Fujifilm GFX100 Pixel Shift support

So the GFX100 now has pixel shift support to produce 400mp images. I downloaded the “FUJIFILM Pixel Shift Combiner” app, however it doesn’t work with Wine.

Any chance this could be built in to RT?

Do you have raw samples? Also upload them here https://raw.pixls.us/

Tried to upload pixls but it didn’t accept it. Here’s a link to a ZIP with all the images

Hi,
would you also be able to upload the resulting dng? Thanks!

I haven’t been able to get it to run on Linux. If you have access to a mac/windows machine the Fujifilm software is here:

Should be able to just drag the 16 images into it.

Thanks for the link. Unfortunately, I only have a win7 vm, and it doesn’t seem to work there…

Here we go. Was able to use the combiner on my work machine

Some progress update:

  • The DNG file you provided (thanks!) can be read just fine by ART (and I suppose RawTherapee as well), but it is already combined into an RGB image. Therefore, you can’t use the automatic motion detection and correction implemented in RT (which ART inherits)

  • However, I managed to extend my make_arq tool to make it generate multi-frame “pseudo ARQ” files also starting from the GFX 100 RAF sequence. This will result in a “pseudo-ARQ” raw file with 4 frames, which can then be “demosaiced” using pixelshift and automatic motion detection. Currently only ART understands this, and only if you use the fuji-gfx100-pixelshift branch that I’ve just created. This is in the fuji-gfx100 branch of make_arq.

  • The nice thing about using make_arq (besides the fact that it is free software and it works on Linux too) is that it allows also to create 4-shot pixel shift files for the GFX 100, i.e. it allows to have the benefits of pixel shift (fewer artifacts/false colours) without the hassle of having to deal with 400 Mpix files, which are a real pain to use, at least on my machine. 4-shot pixel shift also avoids the uncorrectable motion artifacts that happen in 16-shot mode.

  • The downside of using make_arq is that it creates pseudo-ARQ files that are only understood by ART, and that are also completely uncompressed, so huge. A 4-shot file is about 789MB, and a 16-shot one is a whopping 3.1GB – as I said above, not nice to work with.

  • Here are a couple of screenshots just to show that I’m not making things up:

  • Finally, the scene in the shots is really sub-optimal (to say the least) for testing, as it is underexposed, misfocused, and without much detail to begin with… if you have the chance of posting a better picture (ideally properly exposed and with everything in focus), that would be definitely appreciated :slight_smile:

4 Likes

What exactly happened? Was there an error message? I’d like to let @andabata know if we need to look at something.

That’s awesome! Would it be possible to convert the arq into a DNG (or RAF), allowing either 16- or 4-shot?

Apologies for the quick photo. Here is a better photo. Towards the bottom of the book seems to be the best in focus point.

Not sure what happened, but the forum will no longer allow me to post a Google Drive link. So here is the link via another method:
https://pasteall.org/LtMH/raw

@patdavid it said I didn’t check the boxes (I did) or that something else went wrong. I did try to upload a .ZIP with all 16 images in it.

1 Like

Thank you!
(The forum spam filter flagged your posts, so I’ve restored them for you )

I’ll look and see what’s going on.

1 Like

How did you try to upload them ? As a zip ?

Yes as a .zip

We don’t support uploads as a zip. Only single raws

Hi,

some more updates: I’ve merged the fuji-gfx100 branch of make_arq to master, thanks to your new samples I managed to fix the alignment issues and now it should work properly.

you can use ART (still in the fuji-gfx100-pixelshift branch for now, will be merged soon I hope) to generate a linear TIFF which you can continue editing with your preferred tools. In principle it is possible to have make_arq generate a linear DNG, but I don’t have much interest in doing that, sorry… (I am always happy to accept patches though :slight_smile:

One final note: if you are building ART yourself on Linux, consider not linking against tcmalloc, as it seems to go really crazy when you start allocating large chunks of memory, which is what happens with the >400 Mpix files generated in 16-shot mode. I don’t know what it is, but either the default malloc or mimalloc behave much better in this use case.

Merged in art master too, fyi

That was fast, excited to try this out! :slight_smile:

@andabata that was my guess. In this case, all 16 .RAFs needed to be together, so wasn’t sure if it made sense to upload them separately.

Well just to contradict myself, this turned out to be easier than I thought. So now you have a --dng flag to produce a linear, already-demosaiced dng which should be understood by other apps (I only tested with rawdigger though). But if you use art, the arq is better, since it will allow to automatically detect and correct motion, something you can’t do in the dng

They are effective 1 image ? with pixelshit ? So every pixelshifted image gets exported as 1 raw ?

Correct, you get 16 raws.