I thought at first the library would just call out to some web API to do the actual magic, but seems it's actually running the model locally, which is very nice to see.
So it seems there is a bundled models (bundle/models/{small,medium,large}) but I don't see the code responsible for creating those. Anyone knows where the source is located? It's the final piece of the puzzle to figure out if this thing is FOSS or not.
Hmm, I see. That makes the "Open-Source" in the title a bit hard to understand. It is just the library that sits between the developer and the usage of the model that is open source, rather than that the "background removal" part.
How good is the model? I would have expected some kind of hint from the user to help it know what the background is. Something like a selection of where the subject is or what the subject is, at least as an optional fallback.