Denoise, DT versus RT


I like your pp3, being straightforward and effective.
(But the chrominance curve has a big impact here too. 5.5-243 win.)


(Olaf Berthelsen) #45

When I use my original uploaded pp3 then the chrominance curve has a big effect in RGB color space. This is also the case if I set the denoise parameters equal to your settings. If I use yours pp3 the effect is very limited???


(Olaf Berthelsen) #46


The surprising difference in the pictures below is caused by difference in the demosaic settings.



I thought that LMMSE was recommended for noisy images, but the result is clearly not good in this case. Good guidelines will be appreciated…….:o)!

1 Like

(Ingo Weyrich) #47

Well, the false color correction steps setting (which is applied after demosaicing) is not the same in your comparison…


(Olaf Berthelsen) #48

I don’t understand what you are trying to tell me.
The difference between the two images is caused by changing the following RAW settings (by copying the settings from another profile) Demosaic method, Pixel Shift and False color suppression. Where are Pixel Shift and False clolor suppression situated in the raw settings?

Normally I would never try to change these settings since I don’t understand what they are supposed to accomplish. I my standard profile I just changed the demosaic method to LMMSE (recommended in Rawpedia) then the rest came by default.

Since the difference is so big some guidelines would be very nice…


(Sebastien Guyader) #49

Look at the 2nd slider in you demosaicing settings screenshots. You’ll see that you have false color suppression set at 2 for LMMSE, whereas it is set at 0 for AMaZE.
For better comparison, turn both down to 0, or set both to 2. The result could be different from what you observed at first.


(Alberto) #50

fwiw I always use amaze. I did try others in the past, but for my camera amaze always delivers (sometimes I need to tweak the black levels though)


(Sebastien Guyader) #51

Last week I’ve participated to a discussion on dpreview, where we showed that under some circumstances, the LGV algorithm is good to remove Moiré patterns (in that case, the Moiré was caused by the pattern of windings around some guitar strings, and AMaZE was worse).

Edit: but the images were not noisy, so I was digressing a bit


(Olaf Berthelsen) #52


You are absolutely right. Changing the “False color suppression” makes a very big difference both with LMMSE and AMAZE. The results get worse every time the setting is increased both in AMAZE and LMMSE. You would expect the opposite would you not?

The results are best when the setting is zero (and the LMMSE enhancement steps are zero as well) but AMAZE is still the best choice.

What does the “Border” setting do? It’s not documented in Rawpedia.


(Olaf Berthelsen) #53


I also like your result. What is darktable master? It’s not possible to enter these settings in denoise(profiled) version 2.6.0.


(Ingo Weyrich) #54

At the default setting of 4 a 4 pixels border will be skipped because it can not be demosaiced in the same quality (missing pixels at at least one side).
As a consequence, in past, demosaicing full hd video raw files (1920x1080 pixels) always resulted in smaller output (1912x1072 pixels) .
For that reason rt now allows to set the border to zero to get the full sized image (though with reduced quality at the 4 pixels borders).

1 Like


darktable master corresponds to the development version of darktable, so you will have to wait for darktable 2.8 (or to build darktable from the sources) to be able to use the same settings



Here is my take on denoising. Not as good as some of the above, but hey it’s me fooling around using unconventional means, without the help of the usual tools. :stuck_out_tongue_closed_eyes: As usual, the minimalist approach.



(Olaf Berthelsen) #57


Thank you for your response. Im curious to know what those unconventional means are……?




Generally, I treat my processing workflows as a science experiment. I tested the ideas that I had and used the best combination of them to come up with the result that I could appreciate. Will inevitably not generate as many likes but that isn’t the point. The point is the process and having fun doing it, hopefully learning something new along the way.


Objectively and subjectively, there are many views on how noise should look like in a photograph, some of which were talked about in this thread. My goal was not to emulate real life grain nor make the image devoid of noise. I was to make it look pleasant; make the noise less of a distraction from what would otherwise be an excellent image.


I am aware that there are filters, commands and apps made by people who are way smarter and much more proficient than me; but I normally script my own with G’MIC using my own ideas for the thrill and challenge. Just a newbie trying to learn at a breakneck pace. Occasionally, I discuss them openly, usually when I think the ideas are good enough to share to the wider community.


This is a long way of saying that even if I told you the exact details, it wouldn’t be meaningful for general use. I haven’t shared the scripts that I have been working on since they aren’t ready. In this sense, I am similar to @agriggio and his custom RT.

In the PlayRaws, I do however itemize and share the sequence of processing step. Very clearly in fact to give people a sense of how they could achieve a similar result using their preferred tools. Recently, I have decided to not be as thorough, since it takes effort to write up a step-by-step workflow that no one else does. But since you asked, these are the strategies that I used in the final result.

0. Unclipped linear sRGB from PhotoFlow.
1. Preprocessing.
2. Flatten troublesome impulse noise.
3. Guided smoothing with guided filter.
4. Rolling guidance with bilateral filter.
5. Curve manipulation with wavelets.




What I found a challenge was low-frequency noise; I think the synonyms for this might be splotches or speckles. They aren’t as bad in my final result but are still annoyingly present. If you compare with @agriggio’s or @calgarykev’s, you would see that they mostly overcame it.

It is representative of my current math and coding knowledge skills, or the lack thereof, because now we are talking about more advanced techniques such patch- or profile-based denoising. G’MIC has a builtin denoise_patchpca and nlmeans but my computer can’t handle them. G’MIC scripts aren’t optimized like dt’s or RT’s tools.



(Olaf Berthelsen) #60

Hi’ afre

Thank you for your response (How do I copy just your name from the thread?)

I’m impressed by your scientific approach well above my level and skills, so I would not benefit much from a detailed description of your scripts.

I agree that it’s important to have fun(!) editing photos. I also want the photos to be improved and to look pleasant, so that I can share them with my friends and recall the situations in the years to come. The bulk of editing has to be quick, since it is not possible to do scripting or other heavy editing separately on 300 photos form a safari in South Africa.

I concentrate on learning how to use readily accessible, user friendly applications like RT or DT. When I left Adobes software (because I don’t want to pay a monthly fee in all future) I focused on RT since DT was not available in Windows. It is now, so I’m also studying DT. DT has some very nice tools not available in RT.

My original ambition was to understand the tools so well that I was able to look at a photo and decide which tool(s) is needed to arrive at a satisfying result. This has proven to be over ambitious. Experimenting is needed in many cases. But I want to understand what is going on when you adjust a slider.

This thread has demonstrated that you can get a good result in (many) different ways and that some noise is positive in many situations, but not splotches or speckles, of course.


(Sebastien Guyader) #61

@obe Simply add a “@” sign in front of the user’s nickname, without a space (a list should pop up to let you click on the relevant name).


(Olaf Berthelsen) #62

Hi’ Everyone

Thank you for all your interesting responses. What is the result of this long discussion. Well, to me it seems to be:

Both RawTherapee and Darktable can produce fine denoising results. It is obvious to me that RT currently produces the best results on this photo. It’s also easy to get good results using RT. So I agree with:

But I am certainly looking forward to testing DT 2.8………

It has been interesting to observe that the demosaic settings in RT are so important to obtain a good result and to experience the power of the chrominance curve equalizer in the denoise tool of RT.

It’s also interesting to discover that some noise are highly desirable.

I have a DT- preset to use AMAZE. You have an option to specify “color smoothing” and/or “match greens” using AMAZE but this has not improved denoising in this case.

@cribari has posted an illustrative video on denoising in DT. I think that the idea of denoising each channel separately is interesting.

When you approach DT and RT as a new user, you observe many similarities. RT seems more or less to be a subset of DT so why not join forces and focus on one strong application? Following the discussion and testing the posted examples, I have learned that DT and RT are very different “under the hood”.

…….and I’m impressed by the responsiveness of the Forum. It’s really great, thank you…….


(Morgan Hardwood) #63

False color suppression is not enabled by default in any of the profiles shipped with RawTherapee 5.5.