BEN2 -- Background Erase Network
Copyright (c) 2025 Prama LLC.

Maxwell Meyer, Jack Spruyt. "BEN: Using Confidence-Guided Matting for
Dichotomous Image Segmentation", arXiv:2501.06230.

This product includes a deployment-only extraction of software developed at
https://github.com/PramaLLC/BEN2, licensed under the MIT License. See LICENSE
for the full text and PROVENANCE.md for what was taken and what was changed.

The Swin Transformer backbone under swin.py originates with Microsoft
(https://github.com/microsoft/Swin-Transformer), also MIT, and reaches this
package through BEN2's own copy of it.

refine_foreground() in image.py is a blur-fusion foreground estimator from
Photoroom (https://github.com/Photoroom/fast-foreground-estimation), credited
in a comment in BEN2.py and reproduced here with the same two-pass radii.

The model weights are licensed separately from this code, and identically: the
model card at https://huggingface.co/PramaLLC/BEN2 declares "mit" and the
repository is ungated. Prama LLC additionally offers a different, unpublished
commercial model by arrangement. That is an offer of another artifact, not a
restriction on this one -- the MIT grant on the published weights is
unconditional, and MIT permits commercial use.

BEN2 was trained on DIS5K together with a proprietary dataset. mozo takes the
publisher's MIT grant on the weights as stated and does not track the terms of
the data behind them; that obligation was the publisher's, discharged before
publication. PROVENANCE.md records where the chain stops.
