Yep. Not only that but they are also behind the most recent push for age restrictions everywhere: https://tboteproject.com/.
They are allowed, as no one can prevent them 
1 Like
Does Meta’s contribution to the Linux kernel make Linux itself morally questionable?
2 Likes
Looking forward to giving it a go Andrii 
1 Like
This is a strawmen, it’s a completely different paradigm people having problem using meta’s products directly and using something they have had their hands in.
1 Like
Absolutely agree with you. And SAM models are not products of Meta. They are open-source projects contributed by Meta, the same way as many other open-source projects including Linux kernel.
SAM models indeed have some questions related to the origin of training data. That makes SAM models open-weight ones, not completely open-source in terms of understanding of open-source AI. It is still compliment with GPLv3. And this is exactly why we have alternative model for AI object mask - SegNext. It is a bit worse model, but still very capable. And it can be classified as open-source AI model. Every user can make there own choice.
8 Likes
just wanted to share that this works fast and flawless on my system (ubuntu 25.10, excecution on CPU [Ryzen 7 9700X]). For me, this is a very important addition to masks in darktable - many thanks!
1 Like
Though by the same token, they’re free to draw their own lines.
Which one? I want to know.
I should probably clarify that both models take input image in dimensions of 1024x1024. So image features are encoded at this resolution. Mask output is a bit smaller though and further upscaled.
1 Like
I hadn’t realised they were so limited. Are the models used in commercial software as compromised? What about the flow Corridor crew used for their green screen removal project (which I realise requires a green screen)?
CorridorKey
They are not limited. On the contrary, they are pretty capable models. Don’t bother about numbers, just use what they can offer.
The same models are used in commercial software. They can possibly be further fine-tuned for a specific use case, but the architecture stays the same.
Very few very lightweight computer vision models can operate on full resolution image. In most cases we use either downscaling or tiling. That is common practice.
On their GitHub I see:
- Resolution Independent: The engine dynamically scales inference to handle 4K plates while predicting using its native 2048x2048 high-fidelity backbone.
So, they downscale too. Different working resolution, but the concept is the same.
This is Glorious!
I wasn’t too worried about not being able to get a dt version with ai masks, becaue I wanted to see the new toy, but didn’t think I’d be an immediate user.
Well… I absolutely am! 
With feathering, I never found it hard to, eg, mask a person: absolutely no need for hundreds of clicks closely following an edge. But two or three clicks to get a much better mask totally wins that race!
Thank you very much to everyone who contributed to making this new functionality.
2 Likes
Worked great on CachyOS. Had to install onnxruntime-cuda specifically and then rebuild to get my Nvidia GPU to work, but the mask worked exactly as one would imagine it would, thanks! Great work.
1 Like
Works like a charm on my MacOS computers. But I cant get it to work with my Linux laptop. But thats a skill issue for sure… Linux noobie…
I couldn’t get the appimage to work properly. I had to self-build. And, somehow, I couldn’t get a working build from a tar source file. I had to use git, and I have zero experience of git commands and ways. Without hand-holding, I would have no idea what to clone pull, synch, etc. There’s a script earlier in the thread, which I adapted slightly. Give it a try!
I was about to ask why the paint cursor was missing on the Mac build!
Thanks for the update.
Fantastically useful tool. Thanks so much 
What specific problems with AppImage do you have?
1 Like