I’ve been using RawTherapee as my primary RAW processor for quite a few years, and over the past several days I’ve been putting version 5.13 through its paces. What started as a quick look at the new release gradually turned into a much deeper evaluation.
I had already read that some users were experiencing problems after upgrading, but fortunately I didn’t encounter any issues myself. All of my testing was done on a Lenovo ThinkPad P16s Gen 4 (Intel Core Ultra 9 285H, 64 GB RAM, Intel Arc integrated graphics) running Kubuntu 24.04
Rather than reviewing the release notes, I wanted to focus on how the new version behaves in real-world use. Some of the improvements are subtle at first, but after processing a larger number of images I came away with the feeling that 5.13 is a more significant release than it initially appears.
The wavelet improvements were particularly interesting, and I was also surprised by how much the updated denoising workflow has changed my own editing. In many situations I found that I no longer felt the need to switch to external AI RAW denoising software, something I wasn’t expecting before I started testing.
I wrote a detailed article describing my findings, with examples and screenshots. It’s simply the perspective of a long-time user rather than a developer, and I’d be genuinely interested to hear whether others have reached similar conclusions—or had a completely different experience.
Thank you, Paul. It genuinely is meant as a compliment.
What makes it even more interesting is that, before testing the 5.13 Release Candidate, I had been experimenting with a small Python bridge to integrate darktable’s AI RAW Denoise into my RawTherapee 5.12 workflow. I initially thought AI RAW denoising was one of the few things still missing.
After spending time with the new Wavelet-based Presharpening Denoise, I found myself thinking quite differently. AI still has its place for the most demanding images, but RawTherapee 5.13 has convinced me that I need it far less often than I originally thought. For me, that’s probably the biggest compliment I can give this release.
Thanks for the compliment. I really appreciate it.
Just so you know, I am not ‘opposed’ to AI (I’m an old man…). However, given the current state of AI, it struggles to process something if, for instance, there isn’t an image that more or less corresponds to human vision… Here, by enhancing Capture Sharpening, we are venturing into somewhat uncharted territory at this stage of the process.
I came up with this module because a user remarked that, with the advent of AI, RT was somewhat ‘outdated’ when it came to noise processing.
I set myself the challenge of demonstrating that the human brain still has capabilities. I started with Ingo Weirich’s excellent work and devised a module that was conceptually challenging, particularly due to its position in the workflow.
the second using pre-sharpening noise reduction to enable Capture Sharpening to work, and post-sharpening noise reduction for flat areas Young girl noisy image
Note that these two noise-handling methods for Raw files do not aim to eliminate noise entirely, but rather to address the significant component associated with the Raw data itself.
Note also that I use the ‘non-raw’ version of Capture Sharpening - which I’ve named ‘Capture Deconvolution’ (to avoid confusion) within Selective Editing - to restore some crispness to the image after the raw-stage noise processing.
Noise processing is complex and cannot be reduced to a single solution (whether via traditional programming or AI).
Processing can (and should) be divided into several stages:
Raw: before or after demosaicing… performing it beforehand could be an area for future improvement, as it would allow for the optimization of dual demosaicing methods (e.g., Amaze + VNG4).
before Selective Editing - it is currently set to ‘Noise reduction’.
during Selective Editing : the most comprehensive, regardless of the noise-reduction methods used, because it allows for localized processing and takes into account - albeit with difficulty, and for good reason - the Delta E.
towards the end of the wavelet-level processing, which makes it ‘easy’ to combine sharpening and noise reduction
With several noise processing tools:
median
wavelets - either by working in RGB or Lab mode, depending on the case - The decomposition levels, the decomposition method (Daubechies, etc.), and the vanishing moment are important elements.
Fourier with DCT
non-local means
Guided Filter
etc.
With or without additional tools to adjust sharpness
It is a combination of these methods - in terms of their position in the process and the type of tool used - that should enable effective processing.
My stance on AI:
If it involves working “after the fact” - whether at the end of the workflow or on TIFF or JPG files - I am not interested from a developer’s perspective, even though I acknowledge that it may offer some value from a user’s perspective.
If AI is capable of intervening before demosaicing - right in the pixel soup, where there is no reference image - then I’m all for it.
to find new ‘blind’ demosaicing methods, whether combined or not.
finding new pre-demosaicing noise processing methods, whether or not combined with the demosaicing method.
Even though it might seem pretentious - especially coming from an old man like me - I cannot bring myself to use an ‘off-the-shelf’ solution, such as a patch added to RawTherapee. My goal is always to understand and find a solution adapted to the context, rather than simply copy or duplicate what can be found elsewhere.