Immich is a good example of a FOSS project that trained their own models using open data. From what I’ve gathered they were mostly trained on open datasets which require attribution, so believe somewhere in Immich there may be a list of all the attributions.
Either way, it’s a great example on how to do things correctly whilst respecting people’s rights and remaining fully FOSS. They support both category tags(cat, dog, landscape) and facial recognition. It works great and they are often praised by the quality of the tagger.
Its ridiculous that you’d ask obviously rhetorical questions instead of dealing with the point I was making, but since you did, I have a photo I exported in 2021 with Darktable 3.6.0 and I opened it today, made no changes, and exported it again with Darktable 5.2.1 and guess what? The exported image is very different from the original. I’m sure I could go through the modules to figure out why, but that’s not the point – the point is that there’s no reason to expect future versions to render identically to historic versions.
If nothing else, I expect colour science to improve (and it has) and the underlying algorithms to become more accurate (and some have) and bugs in modules to be fixed (and they frequently are).
Deciding whether or not to implement a feature should not be dependent on whether it will be the same in the future.
It would have helped if you read the whole the message you quote part of. The rest of that message made the question a lot less rethorical, as @Pascal_Obry made it quite clear that he will not let modules disappear or change implementation (without proper versioning code, that is).
And if you see a difference between exports from different versions, that might be worth a bug report But of course, then you will have to figure out if the history stacks, and the export settings, are indeed the same.
I have had some instrances where old files misbehaved. Not sure what happened with them, but as I opened them to redo the edits anyway…
Darktable always worked very hard to keep old edits valid, and that worked in most cases. So whether or not a feature will give the same result in the future is a big issue, especially when you have to deal with external libraries or datasets.
In the end, it’s very simple: for any feature, someone needs to code the module and someone needs to maintain it afterwards. If no one wants to implement a feature, it’s not going to happen. If the maintainers and current devs aren’t interested, it’s not going to happen. So far, nobody seems interested in implementing AI features of the kind suggested here… Draw your own conclusions.
If you really insist on having AI features: darktable is open source, so you are allowed to fork it and do whatever you want with the fork (cf. Ansel)
That sounds a sensible approach for anyone so inclined to introduce AI. I am not against AI being introduced but valid problems have been brought up with putting it into the master. I look forward to one of the pro-AI users implementing it in a fork. Hopefully some useful tools will come from it.
I am now retired, but between 1980 and 2000 I worked as a researcher in the field of medical informatics (focusing on knowledge processing based on ontologies and some image processing applications for medical purposes). The recent advances in AI based on neural networks did not come as a complete surprise to me.
I don’t want to go into the impact of AI on our society, but rather focus on image processing. Over the last 12 to 18 months, AI-based image processing has developed to a level that far exceeds my own capabilities in the field of image processing. I’m not talking about noise reduction or other “simple” features (which I would welcome with open arms).
Attached to this post are two examples that, in my opinion, illustrate the problem we amateur photographers are facing (for professional photographers, this is a completely different issue, as I assume that many of them will have to cease their activities in the coming months or years).
Example 1:
The first photo shows a place I recently visited with a friend in Shiga Prefecture, Japan. To my taste, the photo would have been unusable because of the cars and people cluttering the image. With Gemini, it took me less than a minute to create the second photo. Manual editing would have taken me about an hour.
Example 2:
The first photo shows a scene I took close to my home. I found it rather boring and tried to improve it (which took me 15 minutes). The result is the second photo (your taste may differ). The third photo was generated by AI (in less than a minute). On closer inspection, I have to admit that the lighting looks more natural (e.g., reflections on the top and base of the stone lantern) and more interesting.
I assume that we are facing some interesting challenges. I am currently starting a discussion with Japanese photographer friends on the question of whether the use of AI tools can be part of our toolbox or whether we should refrain from using them (linked to the question of what is generally permissible in image processing and what is not).
The fundamental question is: What is the point for me to do something that a machine (AI system) can do better and faster than I can?
That’s probably the same question painters asked when photography arrived…
AI is a tool, and as such everyone will have to decide whether to use is or not.
Where we get real issues is in presentation of such generated or manipulated images as depictions of actual situations or events. I.e. used to mislead/misinform. But even that isn’t new, or restricted to manipulated images. AI just seems to make it so much easier to lie with images.
If you go this way, sell your camera and just generate images using AI. No need to take pictures, just generate them. I’m not against this, but I think those are different approach as generating image is not about photography.
Definitely yes. And the problem is that even for well educated and trained people it will become more difficult to distinguish between reality and generated content.
Pascal, I think you missed my point (which is of course my fault). I didn’t generate images. I used AI tools (hesitantly) instead of the usual tools to process photos I have taken. Is there a point where we should stop? Using the example of the picture with the alley. Is the removal of the people and the cars OK? If yes, doing it in Darktable and GIMP takes me roughly one hour. With AI it takes a minute an the result is better than everything I can do with my usual tools.
I would say, let machine learning do all the things that you do not enjoy doing (in my case, e.g., noise removal, masking) as long as it can do it faster/better than you would.
Keep for yourself the part that you do enjoy. For me, the fun part is shooting. I also like recomposing in post, color grading, styling, b&w conversions and what not.
Do not take the picture if there is too many people
Come back to the place another day/time
Sure, but what the point? That was the point in my first answer. Why changing so many things in the picture, is that still a picture? Of course we all have a different limit on this.
This applies to every edit that we do, doesn’t it? Color grading, b&w conversion… what you get in the end is no longer a faithful representation of what you saw.
The limit depends on how you sell the result. It’s perfectly fine to take a photo and process it heavily until it becomes something completely different, as long as you make it clear that this is a product of art and not a snapshot of reality.
People were doing it also before GANs and CNNs, the only difference is that it required skills and time, which made heavily redacted photos comparatively rare.
Actually, the fact that “fake” is now pervasive will have at least two positive effects IMHO: people are less inclined to assume that something is true just because there is a photo/video of it, and the demand for real, genuine photography from trusted individual and organizations is going to grow.
