You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

These weights are redistributed under Meta's SAM License, a copy of which is included in this repository as LICENSE. By requesting access you agree to be bound by it. Note in particular that the license prohibits use for military or warfare purposes, nuclear applications, espionage, and the development or use of guns or illegal weapons, and that your use must comply with applicable Trade Controls. Access is not available in comprehensively sanctioned jurisdictions.

LibreSAM31large

The image path of Meta's SAM 3.1 (detector plus interactive SAM head), converted to the Transformers SAM 3 layout for LibreYOLO's LibreSAM("sam3.1"). Meta publishes SAM 3.1 only as an original-format checkpoint; this snapshot loads with the stock Sam3Model (text prompts) and Sam3TrackerModel (point and box prompts).

License: read this before requesting access

These weights are not MIT and are not covered by LibreYOLO's license. They are a derivative of Meta's SAM 3.1 checkpoint, redistributed under Meta's SAM License, included verbatim as LICENSE. Anyone who redistributes them further must do so under the same terms and include the agreement. The license bars military or warfare use, nuclear applications, espionage, and guns or illegal weapons; use must comply with Trade Controls; reverse engineering is not allowed; publications must acknowledge SAM.

Meta gates the upstream repository with manual approval. This mirror uses Hugging Face's automatic gate: users accept the SAM License and submit the fields above, without manual review.

Source

facebook/sam3.1 at revision daa63191845a41281374e725f4c9e51c7a824460, file sam3.1_multiplex.pt (SHA-256 0567debeec80ba4ac6369540c6c248025283cb3ff2b92827509e57e2b3541cb6)

Modifications

Only the image path is kept. The Object Multiplex video tracker weights are not included: Transformers has no implementation for them.
Keys are renamed to the SAM 3 layout, query/key/value projections are split, the class-token position embedding is dropped, the CLIP text projection is transposed and the point embeddings are concatenated. The mapping is adapted from the Apache-2.0 Transformers SAM 3 converters and reproduces the official facebook/sam3 weights exactly when applied to SAM 3.
SAM 3.1 dropped the fourth (0.5x) FPN level, which the image path never reads. It is filled with zeros so the stock SAM 3 configuration loads.
No learned value was changed. 1637 tensors.
config.json, processor_config.json, the tokenizer files and LICENSE are copied byte-for-byte from the pinned upstream revision.

model.safetensors SHA-256: f83ea65d102023229c1dd0167d5b889f14e42e1990c768270adf67c6b7d6be6c

Usage

from libreyolo import LibreSAM

model = LibreSAM("sam3.1")
model.predict("image.jpg", text="person")
model.predict("image.jpg", points=[500, 375], labels=[1])