---
license: apache-2.0
library_name: transformers
pipeline_tag: mask-generation
tags:
- libreyolo
- edgetam
- promptable-segmentation
- image-segmentation
base_model: facebook/EdgeTAM
---

# LibreEdgeTAM

This is LibreYOLO's Transformers-compatible mirror of the official EdgeTAM
checkpoint. The raw pickle-based `.pt` file is deliberately not included.

## Source

- Official checkpoint: [`facebook/EdgeTAM`](https://huggingface.co/facebook/EdgeTAM)
  at revision `14d7ecc48c656b94e5184519f698cd5386c5a2bf`
- `edgetam.pt` SHA-256: `ed2d4850b8792c239689b043c47046ec239b6e808a3d9b6ae676c803fd8780df`
- Official source: [https://github.com/facebookresearch/EdgeTAM](https://github.com/facebookresearch/EdgeTAM) at commit
  `7711e012a30a2402c4eaab637bdb00a521302c91`
- Reference Transformers snapshot: [`yonigozlan/EdgeTAM-hf`](https://huggingface.co/yonigozlan/EdgeTAM-hf)
  at revision `c266ce53b3fc00f0f495b583f6a116c4e57f53bb`; its model card declares Apache-2.0
- Conversion implementation: [`huggingface/transformers`](https://github.com/huggingface/transformers) at
  commit `bd37c453544e83eb875ed3608980a1660376007a`, file `src/transformers/models/edgetam_video/convert_edgetam_video_to_hf.py`

## Modifications

LibreYOLO independently converted the model weights from the safely loaded
official checkpoint using the pinned Apache-2.0 Transformers conversion (key
remapping, lossless key/value tensor splitting, and point-embedding
concatenation). It strict-loaded `EdgeTamVideoModel` and checked all
984 resulting tensors for exact equality with the pinned reference.
No learned numeric parameter was changed, and `model.safetensors` was not copied
from the reference.

The non-weight `.gitattributes`, `config.json`, `preprocessor_config.json`,
`processor_config.json`, and `video_preprocessor_config.json` files are copied
byte-for-byte from the hash-pinned reference revision above. Every copied file
is SHA-256 verified. The reference repository declares Apache-2.0 in its model
card.

Generated `model.safetensors` SHA-256: `8858f8e4757b0b96dab8763f296ecffd845efbbbf698f64163cfa20a63d5fff4`.

## Usage

```python
from transformers import AutoModel

model = AutoModel.from_pretrained("LibreYOLO/LibreEdgeTAM", trust_remote_code=False)
```

LibreYOLO users can load it through `LibreEdgeTAM` once the EdgeTAM integration is
installed.

## License

EdgeTAM code and model checkpoints are licensed under Apache License 2.0. The
verbatim upstream license is included in `LICENSE`; provenance and modification
details are included in `NOTICE`.