Pixel Shift vs Regular Capture - quick comparison w S1R

[SSelvidge]

Thought this should be its own thread.

EDIT later: Post 2 and especially post 15 in this thread have more thorough personal analysis from me and serious explanations from another author, respectively.

Digital camera: Panasonic Lumix S1R - f6.9, 1/60th, ISO 100, Sigma 70mm ART.

Holders and Stand: Film in a Valoi 360 with their CS_Lite and light enhancing sheets. All on a Besseler 23C XL III enlarger turned into a copy stand. Took 4-5 tries for optimal focus and alignment (I was trying to be particular and kept bumping things). Best image pair used below.

Film Camera: water-resistant Nikon Action Touch (same camera as L35AW) with a 35mm f2.8 lens, automatic focus, automatic exposure. I selected this image because it was on hand, relatively sharp, and a decent photo that had good grain with limited whites and blacks.

Film: TMax 400, lab developed and scanned. Taken midday under cloudy conditions, early summer 2020.

Pixelshift mode set to also record a single, standard raw file on a 4 second delay trigger from the Lumix Tether software so I didn’t introduce shake. Mechanical shutter, which required me to touch the camera, also introduced shutter shock and so I did not include it. Panasonic outputs same 14 bit file in both E and M shutter modes. Interesting note, a single capture e-shutter image was 25% smaller than the E-Shutter single image + Pixel Shift combo. (~66mb vs 88mb). I used the larger E-shutter image instead.

I made a basic custom linear color profile (gamma = 1… I think?) and inverted using curves with some clipping applied to get it relatively pleasing and did screenshots for speed here. No sharpening, calibration, or anything else beyond white balance and curve was done.

Chose B&W film to show grain easily and color noise in single capture vs none in pixel shift. Did NOT desaturate or switch to Grey Gamma/B&W mode.

Base image comparison, Pixel shift (L) vs E-Shutter single (R):

Left Edge 300/600% crop to see grain / pixels / color noise comparison:

Near Center 400/800% crop to see grain / pixels / color noise comparison:

Look above for yourselves at how the grain, details, color noise and histograms relatively change (or don’t enough for some people).

Now for lab scan comparisons, which I think are of limited utility because so many variables and choices are made by me and by the lab. But some may want to know.

Pixel Shift vs the Lab Scan from back then (not edited to match, but scaled down to match lab scan size). I think I could very easily get them to match.

Pixel Shift Vs Lab scan edited similarly but not exactly. I could keep punching it up but I want to go make dinner!

EDIT added 2 days later: more images below in this thread Pixel Shift vs Regular Capture - quick comparison w S1R - #8 by SSelvidge

Cheers y’all

[SSelvidge]

For the full context of why I did this, below is a linked comment I made yesterday that was the genesis of this post, along with @LABlueBike @VladS and @Harry -

FF Pixel Shift versus MF Scan (Preliminary/Non-Macro)

I have been testing pixel shift vs single capture for a few weeks now on a Panasonic Lumix S1R. Before that, I read about it constantly for a few months off and on trying to understand the value - perceived vs real. I have not had the opportunity to test FF pixel shift vs digital medium formats 44x33mm (Hasselblad/Fuji’s “affordable” MF) or 54x40.5mm (Phase One). They surely do amazing things with single image, 16bit capture. And some of them have pixel shift abilities as well. I am sure you N…

@SSelvidge

“Briefly” on various topics

1. Sharpness: the images are not sharper, per se, in the grand scheme at real viewing sizes and distances. They are as sharp as the original film and lens made them. However, there is more data captured at the grain level and it has much smoother, finer transitions within colors and edges when you go in to 400% or more. A win, IMHO, but not for the reasons most would expect. The images at extreme magnifications appear to be softer. While at normal viewing distances only a little better or not at all. But in the lower resolution single capture images it is simply the effect of hitting the pixel level of the image, creating artificially sharp edges once zoomed in far enough. I find the grain more pleasing in the high res pixel shift images because it’s smoother, more like it would be in a print because light and enlargers dont use pixels. Said another way, more and smaller pixels means those edges have more detail and thus cleaner transitions at typical and 100-200% viewing distances compared to the lower res image. Great for a 60inch wide print or a 50% crop, but of limited or no utility at all for an 8x10inch print or Instagram.
2. Resolution: related to more pixels mentioned already… a clear win on the condition that you want to print it large at 300dpi and/or, especially, crop extensively. Otherwise you just have huge files. Final viewing size and viewing distance is what matters. So subjectively useful depending on your needs.

From the last post in summary (but I had a LOT of detail, too, haha!)

@SSelvidge

TLDR of the longer thoughts below: Pixel shift has great benefits if you can utilize them. But the film emulsion cannot give what it does not have as far as detail goes… so for viewing on nearly any screen and for prints smaller than, say, 16-20" on the long edge I think there is limited usefulness of the technology. At some point, you just “scan” deeper into the grain or dye clouds. Also, alignment and vibration becomes a huge issue compared to single capture.

For me, its the perfect and affordable tool to make higher quality scans of selected images for public presentation and long term preservation. The effort it took to get here was considerably harder than I expected and I am not done trying to perfect it.

For most people I don’t know that pixel shift has strong benefits unless they are printing physical books, cropping extensively, producing large exhibitions or involved in scientific and cultural heritage applications or library and historical archives purposes.

[LABlueBike]

SSelvidge - Thanks for this. It seems to support my own thought in that the benefit is in the size of the file: All other things being equal, one could make a large print with less “magnification” using the pixel-shifted file. As a layperson, I would think such a print would be appear sharper than the single-shot image zoomed to the same size. (AI upscaling might do the same thing or even exceed pixel-shifting, but by then I think you’d basically lose the whole aesthetic of shooting film.)

BTW, that is a nice lab scan; It’s just a matter of contrast and other fine-tuning to get your own scan to match it. I’m sure you’ll get to it after doing your 10,000 other scans.

[SSelvidge]

Briefly… yea.

I think the pixel count is great because it fully resolves the grain and does so with far less jagged edges … because of the pixel count! It’s definitely better but how much better for your average user is the question!

Where I think there is a huge benefit is 2 parts:
1- the full RGB of every pixel while removing color noise and gaining more accurate color! I didn’t highlight that as much as maybe I should have. I notice color noise, even at base ISO, in every camera scanned image. Moreso on deep shadows in slides which I suspect is because of the huge dynamic range slides cover.
2- post editing is much nicer on the pixel shift. Sharpening and color changes are smoother and more accurate.

I have had terrible luck with AI upscalong on film scans, personally. I don’t think it know what to do with grain rather than noise. Always make the grain go crazy in my experience. Maybe it’s better since I tried last year.

It is definitely a great lab scan (shout out to Holland Photo in Austin) but they definitely clip the shadows, let it have some warm tone, and punch it up a lot.

When doing my own work I’ll definitely creat some presets and only get specific on my selects. NLP is great for this.

When working on client images, generally I’m getting them a great base to work off for culling, fine tuning etc themselves.

[Harry]

It’s the first time I 've seen this type of pixel-shift comparison so thanks for posting it here. You’re really exploring the outer limiits of what is possible to extract from 35mm.

For me the ‘Left edge’ comparison shows the difference the best. The downsized pixel-shift has tighter less intrusive grain, to my eye at least. It’s a shame in a way that they didn’t also offer a 4-shot mode to create an image the same size, that would have been interesting also.

Of course 8256 x 5504 pixels is very good anyway, that’s 27.5" wide at 300 dpi, or around 23" at 360 dpi,the optimum resolution for large format Epson printers. It would be interesting to actually print a portion from each file to see if the differences that are visible here actually make it on to the print. In fact I’m not actually sure what goes in to making these zooms in Bridge, Lightroom or Photoshop, or other programs, it is interpolation to an extent I suppose.

Just working on the pixel dimensions you can relate the single shot resolution to a scanner set to 5900 ppi, and the pixel-shift to 11,800 ppi. To put that in context my old Flextight Precision II gives either 6300 ppi or 5000 ppi for 35mm depending on whether the negative is presented with the narrow edge (6300 ppi) or the long edge (5000 ppi). I always tended to use the 5000 ppi because it gave everything I needed and it was 2/3 faster (they are very slow scanners).

Your 70mm Sigma ART must be fantastic to not show up any problems at this kind of resolution, perhaps with how the grain is resolved in the corners, really impressive.

[SSelvidge]

Would me finding a place to load the raw files be useful? Any suggestions as to where because Dropbox has a bandwidth limit now. G Drive?

The Sigma 70mm is fantastic and far better than my Canon EF 100L, but does get weaker at the corners as expected. Still need to run comparisons of apertures. I know most tests put it sharpest at f4 but I do wonder if that has enough DOF in the corners bc it’s not perfectly flat field? F4 may be sharpest but what is optimal balance for the corners? Again, something I need to try.

I also think my NiSi macro rail slips in the vertical position I keep it in, even when supposedly locked down. Not quickly, but over an hour or so it moves ever so slightly.

Anyone have macro rail suggestions that lock great in vertical orientation?

Sometime soon I’ll have my Nikon Scanner 8000 on the rack for testing. The f2.8 only demands absolute perfection so I didn’t use it on this quick test. It’s aperture is why I have dreams of getting a Printing Nikkor or a Pyrite lens because I’m always chasing the dragon on getting better at this.

[Harry]

People do use Google Drive but I was just thinking aloud about how well ‘Zoom’ views correlate with reality and very large prints.

As far as vertical macro rails go then there is the Velmex Unislide, see this post by Graham:

Let's see your DSLR film scanning setup!

Thank you @Graham I had thought of a linear stage for industrial applications but tried with the gear I already have. Not sure how they handle load in a vertical setup. I usually do that too, but the canon 100mm macro, even though there is a tripod collar for it, has a polycarbon housing and I’m a bit reluctant to tighten the collar too much or glue the adapter ring to the lens body, How do you mount your lens? The overall goal in building this rig was to use basic materials that could be mo…

Robert O’Toole liked them, I think you have to work out a way to mount them but I don’t think it’s difficult. I think they appear used in the US, you never see them here in the UK.

https://www.closeuphotography.com/velmex-unislide

There’s also the Novoflex Castel M, it’s very sophisticated for macro photography (and expensive) but does more than is required for this application.

I was lucky to get a Newport 431 combined with a Starrett micrometer. These engineering sliding stages are beautifully made, no play, bearings etc.

[SSelvidge]

My take: On high quality (B&W) film, somewhere between 40+MP and 170+MP you fully resolve the grain with a great lens. I cannot say where though, exactly. And I think the “crunchy” feeling of camera scans that you often see is because of how the grains are not fully resolved… so this is mitigated by combinations of increasing resolution, light source type, and full RGB capture - whether done with a traditional scanner, a camera or otherwise. I imagine that wet mounting and scanning would yield better results in all formats.

Of course, this is all based on what I can see with my own eyes but should still be seen as somewhat anecdotal rather than fully scientific.

---

Extra images for examination - top left corner, near center, and extreme center zoom, and “standard” lab scan comparisons (8MP from a Noritsu). Pixel Shift is on left, single capture or lab scan on right… just as before.

I edited the images to be more pleasing to me AND switched all to Gray Gamma to remove the visible color noise. And did the same to the original lab scan, too. The “standard” 3600px lab scan comparisons are quite striking at the bottom.

Extreme Top Left Corner (made relative at 100/200%):

Near Center (made relative at 100/200%):

t

Near Center (made relative at 200/400%):

Dead center extreme zoom (made relative at 800/1600%):

Dead center extreme zoom of SINGLE SHOT vs LAB SCAN:

Dead center extreme zoom of PIXEL SHIFT vs LAB SCAN:

Pixel Shift downscaled to match vs Lab Scan (both 3637 on the long edge, but slightly different bc lab scan crops much more)

[Digitizer]

@SSelvidge

The “standard” 3600px lab scan comparisons are quite striking at the bottom.

I suppose that you mean 3600 ppi … which is roughly what a 24 Mpixel camera can provide.

If we’re thinking of “line pair” resolution, we’d have to double that, which will require a 100 Mpixel sensor for one shot scans…which will loose some quality due to de-bayering though.

Okay, let’s get a 100+ Megapixel monochrome sensor and scan with a re ally good lens and red, green and blue light, combine these captures to get the ultimate quality. Suffice to say that te ordinary copy stand setup will not be good enough. Maybe this PhaseOne setup will do?

[Graham]

@Digitizer Phase One setup looks good. Column looks a bit weak though, compared to yours and mine?

[Harry]

@Digitizer

I suppose that you mean 3600 ppi … which is roughly what a 24 Mpixel camera can provide.

No, I think it is is actually 3637 pixels on the long side, so only around 9MP. I have been taking great interest in these tests because I haven’t seen the results from pixel shift before, and this S1R gives a 187MP file in-camera from 8 shots.The copy stand is in fact very solid, a converted Beseler 23c enlarger.

[Harry]

Very interesting to see these, I see that DPReview report that the pixel-shift image has less sharpening than normal compared to other competitors so they have used their own custom setting of radius of 0.6 at 200%.

https://www.dpreview.com/reviews/panasonic-lumix-dc-s1r-review/7

The comparison with the lab scan is a bit extreme (!) if I’m correct in thinking that it is ony a 9MP scan. I read that Noritsus in general were capable of 24MP generally and 30MP from the HS-1800.

My first off is that I used a Noritsu LS-600 and HS-1800 for testing

[SSelvidge]

The lab scans are roughly 3600px on the long edge. Yes, the Noritsu is capable of more but I knew I would eventually scan them myself at home. Also the prices for the higher res scan were, to me, outrageous. A roll of film would end up costing $45+tax (scans without prints) in 2020. No thank you! Ha

As for the monochrome sensor and capture of RGB (+IR if we are getting fancy) is something that people are trying and doing. @mightimatti might have something to say on this subject. I just know its not an easy process!

The Phase one setup is surely incredible but $$$$. I am quite impressed with my stripped down Beseler 23C XL! I will do a post on it, eventually. I think it would do just fine for that use case honestly.

[SSelvidge]

Yes, the comparison was extreme! Almost to the point of being useless haha!

I just wanted to show people what a basic (ie inexpensive) lab scan on a Noritsu might come back to them looking like in comparison to camera scanning. The cost payback even with a 24MP is VERY fast considering what higher end lab scans cost at this point.

I checked out the DPReview link (fantastic) and now I have some sharpening tricks to play with. Also its great to note that DPReview often includes Pixel Shift images in the image comparison pages and they show exactly what I would now expect them to vs single capture scans. Great stuff!

[SSelvidge]

@LABlueBike @VladS @Digitizer @Harry et al, Just came across this link from a post in August on this forum that goes more into the value of pixel shift with charts, some math and better explanations than I managed! But came to many of the same conclusions I did. Summed up below.

Pixel Shift Shooting Analysis of the Nikon Z8

and more:

Nikon Z8 Pixel Shift Shooting to Reduce Aliasing

The OP’s post, @Dean.G Nikon Z8 "NEFX" - Pixel Shift & NLP

Important TLDR Summary from the Ed Dozier article about the Z8:

• " The pixel-shift feature, taking 8-shots [makes 45MP - ***see my added footnote!] and combining them into a single shot, resulted in a 26.2 percent increase in resolution. While this may seem underwhelming, it is in fact quite good."
• “The smoothing of color noise may be a bigger factor than resolution improvements. The Nikon Bayer sensor is also called ‘RGGB’, referring to neighboring pixel color sequences. The pixel-shifting operation changes this into something more like Sigma’s Foveon sensor, that stacks all of the color information under a single pixel. The 8-shot and 32-shot sequences add more color-noise smoothing, compared to the 4-shot and 16-shot sequences.”
• " the resolution change using 16 combined shots [makes 180MP - *again, footnote!] was about a 90 percent increase! "

***Interestingly, the Nikon Z8 implementation outputs a standard 45MP file with 8-shots while the S1R does 187MP. The Z8 requires 16-shots to output a similar 180MP sized file. This means the 16-shot gaining 90% on the Z8 could be similar to the S1R 8-shot implementation… I have no way of testing this myself or exploring how the RGGB and color noise affects may be similar or different but it would be worth exploring.

I think I inadvertently stumbled into a crazy value on the S1R - doing all this IN CAMERA is crazy useful and, importantly, very fast. Got it “used” with 125 clicks for way under standard resale prices. I feel very lucky.

[Mark_Segal]

Full disclosure, I have not tried pixel shift with my Sony a7rM4 yet, for a reason I’ll explain in a moment. Firstly, a comment on the examples provided in this thread: very interesting work, but how does one interpret what it means? I see more contrast in the pixel shift samples, but does that necessarily mean the result actually contains finer visible detail than the single shot comparator? And what of colour film which has a different end-product chemistry from B&W film - the former being “dye-clouds” and the latter silver crystals.

My film digitizing set-up resolves in the range of 140~160 lp/mm and when I am in focus, those dye clouds look crisply defined at 11x in-camera magnification seen through a 2.5x loupe. It isn’t clear to me that producing 2 to 4 times the amount of data a pixel shift image requires will result in a commensurate, visible improvement of resolved dye-clouds in the end results, say in a 13x19" or even 17x22" print. Maybe it’s one of those things I just need to spend some time testing, regardless of my a priori concern that it may be a waste of time, especially considering that the sharpness of the composite subject matter may be less than ideal for a variety of other unrelated reasons. It’s the whole system and the way it’s used that determines ultimate photographic sharpness and resolution of detail, not only the chemical structure of the film or the pixel dimensions on a sensor.

[Digitizer]

@SSelvidge

As for the monochrome sensor and capture of RGB (+IR if we are getting fancy) is something that people are trying and doing. @mightimatti might have something to say on this subject. I just know its not an easy process!

More steps indeed, but otherwise fairly straightforward. I’ve posted some info about a proof of concept I ran a while ago. I’ll add a link to said pist asaifi.

[SSelvidge]

Mark, I’ve read a lot of what you write here and elsewhere, so first, thanks for all the effort! It is always very educational. What are the basics of your setup to achieve 140-160 lp/mm? I haven’t fully tested it yet but 47MP/187MP on my Sigma 70mm ART and Nikon CS8000 lens (100mm 2.8 ish) likely have some promise of doing well. Preliminary 70mm quick shot showed me around 87lp/mm on an older Vlads 35mm target. Have not tried PShift yet. Need to!

RE Contrast: I do want this to be educational and useful for folks. Also, I am not sure I’m seeing what you’re seeing! The histograms are the same from 47vs187mp. When compared to the lab scans in the first post, yes there is much more contrast because I was editing them to be more pleasing/similar vs the lab scan. The unedited raw files themselves are as flat as you’d want them to be but I did not present them in their flattest linear form. In the second image example post I had edited the images to be as I would want them printed. Would the raw linear be preferable or more useful?

RE Color Negative: I have also wondered about the differences but I wanted to start with clear, easily seen results on B&W. Color or slides are next. While there may or may not be other benefits I do think for truly large printing, this has great utility. I am looking forward to exploring it more.

Generally speaking my take is that this process is of very limited utility for 35mm printed under 25". That said I don’t think it is less sharp, it just appears that way with the artificial sharpness of the pixel-level contrast. Then once scaled back down from 187MP to similar size as the 47MP file, they are pretty indistinguishable and at regular viewing distances for either 47/187MP they appear the same (as would be expected). For 35mm it is something I only plan to employ for images intended for exhibition and books to be certain, but the workflow is so easy because it is in-camera that it is hardly a bother once it’s all aligned perfectly (I am more exacting than I probably need to be).

I think this is honestly more useful for capturing 120 and larger formats with or without stitching. It will see more use this way especially if the camera allows me to capture raws at ratios of 1:1, 6:7, or 8:10, which I have not tried yet.

I also plan to test the Adobe DNG conversions which seem to create smaller files sizes, as long as they don’t mess with the data appreciably.

I know I write a lot here, sorry!, but this subject hasn’t been explored as much for camera scanning as I would have expected by now. I hope to create debate and discussion so we can all learn and get better skills / scans!

[Harry]

I’m not intending in any way to reply for Mark but he did helpfully go into his method and reasoning in his article " “Digitizing Negatives with a Camera: Revisited”.

I don’t have a Thorlabs target or any other glass etched variant and so like most of us use Vlad’s Test Target. I don’t believe there is a glass-etched target that puts in all the detail across the frame in the way that Vlad’s target does in fact, not to mention the hugely useful Siemens Stars. Vlad himself acknowledges and explains the differences between his target and the glass etched options here and includes a very high res scan of a section of his target:

How good 35 mm test target actually is?

The bottom line is in my view that if your setup reveals detail in a smaller element group than it did before, with a new lens perhaps, a larger sensor or in your case pixel-shift then it is de facto evidence that you have improved the resolution of your setup, whichever target you are using.

[SSelvidge]

I read that long ago, and had forgotten, but at the time I was more interested in the NLP process section! I will refamiliarize myself! I hope I get the terminology correct and have not sounded foolish at any point… Ha!

Thanks @Harry for pointing to the link, which I have shared elsewhere for other parts of the article

If smarter people than me can take some swings at this and write about it, I am totally game