I've been working on background removal models for a while (previously released Snap). Focus is a major improvement, especially for hair, fur, and complex edges.

It's fully open source (Apache 2.0) and runs locally. I also run a paid API to sustain development, but the open source model is complete and functional on its own.

Deliberately unfiltered results (you'll see where it fails): https://withoutbg.com/resources/background-removal-results/m...

Quick start:

  pip install withoutbg

Or run the Docker web UI:

  docker run -p 80:80 withoutbg/app:latest

The model used mean gradient error in training, heavily penalizing missed edges. This produces sharper results but occasionally causes false positives in foreground-background classification. Would appreciate technical feedback on this tradeoff.

GitHub: https://github.com/withoutbg/withoutbg Docs: https://withoutbg.com/documentation/integrations/python-sdk