Perhaps recursively running the masking on patches of the image after first using a whole image mask to define the area to work on?
I feel like there should be some cleverness possible, like pixel shifting the 256×256 input to the model (by ± width,height÷256÷2 px) and seeing where it differs from the originally generally mask. But stochasticicity will prevent this from working since the model won’t generate the same output each time…
To be clear, are you saying the models are trained on one resolution and can’t do higher resolutions? Or does that just kill performance?
Even if the curves don’t hug edges finely, they’re still useful placeholders that can be manually adjusted and feathered. This will be a big time saver on large objects! Can the generated path be scaled? That’s sometimes necessary for feathering to work optimally.
Yes
Correct. Those optimizations are already in place as post-processing step.
1 Like
Interesting idea. Unfortunately, not feasible. It requires re-encoding image on a cropped patch. Re-encoding a different crop produces completely different embeddings - it’s a new segmentation, not a refinement. Results are mostly unpredictable, not an improvement at all. This approach is fundamentally flawed.
This is model with task “mask”, which is used by AI object mask tool.
Lot’s of info here if you are curious
Also if you are curious… several plugin’s from this YT author for integrating Sam2 Sam3 and some denoising and maybe some upscaling??
Should both items listed be active?
No, only one model of the same task can be active.
2 Likes
@Andy_Astbury1 : Now I get what you are talking about. Looking at it.
1 Like
How are these things being done in special effects pipelines?
As I understand it, one AI model is from META, doesn’t that raise any moral questions?
1 Like
Maybe, can you be more specific which moral questions you have with that model from META?
If it really an issue maybe we should reconsider that model?
Fix is on the way.
2 Likes
As a tech person, I can call this company (Meta) and its products one of the worst for people in terms of privacy and data collection, as well as a place for malicious ads and bots. So I have moral doubts about whether Darktable really needs it.
1 Like
Well, exactly for that you can choose one of the alternative segmentation models …
4 Likes
Exactly. It’s on us to stand by our principles. The maintainers of software tools should not be burdened with the responsibility to make ethical choices on our behalf, nor should they force their ethical values on the users.
4 Likes
Would echo that … to 100 %
Maybe not a maintainer group but a BDFL should 100% be allowed to do that.