Thank you for this work and this information. It is very enriching and this approach must be known and valued.
I am not (at all) an expert in artificial intelligence, even though it interests me.
I’m speaking for myself; this may not be the opinion of others… For your information and reminder, I’m sick and old (77 years old), an old Frenchman who discovered computers and C++ upon retirement.
As its name suggests, Rawtherapee is primarily a Raw processing software… which poses many problems when evaluating the image, since there is no native image. I won’t list the entire process, but will focus on the key points that would be affected by AI (of course theirs others…)
- Which demosaicing software to choose, particularly based on noise.
- How to optimize Capture Sharpening – which will basically improve the image’s edges if it is noisy.
- How to denoise (where to evaluate noise) – based on what criteria, with what tools to reduce noise where needed and not where needed.
- How to restore sharper edges to the image after denoising without excessive effects.
- How to differentiate actions depending on the image – for example, highlight the main subject and blur/soften less important areas.
- How to evaluate white balance
- How to evaluate image dynamics (DR), black point and white point values, and which algorithm(s) to use from all those available (Log encoding, Sigmoid, TRC, GHS, etc., etc.)
- How to evaluate the user’s wishes for the final rendering: contrast, saturation, color fidelity, etc.
- How to evaluate possible cropping and deletions, etc.
- Etc.
When I was younger (1993), I worked on the cognitive aspect of expert systems. What makes Mr. X a recognized expert? At the time, it wasn’t about making an exhaustive inventory of his knowledge and skills, but rather trying to model (this is very close to AI) how he goes about it, how he approaches the subject (in fact, how to tell AI to approach the subject). In France, some people call this REX (Retour d’Expérience), which is different from “Feedback” because the key is to identify what makes him an expert and then allow other users (or AI) to learn from this approach.
In March 2024, I tried this, but I think it was misunderstood - my English is poor, and without exchanges and discussions with real people, misunderstandings quickly lead to confusion. You can find the (incomplete) summary of this approach on Rawpedia.
Rawtherapee processing challenge
I think that, as Bill Gates rightly says, there is one area - at least in the near future - where AI will not replace humans: algorithms.
What I wrote above is just algorithms: what to do, what to decide, how to guide, etc.
For those interested, I’ve opened a “captur-noise” branch that covers some of the points mentioned above (Capture Sharpening Raw, Denoise, Capture Sharpening after, etc.).
Developing AI in open source software will face major challenges, because to create AI, you need three things:
- one or more algorithms (the key of the system)
- a computer that knows how to perform/optimize parallelization. As a reminder, Rawtherapee doesn’t use the GPU but optimizes the CPU based on the number of cores.
- a knowledge base (data), which we obviously don’t have.
Thank you for reading, and excuse my bad english.
Jacques
AI-Powered Analysis: Leverages AI to analyze your RAW photos and generate optimized PP3 profiles.
Multi-Format Support: Works with popular RAW formats like DNG, NEF, CR2, and ARW.
Seamless RawTherapee Integration: Automatically processes images or generates PP3 files for manual fine-tuning.
Flexible AI Providers: Supports OpenAI, Anthropic, Google, and xAI, so you can choose the best AI for your needs.
Simple CLI Interface: Easy to use with straightforward commands and options.
Anyways, a lot of stuff like this will be in the document, I just wanted an “AI” tool I can run on a software I know very well, to show an accurate comparison, just by processing the same image manually, and comparing the results.