I use RT/Gimp. I often ask myself if my denoising results in RT could be better. Don’t get me wrong, I’m pretty happy with it, but maybe there’s other/additional software doing better.
I’m a bit unsure how free software compares to payware when reducing noise. Is there a freeware that uses “AI” denoising? This is said to deliver the best results (like in Topaz, but it is 200 bucks).
Anyway, if there is an easy way to enhance my denoising using freeware I’ll gladly go that route.
Have you added G’MIC to the GIMP? Loads of noise reduction options …
I find Iain’s Denoise in G’MIC does a pretty good job. The following video is a bit out of date but it explains the basic principles. https://www.youtube.com/watch?v=pPj_7J4iD_U
Other options are the Blur/Grain & Denoise module in RawTherapee’s Selective Editing, which is great for noisy skies and the like. For a semi-automated solution, the free version of Neat Image is worth trying.
This (Iain’s Denoise) seems to be very powerful. But it probably needs some experience to use it appropriately.
I am a dedicated darktable user. It has many sharpening and denoising options that meet my needs very well. I shoot a lot of 32000 ISO images on my Canon R7. I also like that I can use a details slider in DT to limit the sharpening to details only which reduces noise in smoother parts of the image such as the sky. I can also use this same details slider to denoise the smoother parts of the image while protecting the details so the image doesn’t just become a soft blur as I denoise. I can also use drawn masks to localise sharpening and denoising. Sometimes it might be worth using a parametric mask to exclude highlights from noise correction as noise predominately occurs in the shadows. This masking approach is probably more suited to luma noise correction which can compromise detail.
There is no free option that compares to DxO or Topaz, in my opinion. But those programs can also interpolate and add detail to gain their results. DT is okay but I think RT is a bit better in some regards, especially with regards to impulse noise. I’ve not tried G’IMC but it looks interesting too…
This is good to know. So maybe I might as well stay with RT as the difference to other freeware is marginal.
If you’re interested in AI style stuff, check out chainner: GitHub - chaiNNer-org/chaiNNer: A node-based image processing GUI aimed at making chaining image processing tasks easy and customizable. Born as an AI upscaling application, chaiNNer has grown into an extremely flexible and powerful programmatic image processing application.
I’ve played with chaiNNer. I think the models for denoising produce overly smooth images, which can be used as overlays and mixed with the original. But then I had the same feeling about DxO – it’s very smooth, but looks artificial to me…
Its all taste. For me, as long as I can mitigate chroma noise, I’m good.
I’m working on some older 6400 ISO RAWs from my Canon 1D Mark III back in 2008. I have an image for reference and would really like to see how you denoise it with your freeware tools.
Would it be appropriate to post a link to the RAW file in this thread? And would some of you even want to give it a try?
Of course. But better offer it in the Category PLAY RAW.
@kaha : See here:
A minor note: the software this site supports are not simply freeware – they are not only ‘free as in beer’, but ‘free as in speech’ (libre / free and open source software, FOSS). People don’t just get permission to use them for free, but are also allowed to study them, take them apart, build modifications (of course not everyone has to do that), and publish those. For the average user this means that the developers of darktable (or any other tool here) can’t simply decide one day they go shareware or commercial, or simply stop developing, without giving others the chance and right to continue if they are interested.
Thank you.
I think that the denoising algorithms in Darktable are near the limit what can be achieved from sensor information. They works fine up to and maybe a bit above ISO 25600–32000 / crop factor² on “modern” (= from the last 10 years) sensors.
ML-style “denoising” is actually regenerating the image in a way that is somewhat compatible with the original and matches what human vision knows about the world, eg if it sees a leaf or a feather, it will generate the typical texture of that leaf or feather in a way that fits the image.
This of course can look spectacular and very convincing, but it is, strictly speaking, no longer “denoising”. (Some may say that possibly not even photography 
).
If it looks like the image just with less noise, it’s just another algorithm practically speaking. People who defend that idea should go back to analog or even silver plates to get as close to reality as possible 
There are many examples of faces or other things being generated out of nothing by ai noise reduction. (Ryan Gosling anyone?)
The idea that because an image isn’t absolute truth * any level ot truth is meaningless is extremely simplistic thinking. It’s surprisingly often brought up in photography related discussions. It’s just that the line of argument is a total dead end for all human knowledge.
/* Impossible
The problem is that it may not. ML engines are getting better and better, so outright “hallucination” is becoming quite rare (at least for noise reduction), I think that they will happily impute texture and perceptual sharpness when it is in fact not there.
If someone is interested in doing it, it would be great to shoot pairs of images in low light, with the same aperture and focus locked, on tripod, but one with noise (short exposure / high ISO) and a less noise (long exposure / low ISO, “ground truth”). Then run various engines on the first of the pair and compare.
I wonder if this is also how these models are trained. With series of the same scene with increasing noise levels.
The face was generated by an upscaler though, are there cases of similar problems with noise reduction?