Andrii Ryzhkov
Andrii Ryzhkov
Have the same problem.
Duplicate of https://github.com/darktable-org/darktable/issues/12295
@MikoMikarro I revied you pull request (including previous) and also playing with similar implementation on my own in parallel. I see you injecting SAM model inference in `pixelpipe`. It means...
@jenshannoschwalm what do you mean by "chosen segment" of the mask? Just to align on how SAM model works: There are two modes of SAM inference: 1. Based on prompts....
@MikoMikarro that is not the optimal use of the SAM. Points as a prompt are used to compute mask. Model indeed computes multiple masks - typically 3 with different confidence...
I agree that's specific implementation of the model is not important right now. What important is the interface between AI model and darktable. And it should not be tightly couples...
@MikoMikarro you are absolutely right, SAM has automated mask generation mode which produces multiple masks. But is this mode model deciding what your objects will look like, and user can...
@TurboGit, @jenshannoschwalm can we create a dedicated branch for AI object mask so we all can contribute?
I agree that we need to focus on darktable part now, not AI one. But it is essential to have high-level understanding and alignment on how model works so we...
I agree with you regarding embeddings. Maybe we can think on how to serialize them and save. Regarding your approach, well, you can definitely try if you insist on it....