vietanhdev/segment-anything-2-onnx-models

Segment Anything 2 (SAM 2) — ONNX Models

ONNX exports of Meta's SAM 2 image-segmentation backbones, packaged for direct use with onnxruntime and AnyLabeling.

Why this repo exists

SAM 2 is materially better than SAM 1 on speed and quality, but the official release ships PyTorch checkpoints. ONNX gives you a portable, dependency-light runtime that works in Python, C++, JavaScript, and most embedded targets. These exports are the ones AnyLabeling consumes for its smart-labeling features.

Variants

Each backbone is provided in two equivalent forms — pick whichever fits your loader:

• Raw ONNX pair: <backbone>.encoder.onnx + <backbone>.decoder.onnx
• Bundled zip: <backbone>.zip containing both files

Backbone	Encoder size	Decoder size	Bundle
sam2_hiera_tiny	128 MB	19.7 MB	sam2_hiera_tiny.zip (148 MB)
sam2_hiera_small	155 MB	19.7 MB	sam2_hiera_small.zip (175 MB)
sam2_hiera_base_plus	324 MB	19.7 MB	sam2_hiera_base_plus.zip (344 MB)
sam2_hiera_large	848 MB	19.7 MB	sam2_hiera_large.zip (868 MB)

zip_models.py is the helper script used to produce the bundled zips from the encoder/decoder pairs.

Quick start

pip install huggingface_hub onnxruntime

from huggingface_hub import hf_hub_download
import onnxruntime as ort

repo = "vietanhdev/segment-anything-2-onnx-models"
encoder = hf_hub_download(repo_id=repo, filename="sam2_hiera_tiny.encoder.onnx")
decoder = hf_hub_download(repo_id=repo, filename="sam2_hiera_tiny.decoder.onnx")

enc = ort.InferenceSession(encoder, providers=["CPUExecutionProvider"])
dec = ort.InferenceSession(decoder, providers=["CPUExecutionProvider"])

# Inspect expected inputs:
print("Encoder:", [(i.name, i.shape, i.type) for i in enc.get_inputs()])
print("Decoder:", [(i.name, i.shape, i.type) for i in dec.get_inputs()])

For the full image → mask pipeline (encoder + decoder + prompt handling), see how AnyLabeling wires it: https://github.com/vietanhdev/anylabeling

Use with AnyLabeling

These models drop into AnyLabeling's auto-labeling backend without conversion. See the AnyLabeling docs for the model-config wiring.

Source weights

Original SAM 2 weights and license: https://github.com/facebookresearch/sam2

This repo redistributes the same weights in ONNX format. License unchanged from upstream (Apache 2.0).

Citation

@misc{nguyen2026sam2_onnx,
  author = {Nguyen, Viet-Anh and {Neural Research Lab}},
  title  = {SAM 2 ONNX Models},
  year   = {2026},
  url    = {https://huggingface.co/vietanhdev/segment-anything-2-onnx-models}
}

For the underlying model, cite Meta's SAM 2 paper:

@article{ravi2024sam2,
  title   = {SAM 2: Segment Anything in Images and Videos},
  author  = {Ravi, Nikhila and Gabeur, Valentin and Hu, Yuan-Ting and Hu, Ronghang and Ryali, Chaitanya and Ma, Tengyu and Khedr, Haitham and R{\"a}dle, Roman and Rolland, Chloe and Gustafson, Laura and Mintun, Eric and Pan, Junting and Alwala, Kalyan Vasudev and Carion, Nicolas and Wu, Chao-Yuan and Girshick, Ross and Doll{\'a}r, Piotr and Feichtenhofer, Christoph},
  journal = {arXiv:2408.00714},
  year    = {2024}
}

Acknowledgments

Thanks to Meta AI Research for releasing SAM 2. This repo packages their work for edge inference.