Detail Recovery setting neural restore module

Software darktable
1

Just wondering on recommendations for setting the detail recovery slider in the neural restore module in DT 5.5

image
image534Ă—101 2.1 KB

2

With PR 20854 this slider will be renamed to “strength” and it’s direction will switch.

So with strength = 100% (and detail recovery = 0%) you got the image from denosing model ouput. Sometimes, especially at very high ISO, it may produce over smoothed image with plasticky look. To reduce this unwanted effect we apply blending with the original noisy image. So strength lower than 100% adds proportional fraction of noisy image back to the final result. Raw denoise task uses regular blend, denoise task has improved blending - oroginal image is additionally DWT filtered to bring back only texture.

So, I advise starting with 100% strength (or 0% detail recovery) and reduce it if needed for a particular image.

3

Thanks for the explanation. I wonder if it would be possible to consider having the ability to apply 100% denoising and then having something like an opacity slider to tone it down afterwards. I often find in DT I use the opacity slider for modules to control the strength of the correction. Probably what I am suggesting is totally impractical since I don’t understand the coding effort required for such a request.

I did try the composite module to combine the tiff output with the original raw but that didn’t work. I guess the preview window is the way to go with the current setup. The ability to use drawn and parametric masks to combine this denoising would also be useful, but I presume impractical to implement. Sorry for the noise and thanks for the good work with this.

EDIT: doing two tiff file out puts with the detail recovery slider at opposite ends allows a combination of the two outputs to be achieved in the composite module. This also facilitates drawn and parametric mask options.

4

That basically what it is doing right now. But this is an utility module as of now, which works in both lighttable and darkroom. So full image interactions are not applicable in current implementation.

5

So maybe if it was a darkroom only implementation the ability to do drawn and parametric masks as well as opacity might be possible? Combining two tiff outputs with the composite module was helpful for my test images.

6

It is not that easy to make it IOP module. It creates a new image, not just non-destructive operation. It requires new architectural decisions. Not a quick fix.

7

That is what I thought. It is easy as a user to ask for the impossible because we would like it to work that way. Thanks for taking the time to answer my question and understanding the slider for detail recovery now makes sense to me and helps me maximise the AI denoising.

8

@anry … any idea when the Raw Denoise will be available in a nightly build ?
Just like to play around with it …

Regards Andreas

9

Hopefully within next couple of days.

2 Likes
10

:clap: :+1:

11

It´s now there … cool , will experiment with it !!!
Already found a bug at first glance …
At least on my machine ( Mac Studio M1 running latest Tahoe ) I do get a error message " EOS R1 and EOS R5 II color matrix not found " … every time I do open a module and try to work on the DNG file .
Any idea where it does come from …
It was happening the same way when I started using RawForge plugin from @AyedaO .

Kind regards Andreas

12

I was not able to reproduce this message, but I discovered that some metadata tags in output DNG were missing. Hopefully, this PR will fix it.

13

Hello @anry, Thank you for the PR and thank you @helgeschneider for reporting. I had the same issue and the PR works for me.

1 Like
14

This is what happened when I ran neural restore -> raw denoise without the PR.

I if compiled and installed from scratch everything was fine. The imported raw file ended up like:

  • input color profile -> input profile = standard color matrix, working profile = linear Rec2020 RGB and

neural restore -> raw denoise produced a dng

  • input color profile -> input profile = embedded matrix, working profile = linear Rec2020 RGB.

However my regular darktable installation has an input color profile preset which changes the working profile to linear ProPhoto RGB. The preset applies input color profile -> input profile = standard color matrix, working profile = linear ProPhoto RGB to all raw (and dng) files. This means that my imported raw file becomes:

  • input color profile -> input profile = standard color matrix, working profile = linear ProPhoto RGB and

neural restore -> raw denoise produced a dng + preset is applied:

  • input color profile -> input profile = standard color matrix, working profile = linear ProPhoto RGB.

Going to darktable and moving to the new dng produced the warning: “EOS R5 II color matrix not found”. Manually changing input color profile -> input profile = embedded matrix did not help either. The warning persisted. EDIT: I also got the wrong colors with input color profile → input profile = standard color matrix.

If I apply the PR [AI] DNG writer: emit missing metadata for spec-conformant output by andriiryzhkov · Pull Request #20913 · darktable-org/darktable · GitHub then I can keep my input color profile -> input profile = standard color matrix, working profile = linear ProPhoto RGB preset and I get no warnings. EDIT: and I get the correct color rendering.

1 Like