LibreBEN2b-matte

BEN2 Base background removal, repackaged for LibreYOLO's matte task. It predicts a soft alpha matte at a fixed native 1024x1024 resolution.

from libreyolo import LibreYOLO

model = LibreYOLO("LibreBEN2b-matte.pt")
result = model.predict("product.jpg")
result[0].matte.array      # (H, W) float alpha in [0, 1]
result[0].save("cut.png")  # transparent-background PNG

Source

Derived from PramaLLC/BEN2 at commit 2c99a5da477b5523585bfa5c893888a6e818a8f6, using the released checkpoint from PramaLLC/BEN2 at revision e48a20765fb421d19dcdb0bf3cc61e802ca5ec8f. Copyright (c) 2025 Prama LLC. Licensed under the MIT License.

Training data provenance (upstream): DIS5K and Prama LLC's proprietary 22K segmentation dataset. This repository redistributes only the released MIT checkpoint and does not redistribute training data.

Modifications

State-dict metadata wrap only. Learned parameters are unchanged. The native LibreYOLO fp32 forward matches the released BEN2 Base network with max_abs_diff == 0 for batch sizes 1 and 2. See weights/convert_ben2_weights.py in the LibreYOLO source repository.

The port returns raw logits for LibreYOLO's shared matte postprocessing and does not include BEN2's optional media or foreground-refinement helpers.

License

MIT License. See the LICENSE and NOTICE files.