole discussion needs a different direction. Noise no longer is the main distinguishing factor between systems if good NR is used, which makes it no more than an inconvenience.
Where a true difference remains is in dynamic range (DR). The two-stop difference remains unchanged in spite of NR. A Z7 has about 11.6 stops of DR at base ISO, whereas an OM-1 has about 9.6. There is nothing any piece of software could do about that. When shooting MFT, I find every so often that I get to choose whether to blow my highlights or get pitch black dark areas, but I won't be able to avoid both within a single shot.
When shooting landscapes, I can simply bracket my exposure and blend the shots later, which again (usually) makes DR a mere inconvenience. Shooting moving subjects, however, I have to live with the issue. Shooting an animal in thick underbrush, this is not much of a problem since such scenes usually do not have high DR anyway. Shooting a bird in flight against a bright sky can present a dilemma, however, that I may not be able to avoid. If all else were equal (which it isn't, but that's a different discussion), that would make FF the better choice.
In short, I am of the opinion that giving the state of technology we are at, DR warrants special attention in this debate, whereas sensor noise does not/no more.
Lothar
philzucker
Forum Pro
I'll try to answer that with an analogy first: I have an excellent and quite large Japanese kitchen knife and an also excellent Swiss Army pocket knife. Both cost about the same and are excellent tools. I can cut with both, and sometimes both do indeed cut the same, e.g. apples. What I do most often with them depends on different factors, like sharpness, length of the blade, object to cut, transportability, compactness etc. pp.
It's the same with cameras IMO. I do have FF, APS-C, m43 cameras and a 1/1.31 inch sensor smartphone. If I can plan ahead I choose the camera and the lenses best suited for the job. If I can't plan exactly what awaits me I take something like the above cited "pocket knife" that can do a lot of things, but not all of them as good as specialized tools.
- My photographic "pocket knives" are my smartphone and an APS-C camera with one or two zooms.
- I like to use m43 gear primarily for macro and well lit wild life scenarios with lenses specialized for that.
- FF I use almost exclusively with a few bright primes and especially for high dynamic range and low light scenarios and if there is need for subject separation in tight spaces.
My bottom line is: You can do a lot of different and high quality photography with any system, modern smartphones absolutely included. If you have multiple systems you simply have more options.
Phil
--
GMT +1
Gallery: http://photosan.smugmug.com
Last edited:
Interceptor121
Forum Pro
- Messages
- 12,604
- Solutions
- 8
- Reaction score
- 9,603
Noise reduction applies equally to all formats and therefore is irrelevant to a comparison
in addition in most cases on full frame you don’t need it saving time in post processing
in addition in most cases on full frame you don’t need it saving time in post processing
DeepPrime XD does a nice job on baby’s eyelashes in available light when you need more DoF. That’s on FF.
Andrew
lokatz
Veteran Member
You may need to make up your mind on which arguments to use. In the first statement, you imply that both systems benefit from NR the same way. In the second one, you argue that FF won't even need it, which is a direct contradiction to the first one.
Taking both together, I observe that you seem in full agreement with my main points:
1. MFT benefits from NR as it needs it much more than FF does.
2. Noise is an inconvenience on MFT because it often requires NR, but that's all there is to it. Once you've done proper NR, things are the same on both systems.
Glad we agree.
MEDISN
Senior Member
I thought the final image was shot with FF. I was looking at it on an iPad which allows you to open the image into viewer mode without the auto pop up EXIF window that you get on a computer. I would have never imagined that came from a GM1! Then it made sense with the frame stacking. In the end, I now see why you began you post with the "content matters" pretext. That is an excellent image from a composition and lighting perspective. Well done sir!
No harm, I ask a lot of questionsI do realize now I didn't answer your question about noise reduction...You have answered my question regarding the in camera stacking impact. A lot of good stuff in there, maybe we should have started a separate thread
The answer is I haven't really played with the more recent AI stuff enough to have an up to date opinion. What I will say, admitting it is out-dated, is:
So based on the last few times I checked in on fancy NR packages I wasn't impressed enough to bother using them on my base ISO landscapes (though they clearly work very nicely on a whole host of common images). I should try again, especially since Adobe recently integrated their version into LR.
- The pre-AI fancy NR stuff never really seemed to work well on base ISO mid-tone noise. I think this is because those algorithms were working best on structured pattern noise in high ISO shadows. Photon shot noise of course has no pattern or structure to it, it is truly random, so there is only so much you can do.
- The early AI NR seemed to work best on more obvious structures it was trained on such as feathers, fur, hair, material textures and larger scale structures like buildings. Seemed to do well on larger natural things like leaves when they were large enough to be identifiable. But a lot of landscapes down at the pixel level are almost fractal in nature (especially deserts without trees) and that seemed to not work so well.
. Perhaps I should not have called them NR packages - that might be the wrong term. I am thinking along the lines of opening an image in DxO PureRaw first which does the conversion from RAW to TIFF but with specific modules tuned to each camera and lens. Even with base ISO images it appears to have some effect on sky tones and shadow noise, sharpening, CA, colors, etc. The end result is an image that looks really good in terms of IQ. For example, your beautiful rainbow photo above with the EM5II...I wonder if letting PureRaw do the demosaicing would bring about a clean, color rich sky as a starting point for editing.
Now you've got me thinking about capturing some examples to look at the impact of mid-tone noise and tonality. Something like...
- M43 single shot, let camera meter
- M43 single shot, ETTR
- M43 LiveND
- M43 12 shot median stack
- M43 HHHR (should be similar to above)
- M43 HR (pixel shift)
- FF single shot, let camera meter
- FF single shot ETTR
- FF 12 shot median stack
- FF pixel shift
I will try down at the coast - we have some nice bluffs at the ocean which tend to have some nice shadows in the morning. Will have to work quickly though as the shadows shift. Anyway, will be fun to compare and discuss results!
kenw
Veteran Member
Good point, and I'd say depending on the context it is either an important factor or an irrelevant factor. But of course still one worth discussing!
When comparing sensor sizes NR applies equally to both sensors. In that simple context NR is irrelevant because while NR will make a smaller sensor look better it will at the same time make a larger sensor even better. This context would apply to folks shooting in conditions where they can barely get acceptable results on the smaller sensor and they'd still even like better results than they are getting from the larger sensor. That is to say with NR applied to the smaller sensor there is still visible noise they'd like to get rid of.
On the other hand, if the larger sensor without NR is producing results that need no further improvement, but the smaller sensor without NR is only marginally acceptable now NR is very relevant. In this case one could assert that by using NR on the smaller sensor they have an image for which they see no point to additional improvement from a larger sensor. Applying NR to the larger sensor isn't going to improve it any more, it was already plenty good enough. So in this context indeed improved NR makes all the difference between whether the larger sensor is "worth it" or not.
Of course there is a third case, that in which the smaller sensor is already perfectly acceptable with NR in the first place. In this context again NR is irrelevant - neither of the sensors need it.
So the key point is that assuming we have some "threshold of acceptability" that NR is only relevant in the comparison between sensor sizes for that region in which the larger sensor is above that threshold without NR and the smaller sensor can cross that threshold with NR.
I hope my previous post made clear that where such a threshold would be is a huge "it depends" kind of thing. So for some folks the advent of better NR probably has a large impact on their decision of the benefits of moving to a larger sensor while for others it is irrelevant.
Agree, in the past the side effects of NR were substantial enough that one could for almost all purposes dismiss it entirely in the comparison between sensor sizes.
Agree, right now the better NR solutions are right around the two stops of improvement that is equivalent to the jump between FF and m43 (at least for the base ISO stuff I work with, the improvements might be even more at higher ISO). But again, you can apply that same NR to the FF image. So this improvement in NR performance is most meaningful if one's threshold for acceptability happens to be right at the point where FF without NR is acceptable while m43 without NR is not acceptable.
Thanks for the excellent comparison images of the various AI NR engines!
Again this all comes down to enlargement sizes, viewing distance and threshold of acceptability. Looking at the NR results in your comparison it is abundantly clear that NR greatly improved all of them. Even after NR I'm quite a bit happier with the D850 image than the OM-1 image, but these are different subjects in different light and so forth and so that could be an idiosyncratic observation.
Here I'm going to disagree fairly strongly with the terminology and theoretical assertion. DR and noise are exactly the same thing. Dynamic Range is in fact defined by noise. It is the clipping level divided by the level at which the SNR drops to some arbitrarily chosen threshold. If NR software is improving noise in the shadows then it is increasing DR at the very same time - it is part of the definition of DR.
Now that doesn't necessarily change your practical assertion at all - that with FF you can get acceptable pushed shadows at base ISO that you can't from m43. In my experience a lot of this has to do with RAW processing itself rather than just noise (or equivalently DR). I've run into bonkers color shifts with heavily pushed shadows using LR/ACR profiles. Creating a custom RAW profile with an external tool can fix this, but it is a fight. So in a practical sense this is limitation of "DR" because the shadows fall apart in the m43 image but not in the FF image. But in what the actual definition of DR is (i.e. shadow noise) this is absolutely something that NR software can improve if implemented correctly and if the RAW processor profile can deal with color in the shadows.
I might be misunderstanding you here, or you might have some specific examples that illustrate the point better. My main message here is that assuming RAW processing and NR is working correctly that in reality NR software will in fact improve the DR of shots!
Yep, as a landscape shooter I have many options that don't work for other kinds of photography!
Which is largely why I didn't address NR in my previous post. For me, if I suspect that noise is going to be an issue I just deal with it in the field either through exposure averaging or bracketing. If I'm worried primary about pushing the shadows in a high dynamic range scene then a two or three shot bracket is all that is needed. If on the other hand I'm worried primarily about mid-tone noise then I prefer averaging and will capture eight to ten exposures in the field.
Could NR do the job for me instead? Certainly - but usually for my landscape shots bracketing or averaging is the more certain result with less fussing with sliders and looking for artifacts. Once anything is moving though (e.g. blowing leaves or branches) now the advantage tilts back to the NR solution.
Where I am really glad about newer NR solutions is when things go wrong and I get it wrong in the field. Now NR can sometimes save me!
Again, I'm not so sure of this because from the theoretical stand point it is provably wrong. DR and noise are the very same thing. However, from a practical standpoint in the evolution of AI based NR it is a very interesting question. If it is true that AI based NR algorithms at present don't improve base ISO DR because they fail to reduce noise in the shadows in the RAW file then that is important mark against AI based NR and an issue that should be addressed in the AI models probably via their training datasets.
This is a broader issue with current AI algorithms in general. If the training dataset and the validation datasets are not tagged or evaluated for a particular metric in a particular operating condition then the model will do nothing to address it, or potentially make it worse. So I would absolutely expect a good chance of what you describe happening with shadows and "DR" to occur with an AI model. A black shadow looks better than a noisy shadow - so don't bother applying any NR to shadows, just clip them to black instead. Again, I don't have enough direct experience to assert this is at all what is happening. But I'd be suspicious of that being the case if DR is not improved by AI NR because in fact any other NR technique does improve DR.
So I agree, the DR question with NR does warrant special attention!
As a last thought on AI NR this was a good morning to quickly play with Adobe's new AI based denoise to see how it does with base ISO images. Here is a slightly different crop of the averaging example I posted previously (as usual view the original):
On the left is a single exposure with AI NR at default settings (55 for the only slider Adobe gives you) and on the right is the multiexposure average without NR. You can see the AI NR did a really good job and the results are quite close. You can, however, see that the AI guessed wrong in some spots and introduced green artifacts around the dark rocks near the middle of the crop. In other regions the AR removed some subtle details that the averaged image retained. Really, none of these things at all "breaks" the image. The AI NR clearly made about a two stop improvement with a single click and minimal artifacting. I think this will be most useful to me in mid-day telephoto landscapes where I may still be dealing with mid-tone noise even in FF because of contrast/clarity/dehaze post processing on distant horizons.
Anyway, thanks for the insightful comments!
--
Ken W
See profile for equipment list
No Time To Lose
Leading Member
- Messages
- 560
- Reaction score
- 931
I think the last 3 posts have addressed a question that was forming in my mind on NR - do the modern NR tools help close the gap between M43 & FF sensors? The answer seems to be yes (completely?) when the NR software can reduce the noise either to a practical zero or to the point where it no longer causes offence in the final image.
If that point is not reached, the answer may be “somewhat” if the NR SW can reduce noise more at higher levels than at lower starting points - is this true?
with BIF, I think I understand that the situation is more complicated - the longer reach of M43 may lead to wider apertures (depending on the lenses used of course - let’s base the discussion on OMS 300mm f4 or BWL) and therefore lower ISOs than FF, reducing the noise delta between the initially captured images - at least for longer focal lengths?
If that point is not reached, the answer may be “somewhat” if the NR SW can reduce noise more at higher levels than at lower starting points - is this true?
with BIF, I think I understand that the situation is more complicated - the longer reach of M43 may lead to wider apertures (depending on the lenses used of course - let’s base the discussion on OMS 300mm f4 or BWL) and therefore lower ISOs than FF, reducing the noise delta between the initially captured images - at least for longer focal lengths?
kenw
Veteran Member
Thanks! And yes, a little GM1 can do amazing things with exposure averaging. It is all just about collecting enough photons.
Well, I just now just tried the ridiculously expedient if lazy technique of simply clicking the new Adobe AI "Denoise..." button in the Detail panel and letting it do its default thing with zero further changes to the processing:No harm, I ask a lot of questions
For example, your beautiful rainbow photo above with the EM5II...I wonder if letting PureRaw do the demosaicing would bring about a clean, color rich sky as a starting point for editing.
Adobe's AI Denoise set to 55
My original manual denoise settings of Lum 20, Color 38
So yeah, to me that soundly demonstrates that at least for the mid-tone noise topic I was addressing the new AI techniques work very well at base ISO just as they do at higher ISO. That's great news because even with FF there are times I wish I had exposure averaged (usually B&W conversions through haze at midday).
No doubt playing around with additional features within DxO could probably eek out even more.
If I wasn't clear in my original post, I always tell folks that for me reason to move to FF was simply about expediency. As demonstrated with the GM1 exposure average I could always get FF results from my m43 cameras, just with more hassle. FF just means I can be lazier in the field and in post processing. AI NR is now another tool in the quiver to enhance my laziness
So for me personally I'm unlikely to jump through DxO hoops on a regular basis, I'm too lazy, but of course there's always that mistake that you want to try to rescue.
Yes that would be an excellent thing to do, and to have it in a dedicated thread rather than hidden in our side discussion here!
I've done a fair number of those comparisons myself and pretty much things go exactly as you'd expect if you just consider what the aggregate exposure was (i.e. how many photons did you collect between all the sub-exposures). The one wrinkle I never fully decoded was what was the E-M5II really doing when it generated its HR (pixel shift) RAW files. I mean we decoded the broad outlines (e.g. it transformed into an internal working color space for the stacking where unexpected clipping could occur, and there was a bit of sharpening in its resampling kernel when it created its pseudo-Bayer RAW file) but I never took the time to track down what all was happening noise wise. It seemed that in a practical sense when you tried to avoid internal clipping in choosing an ETTR exposure level that you were getting a bit less improvement than if you had averaged eight normal ETTR exposures. But it was close for all practical purposes.
Anyway, if you have the interest and time I'm sure it would be an epic post like many of your previous HR posts!
--
Ken W
See profile for equipment list
kenw
Veteran Member
I think that is an accurate and succinct summary!
I think that is a "it depends" and one I certainly don't have enough experience to address. I will comment that AI algorithms tend to be non-linear, which is to say they usually have their own internal thresholds. With that in mind it is certainly possible that the magnitude of their benefits differ on how bad/good the starting point is.
I think this goes both ways, and especially for something like BIF you probably need to figure in the AF tracking performance. I'm thinking along these lines:
- If you can carry the weight you've potentially got a wider equivalent aperture available in the monster FF super-telephotos. If you can successfully focus and track those lenses wide open you'll end up with lower noise than with an m43 lens of smaller equivalent aperture. For example, the OMS 300/4 is equivalent to only 600/8 on FF. Meanwhile FF has 600/4 lenses available that will give two stops of noise advantage over the OMS solution, but of course at a huge size, weight and cost penalty.
- Even if you have lenses that open to equivalent apertures if one camera can't track AF as well you may need to stop down that lens to increase the DoF to the point you get enough "keepers" as far as focus goes. That means the camera with poorer AF will end up with noisier photos. So for example a Nikon Z9 could shoot BIF with narrower DoF than say a Nikon Z7 while still getting enough in focus images - same lenses, same sensor noise performance but potentially lower noise results from the Z9 because it could track focus at a wider aperture than the Z7 could.
Interceptor121
Forum Pro
- Messages
- 12,604
- Solutions
- 8
- Reaction score
- 9,603
Nothing has changesYou may need to make up your mind on which arguments to use. In the first statement, you imply that both systems benefit from NR the same way. In the second one, you argue that FF won't even need it, which is a direct contradiction to the first one.
Taking both together, I observe that you seem in full agreement with my main points:
1. MFT benefits from NR as it needs it much more than FF does.
2. Noise is an inconvenience on MFT because it often requires NR, but that's all there is to it. Once you've done proper NR, things are the same on both systems.
Glad we agree.
NR is irrelevant to the comparison both formats benefit the same
in addition because full frame has less noise in many cases you don’t even need to bother using it
there is no contradiction and I don’t agree with your incorrect statements
Things are not the same on both systems even with noise reduction because as explained it makes no difference to the comparison. With full frame you can take shots that even with noise reduction cannot be taken with a smaller sensor
In fact in the claims of the various DxO Topaz the benefits are stated equally regardless of format
--
instagram http://instagram.com/interceptor121
My flickr sets http://www.flickr.com/photos/interceptor121/
Youtube channel http://www.youtube.com/interceptor121
Underwater Photo and Video Blog http://interceptor121.com
Deer Photography workshops https://interceptor121.com/2021/09/26/2021-22-deer-photography-workshops-in-woburn/
If you want to get in touch don't send me a PM rather contact me directly at my website/social media
Last edited:
2ndact scene1
Senior Member
Modern NR (DxO, Topaz and LR) are really changing the playing field. But I think your two points actually demonstrate that its benefits are not equally distributed. If FF in most cases doesn't need NR, then there must be other formats that do need it. The quality of today's NR now gives those other formats the ability to (almost?) match FF.
For FF, in terms of "most cases", I seem to live in the world of "other cases" where NR is a critical tool.
Similar threads
Keyboard shortcuts
- f
- Forum