README.md

---
license: apache-2.0
pipeline_tag: image-segmentation
library_name: onnx
tags:
  - anylearning
  - onnx
  - onnxruntime
  - segment-anything
  - segment-anything-2
  - segment-anything-2.1
  - image-segmentation
  - object-detection
  - instance-segmentation
  - promptable-segmentation
  - edge-ai
authors:
  - Neural Research Lab
  - Viet-Anh Nguyen
---

# AnyLearning labeling models

Versioned ONNX model bundles used by
[AnyLearning](https://github.com/nrl-ai/anylearning-oss) for local object
detection, instance segmentation, and prompt-guided image segmentation.

This repository contains the seventeen models currently offered by AnyLearning:

| File | Model | Download size | SHA-256 |
| --- | --- | ---: | --- |
| `mobile_sam_20230629.zip` | MobileSAM | 36,655,105 bytes | `41aff2660b7531becfee21fb257c49933ddc892c554507bdb775bf504d443942` |
| `efficientvit_sam_l0.zip` | EfficientViT-SAM-L0 | 129,322,579 bytes | `ef48c8f4c6d72da3df0690d9f16aa794035978d23a637acbd24a984220020072` |
| `efficientvit_sam_l1.zip` | EfficientViT-SAM-L1 | 177,189,481 bytes | `e7f6a5a38a00b53c6706ea45e70c5601b2140c57729ad6412c0600fc719634a2` |
| `efficientvit_sam_l2.zip` | EfficientViT-SAM-L2 | 228,296,139 bytes | `be208e948e445b2d3743d17c2cab83564a403b698e90c9cbdfb1ea6273ab823e` |
| `efficientvit_sam_xl0.zip` | EfficientViT-SAM-XL0 | 434,729,250 bytes | `98ccd8ca3a4ff339cdaa29f6f8ff6dad259516f38e7fb093c8b5cbd575d604ea` |
| `efficientvit_sam_xl1.zip` | EfficientViT-SAM-XL1 | 756,188,273 bytes | `316c92d1900b09e102cfbb3585f8da25b6df20c6c6fe0f0ef6552ec3fc0aab6d` |
| `sam2_hiera_tiny.zip` | SAM 2 Hiera-Tiny | 143,293,170 bytes | `f96e764cf19ba63e7870f8d588cfbfe772cda40a13180f19e73100e28f8c92da` |
| `sam2_hiera_small.zip` | SAM 2 Hiera-Small | 169,638,840 bytes | `90aa65e8b9c98cca73916c0b579d0b63139ef60e8b508946c389204819195261` |
| `sam2_hiera_base_plus.zip` | SAM 2 Hiera-Base+ | 302,582,623 bytes | `c30ce8030d9be4ee68afb825285d0d50b706571a671847ba0d7cb36e7aab75c3` |
| `sam2_hiera_large.zip` | SAM 2 Hiera-Large | 843,636,551 bytes | `97d5841ca0827acb3ae54ed896adedb23cc77a4fa23118402efca3e169d20542` |
| `sam2_1_hiera_tiny.zip` | SAM 2.1 Hiera-Tiny | 116,527,977 bytes | `b15c3d3265392901f6dbe45a42e2f64709bf265b15fde899b2439bc04496ba04` |
| `sam2_1_hiera_small.zip` | SAM 2.1 Hiera-Small | 142,921,559 bytes | `5802f9e3b05d41b9a5ce03f32d3935fe6cc94294b77956f37c58b7f2581471ce` |
| `sam2_1_hiera_base_plus.zip` | SAM 2.1 Hiera-Base+ | 272,032,266 bytes | `389aae8f1e552275326e13a10973e822583989dba95aa7c3ffc80022460ac7eb` |
| `sam2_1_hiera_large.zip` | SAM 2.1 Hiera-Large | 805,293,551 bytes | `15f74c68530b0bc9f37d2189394a100b81541f55362b34f55a309ba12b9e1fa4` |
| `rfdetr_nano_detection_1_9_4.zip` | RF-DETR Nano detection | 99,769,055 bytes | `5b130a1c2eb01be3bfbda703367b5d993821690de26f4abbb14a94ee3660c5fe` |
| `rfdetr_nano_segmentation_1_9_4.zip` | RF-DETR Segmentation Nano | 113,602,026 bytes | `133fdb5aed76233a6959addbdb7d64f5132f3cf2b2705a1994c2cbdfc2e93f4d` |
| `dfine_n_coco_956d170.zip` | D-FINE-N COCO detection | 13,979,719 bytes | `f753f6e552632ef1696ec53c92ccb1ca374dbf2f014c6e4ab005ab98efa3a6e8` |

Each promptable-segmentation ZIP contains a small AnyLearning model
configuration and one encoder plus one decoder ONNX model. Each RF-DETR and
D-FINE ZIP contains one static ONNX graph plus its provenance, checksums, and
verbatim upstream license. `MANIFEST.json` records the exact source revision, archive
size, checksum, and expected members. Every transformed or exported bundle pins
each extracted member's size and SHA-256 as well as its source artifact
identity.

## Provenance

The source graph pairs originate from checksum-pinned ONNX exports published by
Viet-Anh Nguyen:

- SAM 2 bundles: source revision
  `071f58077599431edd0e5d2ac52ecca4c78f1cab` from
  `vietanhdev/segment-anything-2-onnx-models`.
- SAM 2.1 bundles: source revision
  `6a3ac868340a3196a349050a6efae22a5acc0330` from
  `vietanhdev/segment-anything-2.1-onnx-models`.
- MobileSAM bundle: source revision
  `9effc01a9e135621d710d49159f1ffb0b6f724dc` from
  `vietanhdev/segment-anything-onnx-models`.
- EfficientViT-SAM encoders and source decoders: immutable model revision
  `a2f0c5929196a13ef1ffc9338d3f9e482e1e0e68` from
  `mit-han-lab/efficientvit-sam`. The published decoders were transformed by
  [AnyLearning's checksum-gated ONNX-only tool](https://github.com/nrl-ai/anylearning-oss/blob/450e4fc7e291d7166b5ac650b7b72c42fac70203/scripts/prepare_efficientvit_sam_decoder.py)
  to expose the four native mask tokens before the official graph's lossy
  single-mask selection. The transform does not load a native checkpoint or
  alter the learned tensors.
- RF-DETR Nano detection and instance segmentation: official checkpoints and
  exporter from `rfdetr==1.9.4`, source revision
  `9b009fa928d6218320439803d1da01869a85c072`. Both graphs use static batch-one
  float32 contracts at opset 17. Native/export parity evidence is recorded
  inside each archive.
- D-FINE-N detection: the official COCO-only checkpoint and exporter at source
  revision `956d1709314c2c6a4df6f34de232054578a7449f`. The checkpoint was loaded
  only by the included restricted weights-only conversion helper. The static
  opset-16 graph has exact native label parity and less than 0.005-pixel box
  drift on the retained landscape/portrait corpus. Objects365-derived weights
  are excluded.

The SAM 2 and SAM 2.1 encoders were transformed by
[AnyLearning's checksum-gated ONNX-only tool](https://github.com/nrl-ai/anylearning-oss/blob/7575ab52cc0a31e5ec3d71b6ec7157dbe042c7d7/scripts/prepare_sam2_encoder.py).
It repairs exact stale shape metadata, removes only unreachable initializers,
requires strict ONNX type/shape inference, and records the source digest. It
does not load a native checkpoint or change reachable learned tensors. The
decoders are unchanged mirrors.

Original model projects:

- [Segment Anything 2](https://github.com/facebookresearch/sam2)
- [Segment Anything](https://github.com/facebookresearch/segment-anything)
- [MobileSAM](https://github.com/ChaoningZhang/MobileSAM)
- [EfficientViT-SAM](https://github.com/mit-han-lab/efficientvit/tree/de7d7733cc0329f391b33f1f459271562ec27bd5/applications/efficientvit_sam)
- [RF-DETR](https://github.com/roboflow/rf-detr/tree/9b009fa928d6218320439803d1da01869a85c072)
- [D-FINE](https://github.com/Peterande/D-FINE/tree/956d1709314c2c6a4df6f34de232054578a7449f)

The SAM and MobileSAM files are unchanged mirrors. EfficientViT-SAM encoders
are unchanged mirrors; their decoders preserve the original learned tensors and
compute graph while exposing the four pre-selection outputs and recording
transform provenance in ONNX metadata. SAM 2 and SAM 2.1 prepared encoders are
paired with unchanged source decoders in deterministic archives.

## Secure and reproducible download

Pin a repository revision and verify the SHA-256 value from `MANIFEST.json`
before extracting or loading a model. Consumers should reject absolute paths,
parent traversal, links, encrypted or duplicate entries, unexpected archive
members, and files whose exact sizes or digests differ from the manifest.

AnyLearning performs the full image encoding, prompt conversion, mask decoding,
and editable-shape conversion. These archives are not standalone applications.

## Intended use and limitations

- Intended for interactive point/rectangle-prompt segmentation in AnyLearning.
- RF-DETR bundles are intended for human-reviewed COCO object detection and
  instance-segmentation suggestions.
- The D-FINE bundle is intended for human-reviewed COCO object-detection
  suggestions.
- Results require human review before becoming dataset labels.
- Quality and latency vary with image content, hardware, execution provider, and
  model size.
- These models can reproduce biases and limitations of their original training
  data.
- Do not use segmentation output as the sole basis for safety-critical,
  medical, legal, or similarly consequential decisions.

## License

The model code and weights are distributed under Apache License 2.0 by their
respective upstream projects. See `LICENSES.md` for source and attribution links.

## Citation

For SAM 2 and SAM 2.1, cite:

```bibtex
@article{ravi2024sam2,
  title={SAM 2: Segment Anything in Images and Videos},
  author={Ravi, Nikhila and others},
  journal={arXiv:2408.00714},
  year={2024}
}
```

For SAM, cite:

```bibtex
@article{kirillov2023segment,
  title={Segment Anything},
  author={Kirillov, Alexander and others},
  journal={arXiv:2304.02643},
  year={2023}
}
```

For EfficientViT-SAM, cite:

```bibtex
@inproceedings{cai2023efficientvit,
  title={EfficientViT: Lightweight Multi-Scale Attention for High-Resolution Dense Prediction},
  author={Cai, Han and Li, Junyan and Hu, Muyan and Gan, Chuang and Han, Song},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  year={2023}
}

@article{zhang2024efficientvit,
  title={EfficientViT-SAM: Accelerated Segment Anything Model Without Performance Loss},
  author={Zhang, Zhuoyang and Cai, Han and Han, Song},
  journal={arXiv preprint arXiv:2402.05008},
  year={2024}
}
```

For RF-DETR, use the citation requested by the
[official project](https://github.com/roboflow/rf-detr/tree/9b009fa928d6218320439803d1da01869a85c072).

For D-FINE, use the citation requested by the
[official project](https://github.com/Peterande/D-FINE/tree/956d1709314c2c6a4df6f34de232054578a7449f).