These are instructions that I followed. And to create that mask in the first picture is started was already with multiple 'click’s. It was then I clicked the small branch the mask changed unexpectedly.
It’s not a bug, and I wouldn’t call it limitation as well. You just need to keep in mind that fundamentally it is a single object selection tool. This object must be visually continuos.
In your case, even if we know that small branch is a part of the tree, it is visually separate. So it is not possible to select both of them in one try. But that is not a problem actually. You just run mask tool one more time and group results.
Thank you for your time and the effort you’ve put into the module.
When the AI function is activated, the image is ‘processed’; this currently runs on the CPU, with MIGraphX selected in the settings. The latest ROCm version also runs without any issues.
MIGraphX is also installed.
Edit: Arch Linux, ryzen ai 395 max with 8060S iGPU
@anry …
Hi just for the sake of interest , why have you chosen the specific SAM2 model ?
SAM 2 offers other models … are there any differenecs between the various models ?
We currently have two segmentation architectures available: SAM 2.1 and SegNext.
SAM 2.1 comes in four sizes (Tiny, Small, Base Plus, Large) that share the same architecture but differ in encoder capacity - larger models produce more accurate masks but are slower. We included Tiny, Small, and Base Plus. Tiny and Base Plus can be manually downloaded from the repository and installed in AI preferences tab. We skipped Large for now since its encoder is ~900 MB and too heavy for interactive use.
I will make a separate post about how to use AI object masking tool. Meanwhile, you can either download nightly build for Windows or macOS, or build darktable with --enable-ai option. After that you need to enable AI features in preferences and download AI models.
Small variant is available in AI preference tab by default.
Tiny can be downloaded manually from Release v5.5.0.4 · darktable-org/darktable-ai · GitHub
Base plus variant will be available in the mentioned repository soon.