Hi all,
When processing raw files taken with a Fuji X-T4, in rather difficult light conditions (a show) with auto-ISO limited to 4000, I experience a severe noise pattern, that no demosaicing algorithm can get rid off.
With RT best 3-pass algorithm, a geometrical vertical-horizontal pattern is quite obvious.
Yes I tried all RT / DT demosaicing algorithms, the V/H pattern remains except for RT fast, the coloured pattern remains in all cases. The âremove false colorâ function is of no help.
On the darktable screenshot you seem to have denoise disabled so this is expected. The camera is denoising the image internally thatâs why you donât see it. Chroma denoise for Fuji is quite good so you lose nothing in enabling even the smallest amount of denoise.
I am aware that high ISO induces amplification noise, but it seems to me that nowadays ISO 4000 is no more considered as âhigh ISOâ, as stated in Demosaicing at high ISOs for X-TRANS - #2 by Claes.
What do you mean by âthere was not enough exposureâ ? (exposure parameters below)
Iâm not a current DT user, I just tried with DT in order to see if it provided other demosaicing algorithms than RT. For some reasons unclear to me, some parameters and particularly the âdemosaicingâ section, are switched off and cannot be switched on : an idea to fix it ?
High ISO doesnât cause noise, low exposure does In your example your your aperture is not as open as it could be, so exposure is obviously lower and thereâs more noise.
Thereâs probably a denoising section in RT where you can achieve the same effect. I rarely use RT so other people might help you better there
Please provide a screenshot. Even better, add a raw and a sidecar (does not have to be from the concert, but it should be taken under similar light, with identical settings).
It turns that I succeded by - as adviced in RT wiki for âhigh ISOâ processing - automatic chromaticity noise reduction and quite aggressive luminance noise reduction.
Do note that you were viewing the image at more than 100%. After that some interpolation sets in which could produce patterns that are not there originally.
Itâs expected that Fujiâs noise will look a bit different, especially since we donât know what algorithm they use internally to demosaic their images (afaik). While Markesteijn is good, their internal method might be better at dealing with those issues.
Is RT making interpolation when zooming raw files ? Beyond 1 screen pixel for 1 image pixel, doesnât it simply duplicate pixels ? e.g. 200% â 4 identical screen pixels for 1 image pixel ? Otherwise, it wouldnât be possible to see the matrix pattern, would it be ?
I am on my phone now, so I cannot see what you are referring to, but patterns are often a side effect of the sharpening with X-Trans. Turn off all sharpening and start playing with them one at a time. I would start with just âcapture sharpeningâ as I find that sufficient for my X-T20 95% of the time.
Bayer is a very different pattern from X-Trans; the demosaicing algorithms are different, and the artefacts are different.
It is certainly possible to get quite close to the in-camera output. If you brighten the image, you âboost the ISOâ, too, so with a 2 EV boost, youâre effectively processing an ISO 16000 imageâŠ
I find that output reasonable for an APS-C sensor.
@aldalalbi: The elephant in the room is that f/7.2 the photo was shot with. Given that you had a 16â80mm f/4 lens, you could easily have collected 3x the light just by shooting at f/4. No amount of post-processing can make that up later.
In your example your your aperture is not as open as it could be, so exposure is obviously lower and thereâs more noise.
It must use integer scaling for that case.