hi all,
I’ve been playing again with different demosaicing algorithms with both DT and RT.
In this case I want to show three results I’ve obtained with DT 3.2.1 and why I think it would be important to have one more demosaic algo in RT.
At low ISO the results are all quite good, even a 100% zoom, for VNG, Mark 3pass and for Frequency Domain Chroma.
The story is entirely different at high ISO (>3200) because the Markesteijn produces horrible artifacts such as small vertical bars and very high color noise.
VNG produces a lot of color noise, but at least it’s more uniform.
The best results are, by far, obtained with the FDC: chrominance noise is much much lower and no artifacts. The only downside of FDC is a small desaturation of very small details that is invisible when slightly unzoomed.
Of course I tried to increase Color Smoothing with Mark3pass, it’s a bit better, but not that much.
What I like of the FDC demosaicing is that the pattern you see is very very similar to the monochrome output with no demosaic at all that, at least for me, is very film-like.
wow, I totally forgot to write about my camera
it’s a X-T30. the sample I used above was at ISO6400.
I really like to shoot at 6400 or 12800 to emulate Kodak TRI-X.
Another way to avoid artifacts is to use passthrough (monochrome)
and then a gaussian low pass filter with radius ~ 0.6.
But then you lose the possibility to apply a specific LUT…
nice. you are zoomed in 1:1 here, right? would you be able to share the raw? i’d like to see how the simple gaussian splatting demosaicing would do on this…
The Frequency Domain Chroma demosaic looks really nice. I hope to see it in both raw editors (not that it personally matters, as I don’t shoot X-Trans).
Nice! Never bothered with other demosaic algorithms (on my Nikon I normally leave it at PPG unless I have to export pics for friends etc then I set it to Amaze; on the Fuji XT2 it’s Mark-1 pass…), but this particular example is quite striking!
I like the idea of having a noise pattern like Tri-X.
Have you tried the 4-pass (i.e. 3-pass + fast) algorithm in RT? It surely helps with the crosshair type of artifacts where they are the most visible (in the flat, low contrast regions).
@Claes: as far as I understand you cannot apply the LUT before demosaicing… it wouldn’t make sense when you still don’t know the RGB values for each pixel, well you’d have something like (127,0,0) on the red pixel etc… but you don’t know the spectral distribution that you get by demosaicing. Ah, that’s why I love Foveon sensors
Happy that FDC is working well with your X-T4! @hanatos: yep, 1:1 zoom. I’ll try to share the RAF on dropbox later if it’s useful, but there is nothing really special here: you can replicate that result by shooting at high ISO something with a high frequency complex pattern like grass or gravel. @sguyader: yeah, on RT I use 3pass +fast all the time (and it improves the results over 3pass alone), but to obtain something similar to FDC I should activate false color suppression. Anyway there are several situations like the blu noisy sky where FDC offers superior grain, much more film-like imho.
@black_daveth: because the simulated grain is nowhere as credible/pleasant as the one X-TRANS sensors produce at high ISOs. What do you use to add grain? Also at high ISO you lose dynamic range and it gives a nice film-like appearance.
Of course here I’m talking about producing pleasant images, not images trying to reproduce reality. In that case I shoot at base ISO.
I use the Add Grain plugin in G’MIC, I prefer the complete opposite approach, maintain as much detail and dynamic range as possible and then break up with grain and compress that range as I see fit after the fact. - when I can experiment and fine-tune.
do you have any finished shots that show the natural X-Trans grain you’re talking about? The FDC example in the first post doesn’t resemble film grain to me at all.
The original example wasn’t meant to be representative of film-like grain. I just said that FDC produces less “digital” artifacts than the other demosaic algos.
The best black and white film-like result I can obtain is by avoiding demosaic at all and using a gaussian filter with radius ~1. Of course you cannot do that with color.
I figured as much, but thanks for sharing that black and white, it doesn’t look too bad (the first one that is), I’ll have to experiment with that myself.
Yeah, the result is very nice too. Maybe the resulting grain is even more tri-x like.
There are hundreds of ways to obtain good results of course, maybe we could have in our programs a few different starting recipes, well optimized by experts, depending on our goals.
It seems to me that DT and RT both come somewhat optimized for the “daylight, maximum resolution, maximum dynamic range” scenario and to obtain good results under less than optimal conditions you’re left in some kind of terra incognita.
