Macro/pixel shift test - thoughts on validity of this approach?

[gerald.d]

Hi all -

Now I've got my Fuji GFX 100S, I thought I'd investigate for myself what the situation is with the pixel shift capability. Obviously based on everything I've seen elsewhere, I don't believe it is fit for purpose (yet).

Of course if one wants to multiply the resolution of an image by a factor of 4***, then there are ways other than pixel shifting that can provide the same (?) result. One of these ways would be to increase the magnification by a factor of 2, and then shoot four shifted images so that they encompass the field of view given from the single image taken at the original magnification. Obviously this in a theoretical scenario - you'd have to shoot some overlapping shots as well to actually make sure you captured everything.

However for the purpose of my proposed test, I don't really need to capture the entire original frame at twice the original magnification - just the center of the fame should suffice.

The CAPcam that I've got my GFX mounted to makes this a pretty trivial thing to test - I have a single checkbox that I can tick in the CAPcam software that will put the camera into macro mode. I.e. set up a shot at 1:2 magnification, check the box, and the lens will move to the right position to capture the image at 2:1 - no other input needed.

And of course, there is an initial magnification that can be used that, when flipping the numerator and denominator, will result in a doubling of the magnification. I'll spare you the maths, but it is 1: (square root of 2), or, 1:1.41.

So the intention is to shoot at 1:1.41, both a regular 100MP shot (for reference) and a 400MP pixel shifted shot. Then, move to 1.41:1, and shoot a regular 100MP shot.

My question is this - is it a valid comparison to take the center of the frame of the 400MP shot, and compare it directly to the 100MP shot that has been taken at twice the magnification?

My lens should be up to the task - it's the Linos Inspec.X L 5.6/105 Float. Specifically designed for optimal performance in the 1:3 to 3:1 macro range, and with a floating element that you set to the magnification at which you are shooting.

Before I set out to do this, I would very much welcome any thoughts/suggestions/criticisms about the approach.

Kind regards,


Gerald.

(*** or I guess, 2, if you're being pedantic - let's not go there)

 

[SCHWARZZEIT]

Hi Gerald,

you should keep in mind the effects of diffraction at those magnification levels.

At 1.41x wide-open your Linos 105 Float will have an effective aperture of 13.43. If your lens had diffraction limited performance at that aperture and magnification you would have a modulation of just about zero at the Nyquist frequency of the 3.76 μm pixel pitch sensor of your GFX 100S. This means the image will not be super sharp on pixel level but still very good by most standards. In case you think about using the pixel-shift feature at 1.41x the only improvement you can expect is the mitigation of any debayering losses. A 4-shot 102MP pixel-shift function would be more efficient at this.

At 0.71x magnification your lens will have an effective wide-open aperture of 9.64 and, again assuming it’s diffraction limited there, the absolute resolution limit in terms of smallest resolved features would be about 2.7 μm in the image plane. With the 408 MP pixel-shift function your image pixel pitch is only 1.88 μm. This is quite a challenge for a lens designed for a 5 μm pixel pitch. Even if you supposedly had a real 408 MP sensor (without the quirks of generating 408MP from 102MP through pixel-shift) in your GFX you would not be able to get the same object resolution at 0.71x compared to 1.41x at 102 MP with this lens. You would need a faster higher performance (at 0.71x) lens for this.

-Dominique

 

[Shashin]

I think you reasoning is good. Simply cropping the center out of the center of a 400MP image might actually be better than doubling the magnification and keeping the original 100MP. It is certainly worth a try.

 

[gerald.d]

Hi Gerald,

you should keep in mind the effects of diffraction at those magnification levels.

At 1.41x wide-open your Linos 105 Float will have an effective aperture of 13.43. If your lens had diffraction limited performance at that aperture and magnification you would have a modulation of just about zero at the Nyquist frequency of the 3.76 μm pixel pitch sensor of your GFX 100S. This means the image will not be super sharp on pixel level but still very good by most standards. In case you think about using the pixel-shift feature at 1.41x the only improvement you can expect is the mitigation of any debayering losses. A 4-shot 102MP pixel-shift function would be more efficient at this.

At 0.71x magnification your lens will have an effective wide-open aperture of 9.64 and, again assuming it’s diffraction limited there, the absolute resolution limit in terms of smallest resolved features would be about 2.7 μm in the image plane. With the 408 MP pixel-shift function your image pixel pitch is only 1.88 μm. This is quite a challenge for a lens designed for a 5 μm pixel pitch. Even if you supposedly had a real 408 MP sensor (without the quirks of generating 408MP from 102MP through pixel-shift) in your GFX you would not be able to get the same object resolution at 0.71x compared to 1.41x at 102 MP with this lens. You would need a faster higher performance (at 0.71x) lens for this.

-Dominique

Hi Dominique -

Thanks for doing the maths - I'm aware that I'm knocking on the door of the limits of the system here, but not entirely sure what to expect, hence the planned test.

If I understand correctly, you highlight two things running up against one another here - diffraction increasing at higher magnifications with the same physical aperture, and smaller effective pixel pitch when utilising the pixel shift of the camera.

I think what you're saying is that I shouldn't expect to see any more detail in either the 1.41x 100MP image or the 0.71x 400MP image than I will see in the 0.71x 100MP image.

Is that right?

Kind regards,


Gerald.

 

[Abstraction]

I don't quite get why the effective pixel pitch would change with pixel shift. It doesn't change when we stitch. This is sort of like stitching in camera with small movements. Why would that affect the pixel pitch?

 

[Shashin]

Pixel pitch is not changing the effective pixel pitch of the sensor. The smaller pixels are a product of the processing, not the pixels themselves. If you are judging based on 100% monitor view, then the difference is just magnification.

 

[gerald.d]

OK, so initial results are in.

• There is significantly more detail in the image at 1.4x 100MP, than there is at 0.7x 100MP. It's night and day. Whatever the maths may be regarding diffraction limited lens performance at a physical aperture of f/5.6 and at these magnifications and pixel pitch, the eyes don't deceive.
• There is barely any more detail in the 0.7x 400MP compared to the 0.7x 100MP. The combing artifacts are there to see in areas of high spatial frequency. The problem manifests itself in a very visible manner in these areas, but I rather suspect it is actually ruining the entire image. As it is currently implemented, my view is that Pixel Shift on the Fuji GFX 100/S is a worthless marketing gimmick.

Crops from the files available here -

Fuji GFX 100s pixel shift testing - Google Drive

drive.google.com

No need to name which is which as it will be obvious once you look at them.

NOTE - I am making these available for personal use only by those who are interested in this subject. Under no circumstances may they be shared or rehosted anywhere without my prior permission. Feel free to embed the images in this thread if you need to for the purposes of discussion.

Kind regards,


Gerald.

 

[Abstraction]

Pixel pitch is not changing the effective pixel pitch of the sensor. The smaller pixels are a product of the processing, not the pixels themselves. If you are judging based on 100% monitor view, then the difference is just magnification.

Right, so I don't see how pixel shifting would tax the lens any more than not pixel shifting.

 

[SCHWARZZEIT]

Hi Dominique -

Thanks for doing the maths - I'm aware that I'm knocking on the door of the limits of the system here, but not entirely sure what to expect, hence the planned test.

If I understand correctly, you highlight two things running up against one another here - diffraction increasing at higher magnifications with the same physical aperture, and smaller effective pixel pitch when utilising the pixel shift of the camera.

I think what you're saying is that I shouldn't expect to see any more detail in either the 1.41x 100MP image or the 0.71x 400MP image than I will see in the 0.71x 100MP image.

Is that right?

Kind regards,


Gerald.

Gerald, in the scenario you described and tested, at 1.41x you’re on the edge of what you can get out of your system with a single shot of a 3.76 μm pixel pitch sensor. I wasn’t saying that you’re not going to see more detail compared to 102MP single shot at 0.71x. On 0.71x your lens should outresolve the sensor by some degree that you should see some benefit from utilizing pixel-shift but it will be a far cry from going to 1.41x. Your test confirmed this.

Thanks for making the crops of your test available. I agree that the combing artifacts from the pixel-shift are ruining the impression of the gain in effective resolution. But you can almost completely eliminate these combing artefacts by applying a Gaussian Blur filter with a 0.5 pixel radius in Photoshop and then resharpen the image. The reduced noise from the 16-shot allows for quite heavy sharpening. However, too heavy sharpening can reemphasize residues of those combing artifacts. You may want to try different blur and sharpening settings to find the right balance. While the result will be nowhere near the level of detail of your 1.41x image it shows that you can get significantly finer details resolved than the 0.71x single shot image.

I don’t think pixel-shift is just a gimmick. In some situations where additional magnification and stitching is not an option it can be a useful tool. But without more sophisticated processing it only works when both camera and subject and everything else in the frame is absolutely still. This is sometimes beyond the control of the photographer and thereby requires additional quality control for each processed shot.

-Dominique

 

[Christopher]

No surprise here. Pixelshift kinda sucks on the GFX and isn’t very useful. While it works well on some other cameras it’s not a great feature on the GFX. I would prefer something like Phase Ones frame averaging 100 times over this not working 400MP mode.

 

[gerald.d]

I don’t think pixel-shift is just a gimmick. In some situations where additional magnification and stitching is not an option it can be a useful tool. But without more sophisticated processing it only works when both camera and subject and everything else in the frame is absolutely still. This is sometimes beyond the control of the photographer and thereby requires additional quality control for each processed shot.

-Dominique

Just to clarify - I'm not saying that pixel-shift is a gimmick per se, I said that "Pixel Shift on the Fuji GFX 100/S" is just marketing BS at the moment.

Hopefully they can fix it in firmware at some point in the future - I'm assured that they are aware of it.

Kind regards,


Gerald.

 

[Will Ophuis]

I was just testing the pixelshift out on my 100s and was pretty disapointed, even parts of the image that are still didn't look sharp? I would be using it to get better colour rendition but looking at the files and artifacts its not worth it, hopefully they can update firmware to improve it.

 

[drevil]

they say most pixelshift solutions create relative soft images, but my guess as there are way more pixels on the gfx100(s) the "issue" becomes bigger as well, just my guestimate