Denoising high-ISO images: a sourdough bread

Bill I just think you are over thinking it but don’t take my word for it…At 100% or 1:1 the every pixel of your image is mapped to a pixel on your display…so this is why 100% is used to look at image detail for tools that modify it…noise sharpening etc…if you go higher or lower with your zoom the relationship is no longer 1:1 there has to be some mapping…its explained better here…so if you are at 15-25 % or so in reality to fit to screen or whatever that number is then things get combined and that is why you can’t see your noise…explained better here… http://www.inzomia.com/main/products/ZoomStudio/zooming_explained.pdf

Even with good old adobe software people can run in to the same situation that you have…

Yes it is and thank you very much. I shoot a lot of high ISO images and it’s been a great improvement

1 Like

As I always say, I’m slow but not fast. I get it now, what you are saying about how it is displayed. What will a printed photo look like, what I see in dt or the jpg.

1 Like

Most probably neither.
It depends on your printer & printer profiles, and inks, and paper, and…

I was interested in the noise I would see. I send jpgs to an online printer.

Echoing what @Claes says you would likely have to do some test runs to know for sure and as with the zoom and display issue the size you print will impact the final result

I have done an RT development of the raw.

RT preview

RT jpg in GIMP

dt preview

dt jpg in GIMP

Not sure how to use that info but looking at this another way and I and not sure how darktable modifies the image when you go to less than 100% size by zooming ie which algorithm but lets say you go to the size drop down …pick small …on my system that is about 12%…scrolling up to about 25% leaves you at about the fit to screen zoom size…so what I think some people might think of as zero zoom really is not …zero would really be 100% in my mind anyway nevertheless now export your image at 25 percent and compare with your standard export done at 100%…You will see it looks like what you see in the display ie 25% display looks like 25% export and the same for 100…so to me DT is quite faithful creating the exported image…you just need to know that when you look full screen this is something like 25% ie sized down so it will not accurately reflect the data…I am not sure how other programs work with sizing and rescaling but DT seems pretty faithful to me??


I think RT’s jpg is more faithful to the preview image. I developed 4 other photos from that trip, and I was very pleased with the resulting jpgs. This one stood out to me immediately as inferior to the preview. Don’t know why that would happen. Time to let it go.

I think (know?) Some effects in darktable that are not ‘true’ to the final result when viewing at less then 100%. Less pixels get fed into the algorithm for previewing , so the outcome changes.

This can be specially true for sharpening, denoising and anything working with local contrast.

Rawtherapee has ‘1:1’ markers next to modules, to indicate that those modules should be judged as 100% or more. DxO applies some effects only at 75% or more, and their heavy denoising will not preview at all.

I believe from memory there was a setting to use a different , cheaper end faster demosaicing method when previewing images zoomed out in darktable?

Could also be that you don’t export to regular sRGB but a different profile, and irfanview is not color managed until you enable it in it’s settings. Although I doubt that will cause stark differences like this.

Of there is no difference at 100% , then it’s a nasty lesson to always remember they anything you see - any program , always - at less then 100% is always an ‘estimation’. Some programs just to out of their way more to make it less annoying or noticeable.

In short (and from memory)

  1. There is the config option defining if you use full demosaicing or a fast path, effectively falling back to PPG while in zoomed in darkroom having worse results with noise.
  2. As the calculated image section is larger than the one you are viewing there is some downscaling. Lancos3 gives overshoots in such cases so you might prefer bilinear.
  3. Some modules require full image data, if they see only parts of the image output will be different. This is not a problem in your case.
  4. Prefer performance over quality should only be switched on using very slow machines.

This all contributes to other output while in darkroom compared to writing files. So what you see is expected and should be understood as a compromise to keep processing requirements under control.

Is this what you are referring to?
image

Also,
image

@jorismak

image

I see that if I zoom to 56%, with the added noise the preview looks very similar to the jpg. When I zoom in at 100%, the noise is so prevalent, that there is no relationship to what the jpg looks like. I can try editing at 56% to see if I can improve on the jpg without making other problems.

I agree from what I see the displayed version in DT reflects the display scaling. The output does not…so if you zoom in to 50% then if you export at 0.5 instead of 1 the exported file will be what it looked like on screen and it not it will look like what it does when you zoom to 100%…if the export scaling factor is 1 then it does matter what the screen zoom is it will not accurately reflect the output unless that zoom is 100…at least from my experimentation…

1 Like

Which setting is that?

image001.png

1 Like

Where is this ‘scaling factor’ ? Or is it something in the print output or something?

image001.png

fwiw this is currently the best i can do in vkdt:

i think quality can clearly be improved, but this one runs the denoising in 19.odd milliseconds on an entry-level gtx1650 max-q (laptop), full resolution image.

i know i felt really strong about darktable's region of interest/scaled/cropped rendering pipeline even at the libre graphics meeting (when discussing librtprocess). but at these speeds i just always process the full buffer now, and display whatever lands on screen.

this does not resolve all the abovementioned issues wrt scaled display (the image still needs to be resampled which may show or hide noise), but at least somewhere in the background, there actually is the same image as would be generated during export. this makes coding modules that require a large filter footprint a lot easier, too.

1 Like

Here’s my attempt with the latest Filmulator.

It’s… orders of magnitude slower than VKDT, alas, but I think the quality is pretty good.
Screenshot_20211113_071410

Everything default except

  • Demosaic Method on (LMMSE)
  • NR On
  • NR Strength 18.96
  • Speckle NR Strength 2.826
  • Chroma NR Strength 29.17
2 Likes

I may not have the time but recent activity here has reignited my interest. Could someone please create or link to a nicely processed image but without denoising for me to play with? I may get back to you in 1/2-1 year but yeah it would encourage me to get back into playing. :slight_smile: