Pixel Shift 16-shots versus Stitching

Revealing real-world examples about the nature of resolution you won’t find anywhere else. Pixel shift certainly works, but there’s much more to it you probably don’t know…

Introduction: setting the scene
Describing the Pixel Shift vs Stitching experiment
Important details about the sample images
Setup Notes
Test with Lens #1 GF35-70mm
Test with Lens #2 Mejiro Genossen FL0530
Test with Lens #3 GF110mm F2
Conclusion and thoughts

Introduction: setting the scene

On 11 Feb 2020, Jim Kasson published a very controversial post: pixel shift does not increase resolution. Quite immediately, it raised the ire of a number of (very experienced) photographers who insisted that they could plainly see far more detail in pixel shifted 16-shot images. So, what gives? Separately, Frans van den Bergh, one of the most respected in the field of MTF and image quality analysis, backs up Jim’s claim in a reply relating to MTF mapper (15 Jan 2021). I quote him, emphasis mine:

Pixel-shifted images are not supposed to produce MTF50 values (or actually entire SFR curves) that are different from those captured using the same sensor in its non-shifted mode, after accounting for scale as argued above. If you see large differences, then it is possibly caused by whatever processing the software does. In other words, pixel-shifting does not increase resolution (e.g., MTF50, or MTF10), but it does reduce aliasing significantly. The shifted images do look better, and may appear to have more detail, but this is a combination of reduced aliasing (less damage to fine patterns), and possibly stronger sharpening applied by the processing software. The best way to think of pixel-shifting is to keep in mind that the low-pass filter effect of the lens as well as the physical sensor photosite aperture remains the same when shifting, only the sample spacing is increased; this explains why the MTF50 / SFR is not expected to change.

Both Jim and Frans are looking at resolution in a very specific way to make this claim, that is, spatial resolution, which is what MTF quantifies. MTF describes how well contrast is preserved at different levels of detail. But it does not take into account the improvements from reduced noise, better demosaicing, cleaner details (from less aliasing, false detail, demosaicing errors), or better color reproduction, all of which could make an image look sharper or more detailed even if actual spatial resolution hasn’t increased.

If you don’t know what is pixel shift, first watch this video by Sony that I consider to be the best illustration of how it works. There are three main flavours of pixel shift, 4-shot, 8-shot and 16-shot. I’ll ignore the first two for these experiments, but I’ll list the differences so you have a clear picture of what they do:
4-Shot
The sensor is moved in 1-pixel offsets in a 2×2 grid. No pixel resolution gain. Colour aliasing is fully eliminated. No elimination of luminance aliasing.
8-Shot (Olympus, Panasonic)
Combines 4-shot with an additional 4 shots shifted by half-pixel offsets diagonally for added spatial sampling. Pixel resolution is increased four times (some interpolation must be involved). Some aliasing remains since only half of all possible sub-pixel positions are sampled.
16-Shot
Uses half-pixel shifts in a 4×4 grid to capture all 16 sub-pixel positions. Pixel resolution is increased four times. Total elimination of all forms of aliasing, but limited by the sensor IBIS positioning accuracy and combining software behaviour/bugs, as are 4 and 8 shot modes.

There’s even 32-shot Pixel Shift now, which is basically 16-shot Pixel shift but with two exposures for each sensor position. The double exposures are used to detect movement errors and for averaging down the noise to improve the signal to noise ratio.

The Pixel Shift vs Stitching experiment

I now own a camera (Fujifilm GFX 100S) that has this pixel shift function so I can test it myself, and in this article I mainly wanted to answer another question I have not seen satisfactorily answered anywhere on the internet apart from some anecdotal talk, by using real-world imaging, not theoretical simulations:

Is it better to use stitching or pixel shifting to gain more resolution for your pictures?

Let me explain what I mean, and set the parameters for the experiment. Take the Fujifilm GFX 100S sensor. In single shot mode, it takes 101 megapixel images, in pixel shift 16-shot (PS16) mode, you get 4x the pixel resolution, 404 megapixels. An object that occupies a height of X pixels in single shot mode will now occupy 2X pixels in the PS16 image. Pixel shift is limited by the total number of shots (camera dependent), while stitching is obviously unlimited by how many frames you can stitch together.

So to make a fairer and more meaningful comparison, I will specify the object of interest to occupy the same number of pixels in both the PS16 and stitching-equivalent frame. To achieve this, for a given setup for the PS16 image, simply double the magnification or double the focal length for the stitching-equivalent frame. In practice, it is a bit more complicated than that, but I will address those issues below. I also consider the practical limitations and issues of each method.

Important details about the sample images

I show a large number of test images in this article. They will take a while to load as they are 100% quality JPEGs rather than the typical compressed ones. If you are viewing this on a device with a high PPI display, you will be served “@2x”, “retina” images, which are 200% upsampled versions of the originals using Nearest Neighbour interpolation, i.e. there’s no difference from simply zooming in to 200% in an application like Photoshop. It is absolutely essential to avoid other resampling interpolators for this kind of comparison. In every other comparison from somebody else I’ve seen so far, the single shot was wrongly upsampled using something like Bicubic interpolation (or worse), which alters the spatial resolution (MTF) of the image! This will give you a false comparison, making the single shot appear much more smoothed and smeary than it really is, exaggerating the benefits of pixel shift. Not to mention many of the comparison pairs were very poorly prepared, often with different and wrong sharpening, strongly mismatched tone curve and camera profile, or heavily JPEG-compressed.

Look carefully at the indicated captions to see if the zoom magnification is at 100% or 200% or even higher. There’s a limit to how large an image I can serve up on this site so I tried to choose the most interesting and revealing crops to compare and show you.

The PS16 images are assembled using the FUJIFILM Pixel Shift Combiner app, which creates a DNG file, which is then processed in Adobe Camera Raw with all settings zeroed (true zero). Single frame images were processed in RawTherapee using DCB+VNG4 hybrid demosaicing with no false colour suppression and all other settings also zeroed (true zero). The main thing is to have zero sharpening, noise reduction, CA removal, and also the tone curve and white balance should be matched. Due to the forced embedded camera color profile in Fuji’s Combiner DNG, there are some unavoidable minor colour differences but that will not affect critical judgment of these examples in any meaningful way.
Update 8 April 2025:
I received criticism that the tests for Lens #1 and #2 are unfair comparisons since different raw converters were used. I would have used RawTherapee for everything except it doesn’t yet support superresolution PS16 files, only 4-shot is. But remember, demosaicing isn’t applied to PS16, so it doesn’t really matter what raw converter is used, as long as no sharpening or any other edits are applied, and the tone curve is a close enough match. Slight differences in tonality and colour is absolutely never going to throw off the far more obvious resolution differences we are seeing here. I’ve since downloaded a fully operational trial version of Capture One which allowed me to re-run all the files through the same software and use the same exact settings. Capture One actually shows a brightness difference between PS16 and single frame results, even though there isn’t any (stupid bug), and thus this criticism was proven to be totally unfounded. The renderings and conclusions are unchanged.

You can think of needing a 3×3 multi-row stitch with 50% overlap between rows and columns to arrive at a stitched result with an identical number of pixels as a PS16 image, i.e. identical pixel resolution. But since I’m just showing cropped samples below, I don’t need all the frames to build the complete stitch, only the central frame with double the lens magnification or focal length was collected. And I would like to avoid adding another variable of resampling which must occur during the stitch assembly process. All the crops are on the near the optical axis of the taking lens, avoiding any unnecessary optical messes that might arise as you move off-axis, plus I’ve taken great pains to ensure the lenses selected have been rigorously tested to exhibit no optical misbehaviours that might affect this testing.

Setup Notes

Camera: Fujifilm GFX 100S, used in single shot (often it is the first shot from the PS16 sequence, so as to ensure exactly the same focus position used) or PS16 mode (2 second delay between frames, timer release 2 or 10 seconds, mounted on Gitzo 3541LS Systematic tripod with Arca Swiss Cube C1 geared head. Set up on concrete, ceramic tiled, floor and target on quartz kitchen countertop or heavy oak table. Care was taken to ensure no detectable air currents, no vibrations from fans or running computers etc. The GFX 100S AF system proved to be grossly inaccurate in these experiments, since the depth of field was often super slim (~2mm only in the case of the GF110mm) and the targets chosen are extremely unforgiving to misfocus. The AF would often front or back focus (usually front) and I had to override it manually. The GF lenses by-wire focusing system also has another awful limitation of almost-far-too-large discrete steps in the motors that drive the focus position movements. You can actually hear the lens go “tick…tick…tick…” as it steps through each jump in focus position, rather than glide smoothly in an infinitely variable manner that the older, mechanically coupled focusing systems do. There were a few times I wished the steps were smaller so that more precise focus could be attained.

Lens #1: FUJINON GF35-70mm F4.5-5.6 WR, representative of a typical standard lens and aperture setting (f/8). Deliberately chosen to induce a slight amount of diffraction on the 3.76-micron pixel pitch sensor. This lens is by no means optically poor, though nevertheless is not winning first prize in any sharpness contest. Thorough evaluation guarantees mine is an excellent copy.

Lens #2: Mejiro Genossen FL0530 110mm F4, a high-end industrial line scan float lens, connected to the GFX 100S by custom adapters from RAFCamera and the Pentax 6×7 Auto Bellows. The image quality of this lens is outstanding and represents what you could expect from similar super-sharp lenses. Very few lenses can hope to be better.

Lens #3: FUJINON GF110mmF2 R LM WR, this was added on 12 March 2025 in response to another (unfounded) criticism I received, which was that the above two tests were plagued by diffraction for the PS16 shots, heavily penalising the pixel shifted results in an unfair manner. The premise was since pixel shifted subframes are shifted by 1/2 a pixel, you must use an aperture that is diffraction limited for 1.88um pixels, not 3.76um pixels (assuming 100% fill factor of the 100S), and so an effective aperture of f/6 is well into diffraction blur territory, rather you need f/3.5 effective aperture or wider. Knowledge of the theory caused me to disagree, but thought it was worthwhile to test it nonetheless, so there can be no further excuses to avoid the truth.

Multiple test shots were taken and scrutinised at 200-400% zoom, not 100%, in Photoshop to select the one with sharpest focus. All sample crops are from single images, not stacked nor stitched and therefore have not been affected by any resampling interpolation, other than demosaicing, or none at all for PS16 crops.

Test with Lens #1 GF35-70mm

I want to get this out of the way first. I’m using the same lens for both the 35mm and 70mm focal length test frames, that way I avoid the messes of choosing lenses with different optical designs that might favour the PS16 or stitching scenario. There’s never going to be a perfectly fair way of doing this, unless you do the test like with Lens #2 below. I am fully aware that zoom lenses do not perform identically across their zoom range. My copy of the GF35-70 is unusually good, well under 15 microns of tilt error and uncharacteristically even in resolution radially about the optical axis throughout the image field, as well as resolving quite similarly at both ends of the zoom range. I will leave you in no doubt:

Below are the comparison crop samples. First up are comparisons of high contrast text details displayed on a smartphone with a 537 PPI display (LG V30+).

PS16 (404 MP) vs single shot (101 MP):

PS16 (404 MP) vs stitching-equivalent (equivalent to 404 MP if stitch was fully rendered due to doubled magnification):

You might think “oh wow”, the stitching-equivalent looks way sharper than the PS16 image, and indeed it is because it actually resolves more in terms of spatial resolution, but it is still exhibiting aliasing typical of AA-filterless bayer sensors. What do I mean? Look below, from the same region, a 400% zoomed view, there’s clearly false colour aliasing around the high contrast text lettering. The PS16 image just barely reproduces the pixel grid pattern of the smartphone display (pretty cool to see that) while the stitching-equivalent frame is entirely unable to reproduce any of these details even though the text structures, which are relatively much more coarse, are rendered with much higher contrast. This so happens to be a perfect demonstration (it was unplanned!) of the conclusions I detail at the end.

Below is an extreme zoomed in crop animation of the simple circle shape taken from the bottom edge of the smartphone display region, with a red line overlay, which demonstrates Jim’s and Frans’ point that the contrast remains the same between single shot (101MP) and PS16 (404MP) images, and the blur width should also be the same. As the single shot is composed of far fewer pixels, it is undersampled and the sub-pixel detail reconstruction is comparatively coarse.

Next I will show comparisons of low to medium contrast organic details that are more commonly found in real world images.

PS16 (404 MP) vs single shot (101 MP):

PS16 (404 MP) vs stitching-equivalent (equivalent to 404 MP if stitch was fully rendered due to doubled magnification), 100% zoom:

PS16 (404 MP) vs stitching-equivalent (equivalent to 404 MP if stitch was fully rendered due to doubled magnification), 200% zoom of a different crop:

Test with Lens #2 Mejiro Genossen FL0530

This additional set of examples was inspired by my recent forays into the macro photography world. These photographers are some of the most technically proficient I’ve come across, apart from the astrophotography ones. Not surprising since each are specialised in imaging the smallest or furthest possible things, both demanding higher magnification than is typical of regular photography.

At these higher magnifications, there is something known as effective aperture that comes into play. It changes with magnification, as extension increases, the light cone has farther to travel and spreads out more, and the sensor “sees” a much smaller region of this cone. This makes the effective aperture smaller than the marked aperture, which is why macro lenses lose so much light at high magnification and diffraction is an oft discussed issue. I shot wide open (physically f/4) for both the 0.5x (for PS16) and 1x (for stitching-equivalent) magnification shots below, making the effective aperture f/6 and f/8 respectively, giving a 0.84-stop advantage to the PS16 image. If it doesn’t seem obvious already, I’m deliberately trying to help the PS16 results slightly to counter my personal bias towards stitching, something I’ve been a huge fan and practitioner of for landscape photography.

The target used is a 4-inch silicon wafer from 1986 by Rockwell International Corp USA. I learned from the late Robert OToole that these make excellent test subjects for macro lenses. It is a bare metal wafer with lots of fine details and especially good at revealing optical flaws. There were two setups in the testing, to position different subject matter on-axis, I identify them as “Crop A” or “Crop B” below.

Crop A, from one of the chips:

PS16 (404 MP) vs single (101 MP), 100% zoom:

PS16 (404 MP) vs stitching-equivalent (equivalent to 404 MP if stitch was fully rendered due to doubled magnification), 100% zoom:

Crop B, from the test structure/die section (I’m guessing, I’m no expert) with tons of super high contrast fine details:

PS16 (404 MP) vs single (101 MP), 200% zoom to better see the tiny details:

PS16 (404 MP) vs stitching-equivalent (equivalent to 404 MP if stitch was fully rendered due to doubled magnification), 200% zoom to better see the tiny details:

Here are the 100% crops from the single shot, PS16 and stitching-equivalent images, showing a fuller view of this very intricate and beautifully pattern section of the wafer.

Update: Test with Lens #3 FUJINON GF110mm F2 R LM WR

Update 12 Mar 2025: A fellow photographer raised the issue with me that the above two setups were severely diffracted for the PS16 results, favouring the stitching-equivalent frames very unfairly. Had I used a much wider aperture, the PS16 would be much sharper, likely sharper than the stitching-equivalent results. This is absolutely wrong, but I was willing to do all the work again and prove it.

I recommend that one starts by reading all of Jim Kasson’s posts in this DPReview thread, so you can understand the theory that since pixel shifting is essentially using the same sensor, with the same pixel aperture and pitch (he explains what’s the difference there), the pixel shifted result cannot possibly be diffraction limited at f/6 if the single shot images are not. You are effectively having a 400% AA-filter over a 1.88um pixel pitch with pixel shifting 16-shots by the GFX 100S, which is essentially arriving back at a 3.76um pixel aperture, so the diffraction limited aperture should be—exactly the same.

It is difficult to re-test with large aperture openings at high magnifications owing to the phenomenon of effective aperture, and available macro lenses with very wide physical apertures and of very high resolution at those aperture that are made for the GFX are virtually non-existent. So, I hatched a different plan: since the theory can be proven at any magnification ratio, I’ll shoot another test sequence at 0.05x magnification for PS16 and 0.1x for stitching-equivalent, and then I can use the gorgeous GF110mm, loaned to me by a friend. The 110mm is far and away the sharpest, highest resolving prime lens Fujifilm makes for the GFX system, so choosing any other lens would only shortchange us of sheer lens resolving power, plus its very wide physical aperture allows us to put the theory to the test.

I’ve posted a great deal of information and high-resolution samples over on the DPReview discussion thread I’ve started, to bring the discussion public and invite more data and views from those I consider to be of higher authority than myself on the matter. I shared samples that are larger than this website can easily show, so I recommend you head over there to check those out. There was a comment on the DPR thread that claims this new GF110mm setup shows much more similar resolution between PS16 and the stitching-equivalent results than the FL0530 tests, therefore proving the original PS16 results were diffraction limited. I made my own crops comparing the visual differences of the resolution gained by the FL0530 0.5x PS16 f/6 and 1x stitching-equivalent f/8 shots, against the GF110mm 0.05x PS16 and 0.1x stitching-equivalent shots both at f/2.8, there was no meaningful difference to my eyes. The hypothesis that the pixel shifted results are diffraction limited is just plain wrong, and Jim Kasson is right. Of course I know better than to tell Jim he is wrong!

The magnification (computed from the tape measure placed in the capture frames) and apertures I shot at are
PS16: 0.0496x, nominally f/2.8, measured by PhotonstoPhotos Optical bench at f/2.87, is f/3.0124 effective
Stitching-equivalent: 0.0939x (it should be 0.0992x, thus giving pixel shift the magnification advantage very slightly), nominally f/2.8, measured by PhotonstoPhotos Optical bench at f/2.87, is f/3.1547 effective

The effective apertures are well within the values demanded by the criticism. I have done nothing to bias the stitching-equivalent shot, rather I have actually given a slight advantage to pixel shift, if anything.

Essentially, my conclusions are the same. If anything, the tests with the GF110mm strengthen the effects already shown above, rather than disprove the theory. Also, this sharper GF110mm allowed me to reveal some truly disgusting pixel shift artifacts, zipper-like artifacts and 2×2 pixel blocking or clumping along high contrast edges. I consider the effect to be insanely ugly, it looks exactly like pixelation and low-quality demosaicing artifacts, which were the exact reasons why I wanted to use pixel shift to begin with. I don’t know if it’s due to Fujifilm’s Pixel Shift Combiner’s limitations, since PixelShifttoDNG does not work on its files, nor does RawTherapee.

Also note that despite the stronger sharpening recipe used to favour the PS16 image below, it never truly catches up to the stitching-equivalent result in sharpness, not to mention the overall level of artifacts put it well and truly behind.

Conclusion and thoughts

Resolution & Sharpness
PS16 Resolution:
– No meaningful change in spatial resolution versus single shot, edge contrast remains the same.
– Resolution of aliased details improves significantly, virtually no aliasing remains due to the accuracy of the GFX 100S sensor positioning, much more accurate reconstruction of fine details.
Stitching Resolution:
– More optical information per pixel results in truly higher actual resolving power. Can’t beat true magnification.
Winner for Maximum Resolution: Stitching.

Pixel-level Quality Factors (Aliasing, Noise, Artifacts)
PS16 Quality:
– Much better luminance and color accuracy, virtually no aliasing.
– No Bayer demosaicing errors.
– Reduced noise through multi-frame averaging.
– No blending or stitching errors.
Stitching Quality:
– Details below nyquist will be aliased.
– Still relies on Bayer demosaicing with its errors, not as perfect as PS16.
– Noise depends on exposure — no averaging benefit like pixel shift.
Winner for Pixel Quality (Pixel level accuracy): PS16

Other important Pros and Cons of each method
Pixel shift:
– Highly vulnerable to subtle camera/subject movement, changes in lighting. – Almost impossible to use outdoors.
– File size disadvantage, 16 shots must be taken, totalling 2.25GB per sequence for GFX 100s. For Sony, raw files are forcibly uncompressed (what a stupid software limitation), meaning you’ll be dealing with 3.85GB per sequence for a7R IV/V, despite less pixel resolution than the GFX 100S!
– Stuck with native aspect ratio of sensor.
– Note that the Fujifilm Pixel Shift Combiner app will throw out a somewhat scary error message “defect detected” if your images have out-of-focus or diffraction blur, but the output DNG is still properly constructed without artifacts.
Stitching:
– Can suffer from blending artifacts, depending on parallax errors, degree of overlap and stitching software among other issues.
– However, with significant overlap between frames, minor subject movement seam errors can be mitigated through intelligent seam carving in the stitching software, making it viable for usage outdoors.
– File size advantage, 9 shots to achieve even greater resolution than pixel shift, ~1.25GB per sequence for GFX100S.
– Free of native aspect ratio of sensor.
– Resolution upside is unlimited.

I learned from Jack Hogan the theory of why doubling the magnification would easily win outright. It is simply this. When you double the magnification, it’s the same as your subject growing 2x linearly in size. So what was 100lp/mm detail is now 50lp/mm. If you are able to get MTF data for your lens, you will know that the lens would resolve significantly higher contrast at half the lp/mm, so doubling the subject size basically taxes the lens’s resolution much less, or you can say you have more resolving power for larger details.

I think the benefits of pixel shifting has merits, but also limits. If the data load is not an issue, it can be very helpful to collect full pixel-level data for digitising collections and archives of all sorts. If you don’t want to mess around with telecentric lenses to avoid parallax issues, it’s also easier to use PS16 for high magnification macro setups than to try and stitch. For everything else, stitching is probably better and more useful.

If you are reading this, feel free to make your own tests! Especially if you have an a7R V. I would love to hear about your results, you can share them here.

Samuel Chia
5 Mar 2025
Updates have been added throughout the article.