Call to advice for fighting color noise

That is not really how it works. The RPU samples are used for regression testing of rawspeed library.
For the camera support, one needs to open a new camera support issue. Just uploading samples to the RPU, while is immensely helpful, is not enough to get the camera supported.

That being said, please do keep submitting samples to RPU, there are still a lot of missing samples.
All new Samsung’s (NX3000+) are missing.

Thanks, Roman. Will comply as instructed.

Here is an example (zoom 100%):

instance 0; wavelet; 100%; strength 1; on the color:

first one plus
instance 1; wavelet; 50%; strength 1; on the lightness:

first two plus
instance 2; non-local means; 50%; strength 1; on the averaging:


Uhmmmmm! The result is lovely! Where did you learn to do that? :slight_smile:

Here: I think the author of the video found the info in the searchable email archive: darktable-dev


RAW of that particular picture I sent you the cropped area of as well as its darktable XMP file are posted here:

You can play with the denoise profiled instances and see how they work.

1 Like

Really nice results, I have a new noise reduction preset :slight_smile:

@Andrius Tried your method today and the results are amazing, even though I had no dedicated camera profile and used generic poissonian preset instead. Excuse the hastily applied mask to focus on the sky, but the results speak for themself!!!

1 Like

@mosaster I am glad it worked for you.
I wish I could take a credit for this but I just copied somebody’s method.
I guess darktable devs have done a great job :slight_smile:


To smoothen out, I like to put an instance on top of the stack using average blending, which just gives it the final touch.

…aaaand of course, YMMV. :slight_smile:

I saved the three instances as a darktable style and uploaded it here:


Wow, that worked great for me // saved them as presets // used them on other images that I was having trouble with and so far, really good results.

Thanks for the NR presets. I’d modify them according to what houz suggests in this text:
“For profiled denoising better use ‘HSV color’ instead of ‘color’ and ‘HSV lightness’ instead of ‘lightness’”.

Thank you for the comment.
I followed the link and finally got here:

I am wondering what “clamping” is

@Andrius See Clamping (graphics) - Wikipedia

1 Like

Thank you afre

This is awesome!
My only real issue with DT was terrible noise reduction. Turns out it was just me being ignorant and didn’t understand the tool.
Thank god for that lol, I can now ditch Lightroom with confidence.

1 Like

Some time ago I profiled my Fujifilm X-Pro 2 in accordance with the dt procedure and was surprised at the good (excellent) results obtained.
When I use either the ‘2-step’ or the ‘3-step’ (mentioned above ) option I see absolutely no difference from that produced from my single step usage.
Possibly it is my camera or the individual profiling … but I do like my simple results.

Probably I should have also mentioned … the X-Pro 2 is already well corrected for noise at ISO 800 by Fujifilm. Noise at that setting is not so different from the base200 setting. I do use that ISO 800 setting for 95% of all my imagery.

A sidenote here : there is no such thing as chroma or luma or colour noise. It’s just noise. How it looks, once it enters our perceptual system, doesn’t change what it is : garbled data on RGB channels.

Similarly, the denoise profile module in darktable doesn’t know any colour or luminance since it operates in camera RGB. That camera RGB has nothing to do with our perceptual system, and will need to meet a colour profile to be remapped to XYZ and then to well-behaved RGB spaces, where chrominance and luminance have a meaning.

Blending the denoise profile in lightness or colour is therefore a theoritical mistake since these are Lab modes. HSV lightness or colour are equally wrong since the conversion RGB → HSV assumes well-behaved RGB spaces, while the camera RGB can be absolutely whatever.

The last changes in denoise profile module (in future darktable 3.0) should make these blending modes irrelevant, since they essentially hide misbehaviours of the denoising algo.