My comparison of sensor size and print quality

[myotis]

I began a very long post on this that included detailed methodology, but it felt far too long for anyone to read, so this is the abbreviated conclusion to my sensor comparison, which included making almost 100 prints.

Over the last few days, I have been making prints (A4 sections from 20" x16" prints) from Fuji GFX 50s files, Nikon Z8 files and Nikon Zf files. I knew there were big differences between these sensors when viewed on the computer, but I wanted to see how these differences translated to differences in a print. I already had an idea from existing prints, but I didn't have any direct comparisons.

Initially printed at Capture One default settings (using QImage One) the GFX prints were easily identified. Not because of sharpness, or differences in the amount of detail, but by the depth, smoothness and subtlety of tonality and colour, along with micro-contrast that made them just "look" better.

In addition, without any adjustments, the GFX prints had more detail in both the highlights and shadows than the Nikon prints.

However, at base ISOs (GFX and Zf: ISO 100 and Z8: ISO 64), with careful processing it was possible to make Nikon prints that were almost indistinguishable from the GFX prints. Careful processing often included "tiny" adjustments such as changes of 1/10 of a stop in exposure, or one unit of clarity or structure, selectively darkening a particular colour (e.g. just the greens) and bringing down highlights.

In spite of the higher resolution, it was much easier to match the Zf prints with the GFX prints, than matching the Z8 prints with the GFX prints and in our "blind" comparisons we actually selected some of the Zf prints as being "better" than the GFX ones.

But the Zf prints might have taken up to eight attempts before getting one as good as the GFX print. And, without having the GFX print as a clear quality target for the Zf print quality, I would probably have accepted any of the Zf prints I made as being "good enough".

These first tests were in bright, sunlit conditions at base ISO, so I did a second batch of comparisons at ISO 800, in dull and more challenging lighting. With these tests, I only used the Zf and the GFX. As I'm unlikely to use the GFX beyond ISO 800 I didn't consider any tests beyond this ISO.

Again it was easy to produce good quality prints from the ISO 800 GFX files, but matching prints from the Zf was more difficult (or maybe I was getting bored with the process by this stage). Subjects with large areas of important shadow and wide dynamic range that printed well from the GFX files I found impossible to match from the Zf files, even though I could still make a "good" print.

Even with the better lit subjects at ISO 800, matching the Zf prints was more difficult than in the previous tests, and I felt the noise in the Zf files was having an obvious detrimental effect on print quality.

So I ran the now excellent C1 AI denoising on the Zf files, which really lifted the quality, and on the better lit pictures I was back to being able to produce prints that were indistinguishable from the GFX prints.

My conclusion is that the GFX makes it easy to consistently produce high-quality prints with minimal processing effort across a range of lighting conditions and subjects.

However, with much more effort put into the processing, most of the time, the Zf seems capable of producing 20x16 prints indistinguishable from 20x16 prints made with the GFX.

So while, I am still going to prefer the GFX for landscape, my brief experiment has made me much more comfortable about using the Zf when using the GFX isn't possible. And while the Z8 is still the obvious choice for wildlife, it also seems it may not be the obvious GFX alternative that I thought it was for landscape (in terms of 20x16 prints).

With further tests, I may of course change my mind on all of this, but I've had my fill of testing for a while. I used to enjoy doing these things but now I just see it as chore :-(

 

[Barney]

Thanks for that, Interesting testing, were the shots all landscape type images ?

 

[myotis]

Thanks for that, Interesting testing, were the shots all landscape type images ?

Photographs of flowers, houses and gardens, some distant trees, grasses and bushes. No people, and no grand vistas, but still house pictures had open skies, as well as distant and close up vegetation. One shot was a telegraph pole with the sun directly behind it, and a hedge and car in front of it. This was one that I couldn't match the GFX quality with the Zf.

 

[Barney]

Good cross section there.

 

[myotis]

Good cross section there.

That's what I was trying for. What I want to do as a follow up, is some of the very high-contrast woodland pictures that I often do, and where I think the sensors will be pushed to their limits.

I have to say I am regularly impressed with the Zf sensor, and I love using the Zf (apart from the articulating rear screen).

 

[myotis]

Just when you thought you knew what you were doing :-(

For as long as I can remember, I have set Capture One to "disable all sharpening" and let QImage do all its clever stuff to give "perfect" print sharpening.

But as (hopefully) a last test, I thought I would compare prints from my 16mp cropped sensor Fuji X100s, with prints from the Nikon Zf. With a smaller sensor, smaller resolution, and not so brilliant lens, I thought this may well be the threshold where the prints from the Fuji would be noticeably worse than those from the Zf: and they were!

I could make them sharper, by adjusting the sharpening setting in QImage, but this didn't improve the detail; it just made them "appear" sharper and actually reduced the overall sense of image quality.

So to digress, but for good reasons, Capture One promotes a "three-stage" approach to sharpening.

1. Capture sharpening, which is the default settings and different for every sensor. It's possible to customise these settings.

2. Creative sharpening, where you selectively sharpen (on layers) specific parts of the photograph.

3. Output sharpening, where you can adjust the sharpening applied on export depending on whether it's for online use or printing. And, in the case of the print sharpening, you can set the viewing distance of the prints and C1 will adjust the sharpening accordingly.

But by choosing the disable all sharpening option, when sending to QImage, all of the C1 sharpening is ignored.

After several attempts, I was having little success with the QImage sharpening (plus changing clarity and structure settings in C1) on the Fuji prints, so I tried using the C1 default "capture sharpening" plus some low-level QImage sharpening.

I've only done this on a single file, but the Fuji print is now looking sharper and showing more detail than the NIkon print.!

I now feel the need to redo all the tests I described in my first post :-(

Having said that, I am fairly confident that what I learned won't change, which is, for 20"x16" prints, it doesn't seem to make much difference to print quality, whether you are using a 16mp cropped sensor full frame camera, or a 50mp cropped sensor medium format sensor (or anything in between).

But, as sensor size gets bigger, the time needed for processing gets smaller.

 

[AndrewFlannigan]

While I understand that some people are determined to achieve "perfect sharpness" and that's fine by me, I long ago decided that a sharp boring picture is no match for a fuzzy but fascinating picture.

Of course, if you can achieve both high technical quality with a really interesting image, then you've ensured your ticket to photographic Nirvana.

 

[myotis]

While I understand that some people are determined to achieve "perfect sharpness" and that's fine by me, I long ago decided that a sharp boring picture is no match for a fuzzy but fascinating picture.

Of course, if you can achieve both high technical quality with a really interesting image, then you've ensured your ticket to photographic Nirvana.

I, of course, agree, but this exercise had nothing to do with achieving perfect sharpness. None the less, I still think it might be interesting for anyone who is concerned about sharpness, assuming they are interested in "print" sharpness, and not comparisons made on a computer screen at 100%.

For at least in terms of 20x16 prints, sensor size didn't make any noticeable difference to sharpness, except for the X100s, which wasn't part of my original comparison, but suddenly made sharpness relevant.

This was probably a flaw in my method, as I didn't actually try to optimise the sharpness. I simply accepted the "default" sharpness from the GFX file as my starting point, because I was more than happy with the sharpness it was giving me.

Using the GFX defaults gave me a "fixed" starting point and as the sharpness affected how the important aspects of the pictures looked, ie the overall look and feel of the print, I still needed to match this aspect of the prints for a fair comparison.

I also knew from experience that the main difference between the sensors, at least for me, wasn't sharpness, but the differences in tonal and colour gradation, which are obvious even when casually comparing prints. This immediately identifies the prints from larger sensors compared to smaller sensors because the larger sensor prints just "look better".

This was the real challenge when trying to make matching prints. With the initial sets of comparative prints, my wife took literally seconds to separate the GFX prints from the Nikon prints. It took lots of tiny changes in processing before I managed to make prints that were difficult, or even impossible, to identify the sensor behind the print.

 

[Garry Edwards]

Excellent information. Can I make a suggestion?

Post it in the "Resources" section, where it can always be found and referred to.

 

[myotis]

Excellent information. Can I make a suggestion?

Post it in the "Resources" section, where it can always be found and referred to.

Thanks Garry, I'm happy enough to do that, but might tidy up the writing first :-(

 

[Clive K]

I am fascinated by this and thank you for sharing the results of your processing experiments.

I have two questions. Firstly regarding the 'highly tweaked' prints from smaller formats being equal or near to equal of the 'lesser tweaked' from larger formats; have you tried comparing the small sensor images against similarly 'highly tweaked' images from larger formats? Could you extend the margin that improving the lesser formats has narrowed?

Secondly. Has anyone other than you and your wife given opinions on the images? I am not doubting your judgements, but repeated viewing of very similar images can play tricks with minds. I appreciate that choices are subjective as seen when digital prints began to replace wet prints, but to let someone view a set of prints with fresh eyes might be beneficial.

 

[sk66]

But, as sensor size gets bigger, the time needed for processing gets smaller.

??

I also knew from experience that the main difference between the sensors, at least for me, wasn't sharpness, but the differences in tonal and colour gradation, which are obvious even when casually comparing prints. This immediately identifies the prints from larger sensors compared to smaller sensors because the larger sensor prints just "look better".

I always wonder about the multitude of variables that come into play with something like this; especially when we are talking about something aesthetic and subjective (i.e. "rendering"). Also, Capture One started out as a Phase One product optimized for MF; so maybe there is something specific to that. And I'm pretty sure that they have worked fairly closely with Fuji in regards to their MF cameras... e.g. it can automatically apply the film sim set in the camera; it doesn't have that kind of functionality/interoperability with Nikon files/picture controls.

I imagine the results would be different if you had started with LR and used the same demosaicing defaults for all of them (e.g. Adobe Standard)... of course, most Nikon users these days would tell you to avoid using an Adobe profile. And TBH, I don't think you can even get to "the same demosaicing" unless you set your own process (e.g. using MATLAB).

I'm also not sure that using one rendition as "the standard" is quite fair/optimal, as opposed to some desired artistic goal/look that they all need adjusted to meet. I.e if sharpness or color accuracy was more relevant to your goal then one of the other sensors might have the advantage (Fuji makes some wild claims about their X-Trans CFA). I mean, of course it is hard to exactly replicate the result of one sensor/processing (especially when you don't know what was actually done, and a different lens was used). But perhaps it is much less difficult to achieve your aesthetic/artistic goals (you've already insinuated that it is).

The larger sensor definitely has the advantage in terms of the amount of light recorded; and it renders a scene differently with the same exposure settings (DOF)... but other than that, I don't know what can actually be attributed to sensor size alone (or even primarily).

 

[myotis]

I am fascinated by this and thank you for sharing the results of your processing experiments.

I have two questions. Firstly regarding the 'highly tweaked' prints from smaller formats being equal or near to equal of the 'lesser tweaked' from larger formats; have you tried comparing the small sensor images against similarly 'highly tweaked' images from larger formats? Could you extend the margin that improving the lesser formats has narrowed?

Possibly, I did actually mention this as a flaw in the methodology. The prints from the un-tweaked GFX file looked really good to me, and from a technical standpoint, but mabe not a creative one, I didn't feel the need for any tweaking, other than minor adjustments in exposure.

And this was, in some respects, the point of the comparison for me, because on screen the GFX files alway slooked so much better than the Nikon files, I was convincing myself, that maybe I should be making a bigger commitment to the GFX system.

Secondly. Has anyone other than you and your wife given opinions on the images? I am not doubting your judgements, but repeated viewing of very similar images can play tricks with minds. I appreciate that choices are subjective as seen when digital prints began to replace wet prints, but to let someone view a set of prints with fresh eyes might be beneficial.

No one else has looked at them, but on every comparison, we independently came to exactly the same conclusions.

I've done tests like this twice before, (when comparing the results from C1, LR and DXO) when I extended the number of assessors to ten people, and it was amazing how consistent people's views were.

But other than "food for thought," there are so many variables that aren't being properly considered, and it's only based on a bit more than 100 prints (now), that it can't be considered as all that definitive.

A statistically valid comparison is way beyond my budget, the time I am willing to put into it, and my expertise. I could do the stats bit, but not sure how to manage and measure the variables.

But I think for me, it has served it's purpose.

 

[myotis]

??

The files from the larger sensor being easier to process was something I expected.

I have extensively used cameras with 1", M43, APS, FF, and (to a much lesser extent) cropped medium format, and there has a been a clear correlation between how easy it is to get desired results and sensor size. It may well be that sensor size correlates with some other factor, and its this "other" factor that is making processing easier, but it's sensor size that is easily identfiable.

And to anticipate your next paragraph, this has been my experience with LR and DXO.

I always wonder about the multitude of variables that come into play with something like this; especially when we are talking about something aesthetic and subjective (i.e. "rendering"). Also, Capture One started out as a Phase One product optimized for MF; so maybe there is something specific to that. And I'm pretty sure that they have worked fairly closely with Fuji in regards to their MF cameras... e.g. it can automatically apply the film sim set in the camera; it doesn't have that kind of functionality/interoperability with Nikon files/picture controls.

Capture One worked "very closely" with Fuji and the Fuji profiles. When C1 was asked about not providing profiles for some of the older Fuji models, they said it was up to Fuji whether they would become available in the future. My GFX 50s and X100s are cameras that don't have Fuji/C1 collaborated profiles, and these gave very different starting points from the Nikon cameras, which use the C1 Pro standard profiles. The two Fuji cameras were also very different from each other.

I therefore splashed out on Cobalt camera profiles, which gave almost identical starting points for all four cameras.

I'm not sure if my assessment of rendering is based on aesthetics, for example I was looking at things like colour gradation where it was obvious that the range and smoothness of colour transitions in something like a pink flower petal were much better than that same petal from the Nikon files.

I imagine the results would be different if you had started with LR and used the same demosaicing defaults for all of them (e.g. Adobe Standard)... of course, most Nikon users these days would tell you to avoid using an Adobe profile. And TBH, I don't think you can even get to "the same demosaicing" unless you set your own process (e.g. using MATLAB).

Could well be, and printing directly from C1, rather than through QImage, might also change the results, but this was a test for my workflow and not meant to be a definitive scientific study on sensor sizes

I'm also not sure that using one rendition as "the standard" is quite fair/optimal, as opposed to some desired artistic goal/look that they all need adjusted to meet. I.e if sharpness or color accuracy was more relevant to your goal then one of the other sensors might have the advantage (Fuji makes some wild claims about their X-Trans CFA). I mean, of course it is hard to exactly replicate the result of one sensor/processing (especially when you don't know what was actually done, and a different lens was used). But perhaps it is much less difficult to achieve your aesthetic/artistic goals (you've already insinuated that it is).

Except, going back to your first question, this was very much the point. I knew from experience that larger sensors gave me a much better/easier starting point for processing than smaller sensors did, but I had never done any structured testing. And, the defaults from the GFX files (apart from minor tweaks in exposure) gave prints that I would consider commercially acceptable with no further work needed. This was not the case from the Nkons or the x100s.

And of course, I was interested in whether obvious quality differences (such as sharpness and detail) visible on screen translated into "unfixable" differences in prints.

And yes, there are a lot of variables, but these are tests for my workflow, and only presented as "food for thought" to others, as I thought it might be useful.

The larger sensor definitely has the advantage in terms of the amount of light recorded; and it renders a scene differently with the same exposure settings (DOF)... but other than that, I don't know what can actually be attributed to sensor size alone (or even primarily).

I have always thought it was related to photo site size, and I thought it interesting that I found it easier to match the Zf files to the GFX, than to match the Z8 files to the GFX.

 

[sk66]

The files from the larger sensor being easier to process was something I expected.

Ah, I was thinking in terms of resolution/file size.

I have always thought it was related to photo site size, and I thought it interesting that I found it easier to match the Zf files to the GFX, than to match the Z8 files to the GFX.

Probably in terms of color... I.e. if you have a point of color (airy disk) that is sampled by a set of all four RGBG Bayer pixels it will be rendered the most accurately. Whereas if you have four points of differing color that all fall on one much larger pixel they won't be rendered at all; they would be combined into a single homogenized point/color. And since the CFA isn't terribly exclusive, what color the mix should be would be very difficult to determine on it's own (so neighboring pixels are used).

Basically, I think as resolution increases (pixel size decreases) technical accuracy improves (color, removal of AA filters, etc), but at the cost of "filmic character." I.e. for a more film like reproduction I might choose to reduce clarity notably (and maybe offset with small increase in texture).

 

[myotis]

Ah, I was thinking in terms of resolution/file size.


Probably in terms of color... I.e. if you have a point of color (airy disk) that is sampled by a set of all four RGBG Bayer pixels it will be rendered the most accurately. Whereas if you have four points of differing color that all fall on one much larger pixel they won't be rendered at all; they would be combined into a single homogenized point/color. And since the CFA isn't terribly exclusive, what color the mix should be would be very difficult to determine on it's own (so neighboring pixels are used).

Basically, I think as resolution increases (pixel size decreases) technical accuracy improves (color, removal of AA filters, etc), but at the cost of "filmic character." I.e. for a more film like reproduction I might choose to reduce clarity notably (and maybe offset with small increase in texture).

I understand the argument, but have also noted that as sensor resolution has increased, it also seems to come with complaints of poorer colour. e.g. when the GFX 50MP sensor was replaced with the 102MP sensor, but as always this is also confused by changes in the in camera processing, and 14 bit vs 16 bit colour.

However, In my example, the GFX 50s has roughly the same resolution as the Z8 (slightly more) but with a pitch size of 5.33 compared to 4.35 for the Z8, and as I understand it, larger photo sites will sample colour more accurately than smaller photosites, so you might expect the GFX of have the potential of "better" colour than the Z8.

And I would argue from my tests that the Zf with a lower resolution but with larger photosites (5.94) is giving colour closer in quality to the GFX, than the Z8 did.

But we are wandering away from the point of my tests, which was, regardless of any measured technical differences between the sensor sizes and resolutions, some of which might be seen on the computer screen; was I benefiting from these differences when making 20 x 16 prints? Not that i actually make prints as big as this, but coming from the film days, 20 x16 always seemed to be the quality standard.

I realise, as @Clive K mentioned, and I had mentioned earlier, that my comparison suffered from not trying to optimise my GFX baseline prints. Had I done this, maybe my conclusions would have been different. But I didn't feel it necessary to improve on the GFX default prints, and it gave a more stable baseline for the comparisons.

 

[sk66]

larger photo sites will sample colour more accurately than smaller photosites, so you might expect the GFX of have the potential of "better" colour than the Z8.

Photosites don't really sample color, they simply record illuminance. The raw file is simply a bunch of brightness values with a header entry telling what the color pattern of the CFA is. And the CFA isn't really R/G/B so much as it is short/medium/ong. I.e. a green filtered (medium) photosite also responds to wavelengths of blue and red light. Demosaicing to a specific color requires additional pixels of different brightness (light past their corresponding filter). I.e. you cannot increase data accuracy with reduced sampling/reduced data.

Larger pixels benefit from more light when they are also on a larger sensor (fill factor also plays a role)... But until you get to the point of viewing a pixelated image, it is light/image area that matters; not light/pixel (which is why larger sensors win regardless of resolution). If you have more pixels/area, they are simply combined in the output. And even a 20" print isn't enough to cause pixelated output. E.g. the CoC standard for image sharpness requires less than 2MP resolved on any sensor size (or negative); which is why resolution is never part of the sharpness/DOF calculation.

But we are wandering away from the point of my tests, which was, regardless of any measured technical differences between the sensor sizes and resolutions, some of which might be seen on the computer screen; was I benefiting from these differences when making 20 x 16 prints?

I understand the purpose as explained later, and it is certainly valid for you/your use... how can it not be?
But all of the other vagaries (ADC bit depth/accuracy, ISO bit depth/accuracy, CFA demosaicing algorithm, etc,) keeps me from accepting it as just a large sensor advantage on the face of it.

 

[myotis]

Photosites don't really sample color, they simply record illuminance. The raw file is simply a bunch of brightness values with a header entry telling what the color pattern of the CFA is. And the CFA isn't really R/G/B so much as it is short/medium/ong. I.e. a green filtered (medium) photosite also responds to wavelengths of blue and red light. Demosaicing to a specific color requires additional pixels of different brightness (light past their corresponding filter). I.e. you cannot increase data accuracy with reduced sampling/reduced data.

I of course know that the photosites are only recording luminance, but I thought the advantage of larger photosites was because they reduced the effects of photosite crosstalk due to optical and electrical leakages between the photosites which degraded colour accuracy and resolution.

I am assuming that as photosite density has increased; they have also found cleverer ways of dealing with this problem, but has it been fixed to the point of it no longer being an issue?

EDIT: I found this, but not yet read it, which discusses this, along with fixes (it's from 2022)

Crosstalk Correction for Color Filter Array Image Sensors Based on Lp-Regularized Multi-Channel Deconvolution - PMC

In this paper, we propose a crosstalk correction method for color filter array (CFA) image sensors based on Lp-regularized multi-channel deconvolution. Most imaging systems with CFA exhibit a crosstalk phenomenon caused by the physical limitations ...

pmc.ncbi.nlm.nih.gov

Larger pixels benefit from more light when they are also on a larger sensor (fill factor also plays a role)... But until you get to the point of viewing a pixelated image, it is light/image area that matters; not light/pixel (which is why larger sensors win regardless of resolution). If you have more pixels/area, they are simply combined in the output. And even a 20" print isn't enough to cause pixelated output. E.g. the CoC standard for image sharpness requires less than 2MP resolved on any sensor size (or negative); which is why resolution is never part of the sharpness/DOF calculation.

Yes, which was one of the reasons why I was making prints for a "real world" comparison, and not making comparisons on the computer screen at a higher magnifications than I was ever going to use. And as I said, sharpness (except with the X100s which I have put down to a poorer lens) didn't play any part in the comparison, because it was more or less equally good at the print magnification used, across all the cameras compared.

I understand the purpose as explained later, and it is certainly valid for you/your use... how can it not be?
But all of the other vagaries (ADC bit depth/accuracy, ISO bit depth/accuracy, CFA demosaicing algorithm, etc,) keeps me from accepting it as just a large sensor advantage on the face of it.

I haven't argued that it's just the sensor size, but that it's the sensor size that is the most easily identifiable factor.

The obvious differences I have casually picked up that are correlated with sensor size are more noticeable than the much smaller and subtler differences I have picked up between different cameras with the same sensor sizes (and resolution). And I have put those differences down to a combination of the factors (and others) that you have listed.

And as some sort of slightly dodgy evidence of the influence beyond sensor and photosite size, as I did no structured testing, my experience with my Nikon 1s was interesting,

In my mind, there was a steady decline in subjective overall image quality from the Nikon V1 to the Nikon V2 to the Nikon V3.

All were the same 1" sensor size but with increasing resolutions: 10MP to 15Mp to 20Mp. However when the 20mp J5 was introduced (after the V3) , it gave noticeably better image quality than the V3, even though it had the same sensor resolution and size.

View attachment 489190

 

[Clive K]

There have been comments about the Phase One, Haasleblad and Pentax cameras producing images of different qualities from the same sensor. There was a suggestion that the image processors were different.

Another can of worms?

 

[myotis]

There have been comments about the Phase One, Haasleblad and Pentax cameras producing images of different qualities from the same sensor. There was a suggestion that the image processors were different.

Another can of worms?

Yes that is "normal" as the sensor is only one of the components in the processing pathway, and while the sensor (in the case of the cameras you mention) are all made by Sony, all the other components are independently made by the camera manufacturers or specified as custom options at the sensor manufacturing stage.

As an aside, currently Hasselblad, Pentax, and Fuji all have cropped medium format sensors, with the Phase One sensors being considerably bigger (full-frame medium format). Hasselblad used to make full-frame medium format cameras and Phase One used to make cropped sensor medium format cameras, but these are no longer made.

I don't think that sharing the same sensor across different cameras has ever meant getting the same image quality across those same cameras.

Nikon is possibly a good example of this, as they often seem to re-use the same sensor for a newer camera model, but still achieve a noticeable improvement in image quality from the new camera.

 

[myotis]

I'm a bit reluctant to post this because comparisons on screen are not the same as comparing prints.

The differences on the prints are much more obvious than they appear on screen, and the differences I mention here are the same differences that occurred across multiple comparisons.

None the less, I think they show some of the subtle differences in "starting point" between the Z8 file (753 in the file name) and the GFX file (942 in the file name).

Both are using Cobalt camera profiles which are very good at giving very similar starting points across different cameras.

The Nikon file has had a reduction in exposure of 0.2 stops to better match the overall density of the GFX file. Apart from that, the C1 default processing settings were used (auto-adjust), which, although I don't normally use, I have found to be very good.

In my final good enough "matched" prints with these two files, I selectively altered the exposure, selectively altered the density and saturation of the greens and added a small amount of clarity and structure. It took me eight attempts to get a satisfactory match. ie when the prints were close enough to begin making mistakes over which camera they came from.

But the purpose of this comparison is not to show how close the results can be, but to show the slightly better starting point you get from the GFX than from the Z8.

The Qimage print to file option was used to make the file I've posted, so both images received exactly the same (or at least similar) quality contribution from Qimage as my comparison prints got.

To sum up the differences between the prints, I consider those from the GFX show slightly better saturation and contrast than the Z8 prints while also keeping more shadow and highlight detail. The higher contrast (micro-contrast) also better defines detail on the GFX prints.

Which, I think you can see from the pictures I've posted.

But, I repeat that the differences between the prints are much more obvious, and at this stage in the comparison, it only takes seconds to separate the Z8 prints from the GFX ones.

Both pictures were taken minutes apart on a sunny day with a totally cloudless sky, so it is unlikely the lighting changed between shots. I used a 35mm to 70mm Fuji lens (at 45mm and f8) on the GFX, and the Voigtlander 35mm f2 Apo Lanthar (at f5.6) on the Z8.

EDIT: I've been looking at these pictures on my laptop (rather than the desktop I use for processing) and I don't think they show the differences I have described in the text above at all well :-(

 

[sk66]

but I thought the advantage of larger photosites was because they reduced the effects of photosite crosstalk due to optical and electrical leakages between the photosites which degraded colour accuracy and resolution.

Getting into realms I don't know much about...

My understanding is that electrical crosstalk is "known;" it's part of what makes the photosites S/M/L instead of RGB. And I believe it's why the green channels are kept separate as GR/GB for demosaicing. That in itself is not a significant problem (it's how our eyes work). But optical crosstalk is variable and unknown, which does make it problematic... i.e. it reduces the resolution and accuracy below what it could be; but that is not the same thing as reducing it below something else. My understanding is that BSI sensors, better microlenses (higher fill factors/less spill), etc, have reduced optical crosstalk; but it can still be a problem when photosites are light limited. Primarily the issue is a problem for what they call "small pixel sensors," which are sub micron sizes.

There's a lot here I don't fully understand... e.g. how does optical blurring of the light play into it (colors mixing due to diffraction, AA filters, etc), and how things like quad bayer and pixel binning work with color contamination (crosstalk). My best guess is that the issues are actually quite minor in the large pixel models (lots of light relative to errors).

This is a paper from 2014 on the issue joint authored by Eric Fossum (inventor of the CMOS sensor)... perhaps I will ask him about this issue directly to see what the current state is.

https://opg.optica.org/josaa/fulltext.cfm?uri=josaa-32-1-28

In my mind, there was a steady decline in subjective overall image quality from the Nikon V1 to the Nikon V2 to the Nikon V3.

All were the same 1" sensor size but with increasing resolutions: 10MP to 15Mp to 20Mp. However when the 20mp J5 was introduced (after the V3) , it gave noticeably better image quality than the V3, even though it had the same sensor resolution and size.

Yes that is "normal" as the sensor is only one of the components in the processing pathway, and while the sensor (in the case of the cameras you mention) are all made by Sony, all the other components are independently made by the camera manufacturers or specified as custom options at the sensor manufacturing stage.

As you noted, there are so many different variables... a change in the CFA dyes, a change in the microlens design, reductions in system noise (a big factor), or just a change in the demosaicing priorities (WB correction levels, white point/black point, etc).

I haven't argued that it's just the sensor size, but that it's the sensor size that is the most easily identifiable factor.

The title/original post gave me that impression; further posts made it clear that was not the case. But I don't know how it's useful as a resource other than as a subjective comparison of your cameras for your use, and for anyone else in the same position.

I'm a bit reluctant to post this because comparisons on screen are not the same as comparing prints.

I can see some differences, moreso if I open the image in PS and enlarge it. But I'm not certain which is which (I suspect the top is the Fuji), or if the Z8 image has been edited for "best match" or if these are the starting points. Either way, I don't think it much matters.

 

[myotis]

Getting into realms I don't know much about...

Nor me, and I will certainly know less than you do. I'm just going by what I have read when people try to explain why larger sensors produce better results than smaller ones. I will read the paper you linked to. Thanks.

My understanding is that electrical crosstalk is "known;" it's part of what makes the photosites S/M/L instead of RGB. And I believe it's why the green channels are kept separate as GR/GB for demosaicing. That in itself is not a significant problem (it's how our eyes work). But optical crosstalk is variable and unknown, which does make it problematic... i.e. it reduces the resolution and accuracy below what it could be; but that is not the same thing as reducing it below something else. My understanding is that BSI sensors, better microlenses (higher fill factors/less spill), etc, have reduced optical crosstalk; but it can still be a problem when photosites are light limited. Primarily the issue is a problem for what they call "small pixel sensors," which are sub micron sizes.

This is mentioned in the paper I linked to.

There's a lot here I don't fully understand... e.g. how does optical blurring of the light play into it (colors mixing due to diffraction, AA filters, etc), and how things like quad bayer and pixel binning work with color contamination (crosstalk). My best guess is that the issues are actually quite minor in the large pixel models (lots of light relative to errors).

This is a paper from 2014 on the issue joint authored by Eric Fossum (inventor of the CMOS sensor)... perhaps I will ask him about this issue directly to see what the current state is.

https://opg.optica.org/josaa/fulltext.cfm?uri=josaa-32-1-28

As you noted, there are so many different variables... a change in the CFA dyes, a change in the microlens design, reductions in system noise (a big factor), or just a change in the demosaicing priorities (WB correction levels, white point/black point, etc).

Indeed, but the reviews I've seen comparing larger sensors to smaller ones (regardless of camera make) often refer to superior colour and tonal gradation, which is what really interested me, because I was aware of the same issue from the film days when comparing different film formats. I was less interested in sharpness, as I was fairly confident that all my cameras would be "sharp enough" for a 20X16 print. But I was very conscious of differences in tonal and colour gradation between different cameras.

The title/original post gave me that impression; further posts made it clear that was not the case. But I don't know how it's useful as a resource other than as a subjective comparison of your cameras for your use, and for anyone else in the same position.

Well yes this was the basis of the comparison, to see how well I could match the results from the GFX with the results from my Nikons. And yes, I used the sensor size as the defining difference between the cameras.

I suspect, If someone buys a Fuji GFX or a Hasselblad X2d, or a Phase One, they are buying it with an expectation that larger sensors will give higher quality results than smaller ones and will set the baseline quality for the other parts of the processing pipeline.

I controlled for as many variables as I could, and then the challenge was to see how far I could adjust the files during processing to correct for all the other factors that might be affecting print quality. I confess to seeing this as the biggest factor affecting image quality that I couldn't control.

And yes, as I've made clear, it's a subjective test of the cameras I own, for my own purposes. But some of my results surprised me, which I thought worth sharing.

I can see some differences, moreso if I open the image in PS and enlarge it. But I'm not certain which is which (I suspect the top is the Fuji), or if the Z8 image has been edited for "best match" or if these are the starting points. Either way, I don't think it much matters.

The clue is in the sentence (near the beginning of the post):

"None the less, I think they show some of the subtle differences in "starting point" between the Z8 file (753 in the file name) and the GFX file (942 in the file name)."

But it's interesting that you have guessed the wrong way round.

When printed, the superiority of the print from the Fuji file is striking. Maybe you have confirmed the benefits of moving the comparison off the computer and onto actual prints.

The lens on the Z8 is clearly a better lens than the "cheap" Fuji zoom on the GFX, so I wouldn't be surprised if it gave sharper results when compared on the computer. I haven't made this comparison.


EDIT a quick google suggests that crosstalk starts to become an issue when photo site centres are less than 3 -4 micrometres apart.

 

[Clive K]

There is a lot that we can take from these tests and it is good of you to share them.

From a layman's point of view while it is incredible that the Nikon results can be made at least as good as the standard Fuji images, the elephant in the room however is how the tweaked Nikon images can compare with similarly tweaked Fuji results. Also we have to consider the mediums used for viewing. If for example the Fuji images are being capped by the viewing mediums; prints or screen, then there might be potential for them to exceed the Nikon if the size of prints for example exceeded the Nikon's pixel count at 300ppi.

 

[myotis]

There is a lot that we can take from these tests and it is good of you to share them.

From a layman's point of view while it is incredible that the Nikon results can be made at least as good as the standard Fuji images, the elephant in the room however is how the tweaked Nikon images can compare with similarly tweaked Fuji results. Also we have to consider the mediums used for viewing. If for example the Fuji images are being capped by the viewing mediums; prints or screen, then there might be potential for them to exceed the Nikon if the size of prints for example exceeded the Nikon's pixel count at 300ppi.

Thanks, and I agree with your comments that my tests leave unanswered questions, but this was by design, to suit my purposes.

I consciously capped the quality at 20" x16" print quality, because I wanted to avoid getting onto a never ending tread mill of whether camera A is better than camera B, rather than deciding quality standard B is more than good enough for my anticipated needs.

And I deliberately didn't compare results on screen because I know from existing casual comparisons that differences, easily identified on screen, aren't visible, or barely visible, on print, and for me, it would have been an interesting but pointless exercise.

To start tweaking the Fuji files, and then also try to tweak the Nikon files to match them, would have been an undertaking way beyond my budget: both in time and money.

My personal take away, is that I will still prefer the Fuji GFX when it's the obvious choice, e.g intimate landscapes on a tripod where image quality is particularly important to me, but I am a lot less concerned than I was, about using the Z8 or Zf, when the Fuji GFX is inconvenient, but I still want the level of image quality I associate with the Fuji GFX files.

But this Fuji GFX image superiority assessment was mainly down to comparing images on-screen, which is what prompted this print comparison.

I still think it's going to be "easier" making high-quality prints from the GFX files than from the Nikon files.

 

[sk66]

Indeed, but the reviews I've seen comparing larger sensors to smaller ones (regardless of camera make) often refer to superior colour and tonal gradation, which is what really interested me, because I was aware of the same issue from the film days when comparing different film formats.

I do believe that "more light" gives a better result... and in that sense the larger format always wins. And you hear similar things about certain lenses rendering colors better/different; or even Nikon vs Sony vs Fuji (etc). And some of that is certainly true. But color and tonal rendition with digital is a tough one for me... I mean, just choosing to demosaic using the Camera Standard profile instead of the Camera Neutral profile makes a much bigger difference in most cases. And adjusting the color calibration applied can be huge as well.

One thing I found interesting is that your GFX50S only has ~65% fill factor. That's extremely low by today's standards (BSI sensors like the Z8's are over 90%, approaching 100% for smaller pixels).
Basically, at least partially due to crosstalk, they only used the center of each pixel (other things affect fill factor as well). So while they are physically larger, they are much less so functionally (~ 3.5um effective). And the increased pixel spacing (effective) also causes an increase in the probability of aliasing (false detail). And yet you still find the results so much more preferable...

I think a lot of it must come down to "naturalness" in feeling/perception simply due to reduced manipulation... although I can't explain it. But take color for example, when you view a flower the perceived saturation, gradation, texture/tonality is not the reality... if you look closer and pay critical attention you would see it quite differently. And I think that's probably what smaller pixels/sensors do in this comparison (and sharper lenses for that matter)... it's like more technical accuracy at the expense of aesthetics and "perceptual reality." The technically perfect image that says nothing and moves noone...

 

[myotis]

I do believe that "more light" gives a better result... and in that sense the larger format always wins. And you hear similar things about certain lenses rendering colors better/different; or even Nikon vs Sony vs Fuji (etc). And some of that is certainly true. But color and tonal rendition with digital is a tough one for me... I mean, just choosing to demosaic using the Camera Standard profile instead of the Camera Neutral profile makes a much bigger difference in most cases. And adjusting the color calibration applied can be huge as well.

Do camera profiles in Adobe affect the demosaicing?

You can choose from multiple demosaicing algorithms in Darktable and Rawtherapee, but don't profiles just apply a tone curve(s) to the demosaiced RGB file.

That aside, camera profiles certainly make a big difference in how a file is rendered. I have made my own in the past (using LumaRiver). I prefer starting with a flat or linear profile over some of the more vivid options, as I find it easier to add contrast and saturation than to take it away. And going back to colour gradation, this was the big improvement I found when making my own profiles, most notably in the greens ( Capture One and Nikon files, probably with a D7000 at the time).

My limited experience (only since beginning this sensor comparison), but as advertised, has found the Cobalt camera profiles to be very good at removing differences between camera sensor renderings, and Capture One has a tool to create "LCC" profiles, which I've never used, to remove (along with other fixes) any colour differences between lenses.

It's interesting that with the Tamron designed, but Nikon built Z lenses, one difference is that the Nikon versions have different lens coatings (along with using Nikon AF motors, additional weather sealing, and different body composites).

There are a lot of variables when it comes to colour!

One thing I found interesting is that your GFX50S only has ~65% fill factor (extremely low by today's standards). Basically, at least partially due to crosstalk, they only used the center of each pixel (other things affect fill factor as well). So while they are physically larger, they are much less so functionally (~ 3.5um effective). And the increased pixel spacing (effective) also causes an increase in the probability of aliasing (false detail). And yet you still find the results so much more preferable...

That is interesting, and I see that in the DPReview thread that Phase One did something similar at one time(maybe still do).

I think the "so much more preferable" is overstating my argument a little.

The prints made from files processed with Capture One defaults (but using Cobalt camera profiles) gave prints that were clearly "more pleasing" from the GFX compared to those from the Nikons. But as I've said in other posts, the prints from the Nikons were still "good prints", which if I hadn't had the Fuji prints to compare them with, I would have been very happy to have as a starting point for further editing. But without further editing, and on close inspection, the prints from the GFX files showed more detail, better colour and tonal gradation, and more contrast and saturation, while not looking harsh or over-processed, and the GFX results, became the target for what I wanted to match with the Nikon files emulating these

The need to create up to 8 prints from the Nikon files, wasn't because the files needed 8 attempts to get an excellent print, it was because it took 8 attempts to get a print that was difficult to distinguish from the print from the GFX file. The exercise was to see whether the print quality easily attained by the larger Fuji sensor was unattainable by the smaller Nikon sensors, or whether, with careful processing, I could match the Nikon prints to the Fuji prints. Which indeed I could do in all but two examples where the lighting was particularly difficult and I couldn't quite make the match within the time I was willing to put into it.

I'm not sure I have even got to the stage of considering my "preferences" for these files if I were making a real print from them, because that wasn't the exercise, which was about whether I could match the more obvious quality advantages of the larger sensor files by careful processing the smaller sensor files.

I think a lot of it must come down to "naturalness" in feeling/perception simply due to reduced manipulation... although I can't explain it. But take color for example, when you view a flower the perceived saturation, gradation, texture/tonality is not the reality... if you look closer and pay critical attention you would see it quite differently. And I think that's probably what smaller pixels/sensors do... it's like more technical accuracy at the expense of aesthetics and "perceptual reality." The technically perfect image that says nothing and moves noone...

I have experienced this when I have spent a long time processing a file and realise I have used up all 16 layers in Capture One, or have a great long list of layers in Photoshop.

By this time I have much better idea of how I want the picture to look, and will save a copy, and start again. This second time around, I can usually get to the same place as before but with far fewer adjustments, and this time around the final picture just looks "cleaner" and more natural.

I certainly agree with you on finding digital colour difficult.

 

[sk66]

Do camera profiles in Adobe affect the demosaicing?

You can choose from multiple demosaicing algorithms in Darktable and Rawtherapee, but don't profiles just apply a tone curve(s) to the demosaiced RGB file.

As I understand it, pretty much anything affecting how you view a raw file is part of the demosaicing... the rendered colors interpreted from the raw data. E.g. if you apply color calibration (calibration tab in LR, color checker profile, etc) the program has to go back and reinterpret the image's RGB data values differently with the new ratios/mix. I.e. Profiles change what "zero" is for the sliders/edits. That's the advantage of working with a raw file... edits are not just a modification applied to "on top"; it's the raw data being reinterpreted which requires re-demosaicing (e.g. white balance can be changed at will). I.e. the preview image is demosaiced, but the actual/final demosaicing occurs at export. It's probably why you can't change the camera style/demosaicing (e.g. Camera Portrait or Adobe Landscape) for images that are not raw files.

But a lightroom profile is not changing the base level demosaicing algorithm (e.g bilinear, AHD, etc); that is more which neighboring pixels (how many, pattern, etc) are used in determining an individual pixel's color (along with different behaviors for edge detection, noise reduction, etc). But using LR's denoise/raw details option in the detail panel does change the base level demosaicing algorithm used; it is then AI based and not a singular algorithm (used to be the "enhance" function in earlier versions); that is also only available for raw files.

Thus import (preview demosaicing), enhance (denoise/raw details demosaicing), and export (final demosaicing) all take extra time.

 

[myotis]

As I understand it, pretty much anything affecting how you view a raw file is part of the demosaicing... the rendered colors interpreted from the raw data. E.g. if you apply color calibration (calibration tab in LR, color checker profile, etc) the program has to go back and reinterpret the image's RGB data values differently with the new ratios/mix. I.e. Profiles change what "zero" is for the sliders/edits. That's the advantage of working with a raw file... edits are not just a modification applied to "on top"; it's the raw data being reinterpreted which requires re-demosaicing (e.g. white balance can be changed at will). I.e. the preview image is demosaiced, but the actual/final demosaicing occurs at export. It's probably why you can't change the camera style/demosaicing (e.g. Camera Portrait or Adobe Landscape) for images that are not raw files.

But a lightroom profile is not changing the base level demosaicing algorithm (e.g bilinear, AHD, etc); that is more which neighboring pixels (how many, pattern, etc) are used in determining an individual pixel's color (along with different behaviors for edge detection, noise reduction, etc). But using LR's denoise/raw details option in the detail panel does change the base level demosaicing algorithm used; it is then AI based and not a singular algorithm (used to be the "enhance" function in earlier versions).

Thus import (preview demosaicing), enhance (denoise/raw details demosaicing), and export (final demosaicing) all take extra time.

I think we may be just thinking about the demosaicing pipeline differently.

I think of the demosaicing as ending with what you are describing as the baseline demosaicing (ie after interpolation). with adjustments thereafter acting on the interpolated RGB (linear) values, including the building of previews and exporting files. The Camera profiles are then used to tone map these linear values, which controls how the preview appears, and allows the adjustment sliders and curves to stay at zero.

On export, the raw processor then uses all the adjustments you have made, along with the camera profile to control how the JPEG is created from the already interpolated RGB values, and not from re-interpolating the raw file.

Denoisng can be different, and I know that color calibration works at the pixel level (rather than a hue level) but I have assumed this is still working on the file after the RGB interpolation, and not on the uninterpolated raw file.

But of course you have now cast doubt on my understanding

 

[sk66]

But of course you have now cast doubt on my understanding

And I can only give what is my understanding... what you are describing is basically pixel (bitmap) editing. What happens when I send the file to PS in order to edit it there, or editing a jpeg in LR. But as a raw editor LR is parametric and only fixes the pixel values on output (final demosaicing).

A raw file has no color space, only something we can call "rgb luminance values." And what those RGB values mean relative to each other depends highly on the intended output... the target color space, white balance, tone curve, etc, etc. For those things and others like edge detection, sharpening, clarity the best results are generated when you start from the raw file (there's a lot more going on than just basic color interpolation; I do not fully understand it all).

E.g. you can sharpen a jpeg/tiff/png/etc, reduce it's noise, change it's saturation/white balance/color space, etc, etc. But the results will not generally/potentially be as good compared to going back and starting at the raw data level. And that's why a parametric editor saves the instructions in a database, but makes no changes until output. And, as I understand it, at export Lightroom applies the edits starting from the raw file and in a sequence which is supposed to be optimal (no idea what that is). If it didn't, there wouldn't be any obvious reason to generate proprietary/limited use preview files instead of something more conventional/useful.

It is certainly true that what you see (the preview) has been demosaiced, and that in order to optimize speed the edits are (usually) applied as pixel based edits in the editing environment. Because of this, in order to get the most accurate rendition you need to be editing based on a 100% 1:1 preview. But you do not have to use/generate a 100% preview, you can even use smart previews (lossy compressed DNG), and your export options are not limited/hindered in any way... perhaps that is the biggest clue now that I think of it.

 

[myotis]

And I can only give what is my understanding... what you are describing is basically pixel (bitmap) editing. What happens when I send the file to PS in order to edit it there, or editing a jpeg in LR. But as a raw editor LR is parametric and only fixes the pixel values on output (final demosaicing).

No, I'm not.

As I tried to explain, rightly or wrongly, I think of demosaicing as describing "part" of the raw processing pipeline: linearisation, black level correction, white balance, scaling, actual demosaicing, (and colour space assignment, sort of).

And to go back to my original comment which started this, I think the adobe profile, or any profile is applied "after" all these steps are completed and don't affect how the raw data numbers with one luminance value (R,G OR B) per pixel is interpolated into three luminance values per pixel (R,G AND B).

The RGB values you actually view on screen (via the preview) are controlled by the adjustments you make on top of these interpolated values. Including switching between profiles.

And it's these adjustments that are saved as an instruction set in your raw processing software catalogue.

It's only when you export as a JPEG that the changes you have made in how you "view" the already demoasiced pixels are baked into a new file with new pixels, which are given values that match the ones you were viewing at the time of export.

This means, that any manual changes you made to the photograph can now be viewed (as a JPEG) without needing to refer back to the original interpolated values.

So I don't believe that switching between profiles means you re-run the demosaicing, it just changes how you view the already demosaiced values.

At least this is how I understand it :-(

A raw file has no color space, only something we can call "rgb luminance values." And what those RGB values mean relative to each other depends highly on the intended output... the target color space, white balance, tone curve, etc, etc. For those things and others like edge detection, sharpening, clarity the best results are generated when you start from the raw file (there's a lot more going on than just basic color interpolation; I do not fully understand it all).

E.g. you can sharpen a jpeg/tiff/png/etc, reduce it's noise, change it's saturation/white balance/color space, etc, etc. But the results will not generally/potentially be as good compared to going back and starting at the raw data level. And that's why a parametric editor saves the instructions in a database, but makes no changes until output. And, as I understand it, at export Lightroom applies the edits starting from the raw file and in a sequence which is supposed to be optimal (no idea what that is). If it didn't, there wouldn't be any obvious reason to generate proprietary/limited use preview files instead of something more conventional/useful.

It is certainly true that what you see (the preview) has been demosaiced, and that in order to optimize speed the edits are (usually) applied as pixel based edits in the editing environment. Because of this, in order to get the most accurate rendition you need to be editing based on a 100% 1:1 preview. But you do not have to use/generate a 100% preview, you can even use smart previews (lossy compressed DNG), and your export options are not limited/hindered in any way... perhaps that is the biggest clue now that I think of it.

 

[Terrywoodenpic]

The old adage that a good big one always beats a good small one once again proofs true.
However it is also true the ultimate quality is not the only criteria when choosing a camera

 

[myotis]

The old adage that a good big one always beats a good small one once again proofs true.

Unfortunately, I don't think it's as simple as that with digital :-(

I also think the other half of that idiom remains important "all things being equal"

One of my most useful takeaways from these tests, is that all three cameras, albeit with different levels of processing effort, could produce virtually indistinguishable 20 x16 prints.

And where tiny differences existed, they were probably down to my processing rather than the camera used.

However it is also true the ultimate quality is not the only criteria when choosing a camera

This, I suspect, always has been, and always will be, true.

 

[sk66]

Gone way off topic now... sorry.

And to go back to my original comment which started this, I think the adobe profile, or any profile is applied "after" all these steps are completed

Yes and no. While editing most adjustments are made without reference to the original raw data... like layers in photoshop. But many edits do depend on the original raw data. A few I know of (Lightroom specific).

•Changing the base level/demosaicing profile, which are the Adobe/Camera specific profiles. These require the raw data in order to get the color science correct during demosaicing, along with other "picture control" (Nikon) settings like sharpening/clarity/tone curve/etc (I don't know the specifics). And they are only available when working with a raw file. Other profiles which are "color only" are also available, and only those profiles are available for non-raw file types.

•White balance temp changes. When working with a raw file the temperature edits are in kelvin and are applied at the raw data/demosaicing level. If the raw data is not available the temp slider becomes a +/- temp filter applied "on top."

•"Enhance" functions (Denoise/Raw details now part of the detail panel). They are an AI based demosaicing algorithm rather than the default... it is also only available when the raw data is available;because it is applied at the raw data/demosaicing level.

•Smart preview editing. Using a compressed DNG for editing (2560px long edge). Because DNG is a type of raw file the raw-only edits are still available. And you are not limited to 2560px export if the original raw file is available. But if the raw file is unavailable, you are limited on output size (because you cannot reverse the demosaicing/compression that was applied).

There are probably some other edits which are applied differently when the raw data is available but which appear to be functionally the same.

A lightroom preview is not an image as such... it is just a set of instructions in a database to be applied to the raw data. Those instructions do generate the image view you are working with while within lightroom; but that is not necessarily the final demosaicing applied to the raw data when an image is exported. If you work with anything other than a 1:1 preview file the output results may differ from the previewed results (and maybe even if at 1:1).
If what you are working with is not a preview then it is a final demosaiced image, raw/demosaicing edits are not available, and the output should be WYSIWYG.

Enhance functions are the only edit I know of which pauses editing and re-demosaics the raw data at the time the database entry (edit) is made. I don't know if the other raw-only edits do or don't; and I don't how far back into the demosaicing process they have to go when they are applied. But because certain demosaicing steps must be done in sequence, and some other edits are better if applied in a specific sequence, the database entries (edits) are applied to the raw data "in order" when an image is created/exported (I don't know what that order is).

 

[myotis]

Gone way off topic now... sorry.

Yes and no. While editing most adjustments are made without reference to the original raw data... like layers in photoshop. But many edits do depend on the original raw data. A few I know of (Lightroom specific).

•Changing the base level/demosaicing profile, which are the Adobe/Camera specific profiles. These require the raw data in order to get the color science correct during demosaicing, along with other "picture control" (Nikon) settings like sharpening/clarity/tone curve/etc (I don't know the specifics). And they are only available when working with a raw file. Other profiles which are "color only" are also available, and only those profiles are available for non-raw file types.

•White balance temp changes. When working with a raw file the temperature edits are in kelvin and are applied at the raw data/demosaicing level. If the raw data is not available the temp slider becomes a +/- temp filter applied "on top."

•"Enhance" functions (Denoise/Raw details now part of the detail panel). They are an AI based demosaicing algorithm rather than the default... it is also only available when the raw data is available;because it is applied at the raw data/demosaicing level.

•Smart preview editing. Using a compressed DNG for editing (2560px long edge). Because DNG is a type of raw file the raw-only edits are still available. And you are not limited to 2560px export if the original raw file is available. But if the raw file is unavailable, you are limited on output size (because you cannot reverse the demosaicing/compression that was applied).

There are probably some other edits which are applied differently when the raw data is available but which appear to be functionally the same.

A lightroom preview is not an image as such... it is just a set of instructions in a database to be applied to the raw data. Those instructions do generate the image view you are working with while within lightroom; but that is not necessarily the final demosaicing applied to the raw data when an image is exported. If you work with anything other than a 1:1 preview file the output results may differ from the previewed results (and maybe even if at 1:1).
If what you are working with is not a preview then it is a final demosaiced image, raw edits are not available, and the output should be WYSIWYG.

Enhance functions are the only edit I know of which pauses editing and re-demosaics the raw data at the time the database entry (edit) is made. I don't know if the other raw only edits do or don't; and I don't how far back into the demosaicing process they have to go when they are applied. But because certain demosaicing steps must be done in sequence, and some other edits are better if applied in a specific sequence, the database entries (edits) are applied to the raw data "in order" when an image is created/exported (I don't know what that order is).

I agree with most of this, which hopefully was obvious from my other posts. There are a couple of things I'm not sure about, but I don't want to get diverted from the question on whether the camera profiles manipulate the demosaicing algorithm.

Everything I've read in the past, says they don't, so either we just agree to disagree or one of us has to find some definitive evidence.

Google has found some evidence that supports my understanding, but not from what I would call definitive sources.

Edit: I'm really struggling to find any definitive information, but here is how Julieanne Kost describes profiles:

PROFILES

A profile is a set of instructions that is used to render a photograph, converting it from raw camera information into the colors and tones that we see.

• Every raw image must have a profile applied.
• Profiles are nondestructive and can be changed at any time without any loss of quality.

Julieanne Kost's Blog | The Power of Profiles in Lightroom Classic

In the video below, you’ll discover how to work with profiles in Lightroom Classic to help establish a neutral starting point or apply creative effects to your images.

jkost.com

 

[sk66]

Everything I've read in the past, says they don't, so either we just agree to disagree or one of us has to find some definitive evidence.

I suppose it depends on at what point you consider "demosaicing" complete... but most authorities include color normalization (white balance, color calibration, color intent) as the final stage. And as I understand it, camera/adobe specific profiles affect the color intent/normalization. Many (most?) authorities say color space/tone curve is also part of the demosaicing process for display; I think that is quite arguably after the fact. But I don't know of any authority that has demosaicing as the (first) interpolation stage alone.

I can't think of any reason why an edit would require access to the raw data if it wasn't applied somewhere in the "raw data conversion chain"... and to the best of my understanding, that is "demosaicing." It makes no logical sense that a profile/edit would be disabled for a non-raw image if the demosaiced/displayed data is sufficient. But yeah... finding something definitive/authoritative/reliable seems to be extremely problematic (I've tried AI a few times, with different results each time). most everything seems to be simplified/generalized "for the web" (probably because it is quite complex in reality).

 

[myotis]

I suppose it depends on at what point you consider "demosaicing" complete... but most authorities include color normalization (white balance, color calibration, color intent) as the final stage. And as I understand it, camera/adobe specific profiles affect the color intent/normalization. Many (most?) authorities say color space/tone curve is also part of the demosaicing process for display; I think that is arguably after the fact. But I don't know of any authority that has demosaicing as the (first) interpolation stage alone.

If you go back many posts, I made the same point, when I said that I saw the "demosaicing" as a distinct step within the raw processing pipeline where the separate R,G and B luminance values for each pixel were interpolated into combined into R,G and B luminance values for each pixel, which were picked up by the next step in the processing pipeline.

So I thought we were simply arguing on whether the camera profiles affected the demosaicing algorithm (your argument) OR acted on the interpolated (raw) RGB luminance values produced by the demosaicing algorithm (my argument).

Which, to be frank, has left me a little puzzled about the details you have gone into in some of your post, as I thought I had explained this more than once

I can't think of any reason why an edit would require access to the raw data if it wasn't applied somewhere in the "raw data conversion chain"... and to the best of my understanding, that is "demosaicing." It makes no logical sense that a profile/edit would be disabled for a non-raw image if the demosaiced/displayed data is sufficient. But yeah... finding something definitive/authoritative/reliable seems to be extremely problematic (I've tried AI a few times, with different results each time). most everything seems to be simplified/generalized "for the web" (probably because it is quite complex in reality).

The edit needs access to the raw file, because I don't "think" anything is "saved" until after the tone curve has been applied, and a preview is created. But the preview only contains a subset of the raw data information, so any editing instructions are referring back to the raw interpolated RGB values, before re-rendering the preview after any user adjustments, and of course before exporting a JPEG.


EDIT Darktable history shows the steps before *normal user* editing as:

1. Raw black/white point
2.Demosaic
3. Input color profile (seems to convert the original device colour space into a working colour space)
4 output colour profile (seems to convert between colour spaces)
5. display encoding

** I say normal user editing as Darktable gives you a choice of algorithms and some control over the some demosaicing parameters

demosaic

Technical information description reconstruct full RGB pixels from a sensor color filter array reading. purpose mandatory. input linear, raw, scene-referred. processing linear, raw. output linear, RGB, scene-referred. Control how raw files are demosaiced and optionally apply capture sharpening...

docs.darktable.org

 

[TimHughes]

A fascinating read, thank you @myotis.

Just checking if my simple understanding is correct - before processing, GFX files look better than Nikon files. With some processing effort, Zf prints could be matched to the GFX more easily than Z8 prints could (despite the Z8's higher resolution). But we don't have a comparison of processed GFX vs processed Nikon.

If so, this matches my experience with GFX and various Canon bodies, and it's why I prefer GFX for studio and landscape where image quality matters most, and keep Canon for live events and fast action where capturing the moment takes priority.

 

[myotis]

A fascinating read, thank you @myotis.

Just checking if my simple understanding is correct - before processing, GFX files look better than Nikon files. With some processing effort, Zf prints could be matched to the GFX more easily than Z8 prints could (despite the Z8's higher resolution). But we don't have a comparison of processed GFX vs processed Nikon.

Yes, that is correct, it may well be that with careful processing the GFX will end up giving print quality much better than the Nikons. But for my purposes, that didn't really matter, because I had really answered the questions I wanted to answer.

I didn't want to get into the never-ending task of tweaking the GFX prints a bit, and then tweaking the Nikon prints a bit, and then re-tweaking the GFX files a bit, and so on, and so on and....

In the past, I've spent days and days doing this when trying to decide if C1, LR or DXO gave the "best" results, or which denoising software gave the "best" results, and this tiime, I decided to use prints from the GFX with default processing, as my standard. Especially as they were so obviously better straight out of the printer.

If so, this matches my experience with GFX and various Canon bodies, and it's why I prefer GFX for studio and landscape where image quality matters most, and keep Canon for live events and fast action where capturing the moment takes priority.

 

[juggler]

Interesting. I see a lot of people raving about GFX images but usually only when they've zoomed in to a single eyelash. And frankly, who cares? But this is a practical test.


I wonder - have you tried any form of blind randomisation? eg print comparable images, but then have someone else label them & shuffle them before you get to look? That would be illuminating.