sam2-lite
Copyright 2026 pablopigue

sam2-lite is an independent project; it is not affiliated with, endorsed by or sponsored by
Meta. "SAM 2" refers to the original model by Meta FAIR, on which this work is based.

The parts of sam2-lite are under different licenses:

1. Source code: Apache License 2.0, except for the files listed in 2 and 3. A copy of the
   license is included: LICENSE in the source repository, LICENSE-APACHE-2.0 in the model
   repository.

2. src/sam2lite/eval/third_party/sav_benchmark.py: SAM 2's J&F evaluator, vendored unmodified
   from https://github.com/facebookresearch/sam2, with its licenses in the
   same folder:
     LICENSE                BSD 3-Clause License, SAM 2 Eval software (Meta Platforms, Inc.)
     LICENSE_DAVIS          BSD 3-Clause License, DAVIS (davis2017-evaluation)
     LICENSE_VOS_BENCHMARK  MIT License, Rex Cheng (vos-benchmark)

3. Media, not under the Apache License: the example videos (app/examples/*.mp4) and the README
   animations (assets/demo.gif, made from swans.mp4; assets/hard_case.gif, made from
   train-viaduct_teacher_vs_student.mp4) are clips of Wikimedia Commons videos licensed under
   CC BY 2.0 and CC BY 3.0: cut, re-encoded and, where noted, with masks drawn. Titles, authors,
   sources and licenses: app/examples/ATTRIBUTION.md.

4. Models (https://huggingface.co/Pablopigue/sam2-lite): Creative Commons
   Attribution-NonCommercial 4.0 International (CC BY-NC 4.0),
   https://creativecommons.org/licenses/by-nc/4.0/. For non-commercial research use only,
   because of the data in 5 and 6. They contain:
   - SAM 2.1 Hiera-tiny weights by Meta Platforms, Inc., Apache License 2.0,
     https://github.com/facebookresearch/sam2. Modified: the memory attention (fine-tuned to
     attend to 3 memory frames instead of 7) and the memory temporal encodings (3 of the 7
     kept). The other SAM 2.1 weights (memory encoder, prompt encoder, mask decoder, object
     pointers) are included unchanged and remain available under the Apache License 2.0.
   - timm weights timm/mobilenetv4_conv_medium.e500_r224_in1k (Ross Wightman), Apache License
     2.0, https://huggingface.co/timm/mobilenetv4_conv_medium.e500_r224_in1k. Modified: the
     image encoder starts from them and was trained by distillation.

5. ImageNet-1k: the timm weights in 4 were trained on it, and its terms of access allow only
   non-commercial research and educational use (https://image-net.org/download.php).

6. DAVIS 2017 (https://davischallenge.org; Pont-Tuset et al., 2017; Perazzi et al., 2016):
   CC BY-NC 4.0, as stated in the Terms of Use of the official dataset package. About half of
   its sequences (31 of 60 in train, 15 of 30 in val) come from third-party online sources,
   mostly YouTube, with their own terms of use (listed in the package's SOURCES.md). The models
   were distilled on DAVIS 2017 train frames and evaluated on DAVIS 2017 val. No DAVIS videos,
   frames or annotations are redistributed.