I got curious about the maths some time ago, there’s a Wiki article here. I had never really thought about the maths behind them but looking at those tables you can see that each group differs in size from its neightbour by a factor of 2. There are 6 elements in a group so each element differs in size from its neighbour by a number called the ‘sixth root of 2’, or 1.12246204831.
A consequence of this is that if you change your sensor to one with double the pixels but the same physical size, 24MP to 48MP for example, then theoretically you will move up by precisely 3 elements. This is irrespective of the size of your sensor though clearly the larger MP counts are only there for full frame and medium format. It also means that for a sensor with a specified number of MP there is an absolute limit to the lp/mm that you can resolve.
It would be interesting to see if pixel-shift achieves a real increase in perceived resolution on Vlad’s Test Target. I suspect that it will but that it won’t be the large difference indicated by the quadrupling of pixels on the S1R.
I do have Vlad’s target as well as my expensive Thorlabs target. I have also read his article you linked from “fiom4ever-digital.blogspot.com.” I agree with his statements therein that a good chrome-on-glass target can read higher resolution than can his target, and I also agree that his target is a practical and more way to measure corner to corner alignment of a digitizing set-up. Of course first one has to get to an alignment, and for that either a Zig-Align attachment or a Versalab Parallel help one do a fine job of that. I have also found that the Thorlabs target can read higher resolution produced on my system than I achieved with his target, but both could demonstrate high resolution results, sufficient for many purposes and much cheaper in the case of his product. I should also clarify that the person named “Mark” he quotes at the top of his article is not me - it’s another “Mark”. There are quite a few of us “Marks” around!
SSelvidge - thanks for the additional explanations of what lies behind your comparisons; that helps interpret the outcomes.
You raise some interesting questions here which I agree deserve more research than at least I have seen published to date. Perhaps I should do some of this work and report on it, as my digitizing set-up does allow me to photograph a variety of media sizes using a wide range of magnification ratios, and stitching whether in Lr of Ps has become very good at handling such outputs. But it’s time-consuming work to both do and write-up, and I have a lot on my plate just now, so it will await another day; but my interest in pursuing the subject is “tweaked”, and for that I thank you.
This week I hope/want to shoot more images of my Vlads Target (using multiple lenses if I can muster the time) with the intention of making a Google Drive folder or WeTransfer to share the raw files results. This is a larger task as I would want to compare apertures while I do it - two birds situation.
And buying an alignment tool is something I should do but have been resisting!
Well, notwithstanding what I just said above, since I’m digitizing some colour negs this afternoon I decided I should give pixel-shift a run-through on a photo where the “grain” shows well. Having now done the comparison between a single-shot and a 4-exposure pixel shift image (creates a Sony “ARQ” file), I see absolutely no difference in outcomes between them and must attribute this to the likelihood that my capture set-up is squeezing all the detail that can possibly be squeezed from this film with a single exposure, so there simply is no value-added to be had from more data. I’m providing a screen grab of a portion of the photo at 100% magnification in Lr showing the ARW and ARQ files (single and pixel-shifted resp.), so you’ll see what I mean. At least as far as I’m concerned, I don’t think I’ll be spending more time on this.
Now a caveat - this is specific to my set-up and a sample of one, so it’s probably not telling a universal truth - only one for this case.
BTW, this was a handheld flash-supplemented shot under low lighting conditions made in one of the Mameluke tombs in the City of the Dead, Cairo, in the year 2000 using Fuji ASA 100 in a Nikon N70/w. kit lens.
Oh - and I forgot to add: converted in NLP, “Basic” Color Model, “Lab Standard” Tone Profile, HSL “Natural”, White Balance “Auto Neutral”; No adjustments made. The camera set-up: Sony a7rM4 on Novoflex bellows with Schneider APO Digitar Macro lens at its recommended maximum aperture of f/5.6, magnification ratio 1:1.
Thanks for that - that’s a very well-done article, even if it was not in the context of scanning (with a macro lens).
The article inspired me to do a 16x pixel shift of a recently taken 120 format Ektar 100 photo. The improvement is now very evident: Comparing the shifted raw file at 200% to the single capture raw at 400% (to bring in a similar screen view): Shadow detail is vastly superior; color is better; noise is improved, and grain/dye seems smoother.
Given that shadow detail is an area in which film shooting is more challenging than digital capture, the unmistakable improvement in the pixel-shifted file looks helpful. (The other single-shot deficits can be easily addressed to make perfectly fine photos: The color can be tinkered in LR’s and NLP’s color tools; grain/dye harshness can be smoothed by nudging LR’s Clarity down a bit; and noise can be improved using the LR’s Manual Noise Reduction.) Though NLP does, in fact, bring in some of the shadow detail in the single-shot file through boosting Darks and Blacks, it comes at a cost in noise. So, pixel shifting seems to offer some helpful benefits “out of the box” for the occasional high-value photo.
Thanks all for pursuing this discussion after my original post. I still have a lot to learn about NLP’s tools and my LR skills can also stand improvement. Nevertheless, I’ve learned a ton through this discussion!
So many elements for us all to consider but I trust yours are all top notch:
emulsion, storage, age & the taking lens on the film
digital capture lens, camera resolution, alignment + stability + focus.
As best I can tell we should all theoretically see:
4x PS - full RGB capture + color noise mitigation
8x PS - adds resolution gains + file size
16x PS - what happens here? more resolution gains + file size?
and so on for 32x or more
at some point you reach the end of the lens/sensor/real world conditions (but this is physics I don’t have a grasp on).
I do wonder why you didn’t see any color noise improvement, as expected, in your quick 4x PS test. Maybe the dye cloud/grain hides it so it doesn’t matter? Or the automatic color noise reduction is dialed in differently for the Sony A7Riv compared to my experience with Canon R5 or Panasonic S1R?
This has me wondering how the 4x/8x/16x/etc PS shots, software, etc differ from brand to brand. For example, in a previous link above, the Nikon Z8 16x Pixel Shift is a similar sized file as the S1R’s 8x PS, both more than 180 MP which seems odd.
@LABlueBike I am glad it was useful. I’d love to see your examples when you get a moment. More information is always great.
These links/site may be relevant to this discussion. The first is on the effect of Pixel Shift mode for the S1R (but there are others) and the second is the general homepage with lots of charts and in-depth explanations of the analysis elsewhere on the site… a lot of which is more engineering and science than photography.
Relevant info in screenshots below but you can make your own with many camera models - there is a A7Riii Pixel Shift specific page as well but nothing newer.
Below: Note the Pixel Shift (HRM in the image) peaks much higher than the regular capture compared to nearly all cameras and compares favorably to a top of the line Phase One. Unless each brand does it entirely differently, this is likely to hold true for them as well and is shown on the Sony, Olympus, etc pages.
Recall that the capture device is one of the finest sensors on the market. At ISO 100 there is no visible colour noise, whether making the usual digital photographs or photographing a piece of film.
There can be no more gain of resolution than the film itself holds. When the dye clouds are sharply defined from corner to corner and the pattern of them remains unchanged with vs without it doesn’t get any better because it can’t regardless of how much pixel shifting one does. But yes, file size grows exponentially. That ARQ I produced yesterday was almost half a gigabyte because it was produced from a 61 MP sensor and uncompressed. This is far beyond anything one needs for making even very big high quality enlargements.
My results are of course specific to the equipment and software I used.
The most important guidance I get about what I’m dealing with is from looking at the results. Sure, on a theoretical basis one could go on experimenting, but honestly at least in my case I don’t think it would be a productive use of time.
I didn’t mistake the ‘Mark’ in Vlad’s post with you in fact and I’m glad that his question prompted Vlad to make this post as I think it is a very fair appraisal of his Test Slide. I can see that when you are able to resolve 128/144 lp/mm with your 80mm Apo-Digitar and the Sony 61MP sensor then you are slightly beyond what is possible to resolve with the VTT as Vlad has demonstrated, though I’d still find it very useful for evaluating the corners (they must surely not reach those particular heights?). The Siemens Stars are also a clear way to indicate if there are any problems with alignment, but I guess you don’t have those anyway with your optical and laser methods.
In this community of ‘camera scanners’ it has also proved itself as an invaluable way to see what is possible from different setups, and to confirm that users are getting the most from their system, based mainly on the pixel dimensions of their sensor.
Very interesting to see your 4-shot pixel-shift test on your Sony. I suppose it might be possible that the 16-shot mode could compare in quality with stitching when copying medium format as from your article that is sometimes your practice. It’s a disadvantage that you need to take the separate exposures into Sony’s own software though, clearly it adds a bit more to the workflow compared to Panasonic’s solution. I appreciate that you don’t have the time to test this though.
@Mark_Segal, First, I am not arguing that your results or intuition is wrong! And I don’t need to take up more of your time if you’re otherwise busy
Second, I am quite surprised the A7Riv does not show any color noise at base ISO. Is that is a well-known feature? I will search some. This is worth exploration. And I need to make a publicly accessible folder of my RAW camera scans for people to look for themselves.
My R5 (in single capture) surely shows base ISO color noise upon very close inspection on both slides and B&W - which I can solve increasing the color noise reduction slider to around +35 in adobe camera raw/LR. The S1R does, too, as I have shown. It uses the same sensor as the Leica Q2 and SL2. Leica and Panasonic are tech partners, and all 3 use the same Sony based sensor.
I can’t speak for other cameras. Perhaps this means I need to ETTR more or I am making a technical error? Or have I mistaken ETTR on negative film for years and I need to take into account the inversion?! Wouldn’t that be something!
Your large file size surprises me, too. My 8x PS raw images are made in-camera in a few seconds and end up around 335mb at 16,736x11168px. It takes my machine longer to open one in camera raw than it does to complete the image from start to finish.
But… this is surely why I’m now wondering what effect each brand’s implementation has on the utility of actually using it or perfecting utilizing the technique/technology. And now I’m reminded of @VladS’s thoughts in another post as he mentioned RawDigger re: what is Panasonic or Sony or Nikon doing and not doing to their raw images under the hood, so to speak.
This is all so fascinating to me. I hope it is to others as well
Here are the two Ektar 100 files scanned with the Z8/105 macro, ISO 64. Top photo is the single shot; bottom photo with pixel shift x16. Both have the same temp (2700) and tint (-20) in LrC, but the single shot has a weird blue-green tint to the overcast sky. Oddly, the skin tone is more pleasing out of the box with the single shot. (But when shooting people with Ektar 100, adjusting skin tones is usually a given, right?) Importantly, both files have the “Dark” and “Black” boosted in NLP by 15 in each, but you’ll see in the dark trees on the right that shadow detail is much better in the shifted photo. This even though my wife’s black jacket responded slightly better to the NLP adjustments in the single shot than in the shifted photo. I’m sure the jacket can be improved in further adjustments; the dark tree, not so much.
@SSelvidge and @Harry and @Mark_Segal . Thank you all for having this in-depth discussion. Those are exactly issues we all dealing with and wondering about. Also I really understand how much of you valuable time you spent writing these posts down. I really mean it. I very much wanted to explore this topic myself but lack ,organization and proper camera. Now here is what I want to add to discussion and hear what you think about it.
So my hope was that pixel shift can solve the following conundrum. The size of dye clouds or silver crystals in the processed emulsion is 3-10 micron per literature. The pixel pitch is around 2-5 microns depending of specific sensor resolution. So my concern is that with good lens we are throwing Bayer array based demosaicking algorithm out of the water. The typical algorithm does not expect such high spacial frequency image as we have when we scan film. the typical guessing in demosaicking is based on assumption that nearby pixel have similar colors or that any change in color takes few pixels involved. We present completely different image. Now, I started playing with different demosaik algos in Darktable - where you are able to choose the algo. I wanted to write article about what I found , but never had time to finish it. The bottom line was that of course algo which used minimum number of pixels showed somewhat better details for scanned images. Only one algo would preserve the speckle in the image if speckle is one pixel in size. Not that I want to preserve one pixel, I want that pixel not affect pixels nearby. Don’t take my word for it - just run and see yourself if you can observe the effect.
Now let’s pivot to pixel shift. My thought was that with pixel shift and proper demosaicking we should no longer see adversarial effect of nearby pixels. It’s like we have Faveon sensor emulated. So each pixel speaks for itself only. My feeling was that this sort of processing should spare us from the perceived noise in image - where noise is not random noise of electrons wandering around, but rather demosaicking running amock as the image is not what algo is intended to process. Any thoughts on that would be appreciated.
The second thought I was having is about inversion process of grainy images. It would be great if @nate could chime in here. So here is my thought experiment. As far as I understand the inversion process is based on certain actions taken based on the shape and location of histogram for each channel. One way or another the inversion process most likely involves normalizing the histogram for each channel and applying certain curves to make them rich cinematic or lab standard. Now the point I am trying to make is that grainy structure, color noise can deform histogram as colors distribution varies widely in grainy image vs smooth image. So I even tried to intentionally blur the image to rid of noise and average the colors across maybe 10 pixels. Then I would invert the image, save the LR settings and apply them again against not blurred image. The idea would be that inverted image would be “better”.I tried that with literally couple of random shots and cannot see much of the difference . But that means nothing , as I really did not try to do this in properly set study. I am throwing this idea out - if anyone can check it out or just argue why this should not work - I am all ears
On the side, let me note that I am very glad that my targets help you in some way and I contributed something to the community. Thanks!
I think the principal of ETTR remains correct provided one takes care to not clip because it maximizes the potential signal the photosites can receive. But just let us be aware the “brightest” data points will end-up in the shadows, not the highlights, once the negative image is inversed in NLP. If anything that should help to minimize digital noise in the shadows where the problem is usually most apparent. So an advantage.
The large file size is not a surprise because it is blending 4 photos that are each 123 MB raw - from a Sony a7r4. This camera produces very large uncompressed raw files.
On colour noise, I repeat - at low ISO I see NONE from the a7r4. I did see A HEAP of it on some photos I made in the Volklingen Hutte last year where conditions were like in a darkened cinema and the ISO was set to 6400. Perfect for eliciting and seeing colour noise in spades, but it takes those kind of extremes with this camera. Will never happen digitizing film in a studio set-up at ISO 100.
I think the single shot is an all-round better photograph in respect of tonality and colour rendition. The sky may be a bit cyan but that is so easily corrected with a sky mask and a slight leftward shift of the tint slider in the colour balance panel of LR. The shrubbery in the background isn’t sharp enough for the very slight differences of detail to matter - and those differences are due to contrast, not resolution. The shifted image is on the whole more contrasty than the single shot, and I think not to its benefit, but that is a matter of taste.
Vlad, it’s possible to over-think all this. Focusing on colour negs, the developed image is made-up of its dye clouds, as we know. Those dye clouds are of irregular size and shape and random placement, contrary to pixels on a sensor which are of uniform size and shape and arrayed in a grid. The dye clouds can be anything from 10 to 30 microns according to Tim Vitale’s research, while a photosite on my Sony a7r4 is 3.76 microns. So on the whole, it will take several pixels to photograph a dye cloud, and they will not coincide. Hence for accurate enough rendition of the dye clouds one depends on the quality of the demosaicing of the values in the photosites. The Bayer matrix affects colour mixing, but each photosite still retains its own luminance value. The signal processing that goes on in the sensor and the demosaicing that happens afterwords in processing software are complex to the extent that those who aren’t specialized in the physics and the math used in these processes, including me, can’t say exactly how all this comes together, so I would prefer not to waste time trying. And of course there is also designers’ judgment that goes into all this, so it should be expected to vary from one camera and software to another.
The histogram measures per channel luminance, showing the distribution of pixels across the luminance range. How much this distribution is affected by sensor noise I have no idea and can’t find out unless I expose those negatives at ISO levels I would never use, so at least for me, not something to spend time on.
Remember that in a colour negative the dye clouds ARE the image, so blur them and you blur the photo. Given that the apparent “grain” in a print at least up to 13x19 inches will appear to be much less than what you see in a highly magnified display image, it’s better to leave the “grain” alone and not blur the photo in all those areas where detail matters. I don’t mind blurring the sky and human skin to lessen the appearance of the dye clouds in those particular areas, but for all the rest I’m more concerned to preserve detail. Some of this is a matter of user taste.
So setting aside theoretical considerations that I’m neither trained nor experienced to understand, I look at results of various capture and processing approaches and see what I think delivers the highest quality result, as defined by the parameters I choose for assessing image quality. So far, as I’ve mentioned above, I don’t see that pixel-shifting contributes anything to enhance quality, but that was a sample of one. If I were to go back to it with other kinds of images and see differently, I’ll be back here to advise, but it won’t be soon.
Hi Mark - Thanks for that. Since posting, I tried “Linear Flat” and “Linear Deep” in the NLP Tone Profiles and saw instant improvement all across the shifted photo, especially the skin tone.
But yes, we should all appreciate that pixel shifting is not necessary to get wonderful scans out of our quality camera/lenses using NLP’s adjustments.
I’m still learning. Indeed, just in the last hour I learned why most of my LrC adjustments are “backward” and why it is better to make such adjustments in NLP. Coming off of my previous use of a Nikon 9000 (Rest in Peace) and VueScan TIFFs, I’m enjoying NLP’s adjustments immensely.
I have made a G-Drive link for anyone to analyze. And did DNG file size tests (info below) but have not analyzed the data. Not quite sure what I would look for beyond the obvious.
It includes the raw files, lossy compressed and uncompressed DNGs made from ACR within Bridge (newest update), low res lab scan reference, and jpegs made from my raws. I also included a quick readme with some extra info and the very basic linear color profile I made using DNG Profile Editor. I am open to suggestions on how I might do any step in this whole process better. @VladS@Digitizer@Harry@LABlueBike@Mark_Segal
If anyone wants to add their single capture/pixel shift capture files to the pile, let me know and I will send you a private link to do so. I think some color neg and slide films would be great to peek at, especially from various brands. No need for DNGs at the moment.
FILE SIZES - For my example image used in this thread the sizes are as follows for Single capture
RAW - standard capture - 66mb (not shared, likely uses slight compression)
RAW - single image from within PShift - 85mb
DNG - lossy compressed - 13mb
DNG - uncompressed - 65mb
8x Pixel Shift
RAW - in-camera - 339mb
DNG - lossy compressed - 35mb
DNG - uncompressed - 242mb
As long as the uncompressed data is not modified somehow, thats a nearly 30% file size decrease. If the lossy compressed ends up being fine… I would be surprised but happy haha
Your reasoning about the effects of the Bayer filter brought to mind this old (2010) example from Tim Parkin of On Landscape, “Where have the berries gone”. Details from the same image, one from 5"x4" film and the other from a Canon 5DMkII. I still use my 5D MkII from time to time, what a camera!
I ran the files through DxO PhotoLab with DeepPRIME noise reduction and compared exported 16bit TIFFs inLightroom Classic. I saw slightly more detailed results with pixel shifted captures, but I’d say that the advantage is not worth the bother - unless you really want to pixel peep and/or print 50in high. With 20in vertical size prints, I got this at 300%:
