---
license: other
library_name: onnxruntime
pipeline_tag: image-segmentation
tags:
  - segment-anything
  - sam
  - sam2
  - sam3
  - mobilesam
  - efficientsam
  - onnx
  - onnxruntime
---

# SAMExporter validated ONNX models

Real, structurally checked ONNX artifacts for
[`samexporter`](https://github.com/vietanhdev/samexporter). These files are
published by NRL.ai so the AnyLearning and SAMExporter documentation can use
stable, revisioned model links.

## Available artifacts

| Directory | Prompt types | Source | Terms |
|---|---|---|---|
| `sam_vit_b/` | point, rectangle | Meta SAM ViT-B | Apache-2.0 |
| `mobile_sam/` | point, rectangle | MobileSAM TinyViT | Apache-2.0 |
| `efficient_sam_ti/` | point, rectangle | EfficientSAM-Ti official ONNX | Apache-2.0 |
| `efficient_sam_s/` | point, rectangle | EfficientSAM-S official ONNX | Apache-2.0 |
| `sam2_1_hiera_tiny/` | point, rectangle, mask refinement | Meta SAM 2.1 Tiny | Apache-2.0 |
| `sam3/` | text, positive/negative geometry | Meta SAM3 | SAM License (included) |

Each directory contains its applicable upstream license. This repository uses
`license: other` because SAM3 is distributed under Meta's SAM License while
the other five families are Apache-2.0. Use and redistribution remain subject
to the license in each directory. No source checkpoints or training datasets
are included.

## Download

Download only the family you need:

```bash
hf download nrl-ai/samexporter-onnx-models \
  --include "mobile_sam/*" \
  --local-dir output_models
```

Or download every family:

```bash
hf download nrl-ai/samexporter-onnx-models --local-dir output_models
```

SAM3 uses ONNX external data: keep `sam3_decoder.onnx` and
`sam3_decoder.onnx.data` together in the same directory.

## Validation

- Every ONNX graph passes `onnx.checker.check_model(path)` with external data
  resolved from its published directory.
- SAM ViT-B, MobileSAM, SAM 2.1 Tiny, and SAM3 were exported from real official
  checkpoints; EfficientSAM-Ti and EfficientSAM-S are the official upstream
  split ONNX pairs.
- Point and rectangle results were run with ONNX Runtime CPU on the landscape
  truck and portrait plant images in the SAMExporter repository.
- Full-resolution overlays and sibling resource logs are retained in
  [`visual_results/`](https://github.com/vietanhdev/samexporter/tree/main/visual_results).

Exact file digests are listed in `CHECKSUMS.sha256`. The export provenance is
listed in `PROVENANCE.md`.

## Runtime

```bash
pip install "samexporter[runtime-cpu]"

python -m samexporter.inference \
  --sam_variant sam \
  --encoder_model output_models/mobile_sam/mobile_sam.encoder.onnx \
  --decoder_model output_models/mobile_sam/mobile_sam.decoder.onnx \
  --image image.jpg \
  --prompt prompt.json \
  --output result.png
```

See the
[`samexporter` documentation](https://github.com/vietanhdev/samexporter#readme)
for family-specific conversion, provider, prompt, and performance guidance.