@maboleth Don’t know if you already tried this but I found I got better results in RT with luminance noise reduction by setting strength to 100 then moving the details slider to a level with an acceptable level of noise and detail. I’m happy with the results, but might not be good enough for you?
You should try then darktable 3, the new profiled denoise is much improved (including auto modes).
Yes, I do the same. Generally my goal in removing noise is making it less prominent, but also having it uniform across the whole image.
That means I do not opt for selective noise removal, like 100% noise-free sky then 50% noise building. That looks artificial and overprocessed as hell. Or aggressive noise removal with severe artifacts - that looks like a badly compressed JPG phone image quality.
Second that. it was even further improved in 3.0.1 with more control over chroma/luma noise separately.
Relevant video by rawfiner (the developer of the module): https://www.youtube.com/watch?v=7ZhbeXpx2W8&t=2s
Sure, I use NeatImage for Linux - awesome results and really nice upgrade policy too. It is the only non-opensource software I actually use and paid for.
OK I did not read the whole thread BUT noise reduction in both RT and dt is far superior to Adobe and many other commercial noise reduction. You need to know how to use them, but it is not so difficult.
Just a little info: Two days ago, I added a new denoising filter in G’MIC, based on convolutional neural networks, which works quite well.
It’s only available in the development version (3.0.0_pre) for the moment, but you can already test it because binaries are available.
For me profiled Denoise in DT is far better than RT basic tool for keeping details and it can be adjusted but rarely needs to…also from how you handle noise you might like the new diffuse model as you can manage and incorporate grain. There are an excellent series of videos made by rawfiner the author on the various ways to use numerous denoising tools in DT…
You don’t say much about what file type you are using, but you might want to play around with the demosaic algorithms. Particularly try dual demosaic options. They can affect how noise is able to be handled later in the pipeline.
Later reply but I use NeatImage for Linux
There are many noise suppressors on github. Most of them are under MIT, Apache or BSD license.
From Intel there is a noise suppressor,
[GitHub - OpenImageDenoise/oidn: Intel® Open Image Denoise library]
[Intel® Open Image Denoise]
with command line API. This is easy to use, but you have to convert the images to pfm graphics format and little endian.
E.g. ‘convert input.jpg -endian lsb output.pfm’ or with GIMP.
The call then:
$ oidnDenoise --hdr Noisy.pfm -o Denoise.pfm
Have you tested it with photos afaik it was invented only for images rendered with ray tracing and real photo have different noise.
The Readme says Open Image Denoise is meant for noise from rendered images/ray tracing. How well does it work with real world digital camera images?
I had taken a screenshot of an example image and denoised it. This worked as shown.
But the test with an own photo and high iso noise failed.
I guess the photos still need to be trained or are trained on standard noise like Gauss etc…
The program G’Mic [https://gmic.eu/], the filter Denoise works with neural networks. But is very slow …
When looking for an article on dpreview this open source program jumped into my eyes:
Sadly it is only for Apple
If you’ll read the comments, there are links to other implementations of the same algorithm:
HDR+ Pipeline - I believe at this point someone added DNG export, bypassing the tonemapping steps, I am not sure. I hacked TIFF export which I manually tagged to DNG ages ago, but it wasn’t exactly production-ready. HACK: Save aligned/merged Bayer image to TIFF file · Entropy512/hdr-plus@135273b · GitHub
The same author also has GitHub - martin-marek/hdr-plus-pytorch: A PyTorch implementation of HDR+ with CUDA support. - note that it only implements the align-and-merge core, and the “demo” script saves out a postprocessed TIFF, plus also re-quantizes the output to at best 14 bits resolution for most cameras (e.g. white level < 16384) instead of rescaling the output to make better use of int16. A variant of that that exports via PiDNG is on my TODO list - GitHub - schoolpost/PiDNG: Create Adobe DNG RAW files using Python. Works with any Bayer RAW Data including native support for Raspberry Pi cameras. - maaaybe some progress on that next week? But as many on the RT team can tell you, I am a HORRENDOUS procrastinator.
Thanks for the infomation. But noise reduction by multiple exposures has been a common practice for a long time.
This can be done with ImageMagick [https://patdavid.net/2013/05/noise-removal-in-photos-with-median_6/] or Enfuse. There are also some astro programs that work with this method.
I’ve had good results running Topaz Denoise with wine.
With low ISO noise, i.e. real noise and no syntetic noise, good results are also achieved with GMIC (from version 3) and GIMP-ML.
Please see Vergleich-Entrauschen — ImgBB
The photo software digikam has in its image processing the possibility to work with GMIC.
Also for RAW image development
Image → open → enhance → GMIC