SAMExporter — Export SAM Models to ONNX

SAMExporter is the tool used to export the Segment Anything family of models into ONNX format for deployment in AnyLabeling and other applications. It supports the full model family: SAM, MobileSAM, SAM2, SAM2.1, and SAM3.

GitHub: vietanhdev/samexporter (opens in a new tab)
PyPI: pypi.org/project/samexporter (opens in a new tab)
Pre-exported SAM3 ONNX models: vietanhdev/segment-anything-3-onnx-models (opens in a new tab)

Supported Models

Model	Prompt Types	Speed	Accuracy
SAM ViT-B	Point, Rectangle	Fast	Good
SAM ViT-L	Point, Rectangle	Medium	Better
SAM ViT-H	Point, Rectangle	Slow	Best
SAM ViT-B/L/H (quantized)	Point, Rectangle	Faster	Slightly lower
MobileSAM	Point, Rectangle	Fastest	Lower
SAM2 Hiera-Tiny	Point, Rectangle	Fast	Good
SAM2 Hiera-Small	Point, Rectangle	Medium	Better
SAM2 Hiera-Base+	Point, Rectangle	Medium	Better
SAM2 Hiera-Large	Point, Rectangle	Slow	Best SAM2
SAM2.1 Tiny / Small / Base+ / Large	Point, Rectangle	Same as SAM2	Improved SAM2
SAM3 ViT-H	Text, Point, Rectangle	Slow	Open-vocabulary

Installation

Requires Python 3.11+.

pip install torch==2.10.0 torchvision==0.25.0 --index-url https://download.pytorch.org/whl/cpu
pip install samexporter

For model export with ONNX simplification (Linux/macOS or Windows with Long Path support enabled):

pip install "samexporter[export]"

Note: The [export] extra installs onnxsim which requires building from source on Windows. It is only needed when exporting models with --simplify. Inference from pre-exported ONNX models does not need it.

SAM / MobileSAM

Download checkpoints

original_models/
  sam_vit_b_01ec64.pth   → https://dl.fbaipublicfiles.com/segment_anything/sam_vit_b_01ec64.pth
  sam_vit_l_0b3195.pth   → https://dl.fbaipublicfiles.com/segment_anything/sam_vit_l_0b3195.pth
  sam_vit_h_4b8939.pth   → https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth
  mobile_sam.pt          → https://github.com/ChaoningZhang/MobileSAM (weights/)

Export

# Export encoder (example: ViT-H)
python -m samexporter.export_encoder \
    --checkpoint original_models/sam_vit_h_4b8939.pth \
    --output output_models/sam_vit_h_4b8939.encoder.onnx \
    --model-type vit_h \
    --quantize-out output_models/sam_vit_h_4b8939.encoder.quant.onnx \
    --use-preprocess
 
# Export decoder
python -m samexporter.export_decoder \
    --checkpoint original_models/sam_vit_h_4b8939.pth \
    --output output_models/sam_vit_h_4b8939.decoder.onnx \
    --model-type vit_h \
    --quantize-out output_models/sam_vit_h_4b8939.decoder.quant.onnx \
    --return-single-mask

Batch convert all SAM models:

bash convert_all_meta_sam.sh
bash convert_mobile_sam.sh

Inference

python -m samexporter.inference \
    --encoder_model output_models/sam_vit_h_4b8939.encoder.onnx \
    --decoder_model output_models/sam_vit_h_4b8939.decoder.onnx \
    --image images/truck.jpg \
    --prompt images/truck_prompt.json \
    --output output_images/truck.png \
    --show

SAM1 truck inference result

SAM2 / SAM2.1

Download checkpoints

cd original_models && bash download_sam2.sh

Install SAM2 PyTorch package

pip install git+https://github.com/facebookresearch/segment-anything-2.git

Export

# SAM2 Tiny
python -m samexporter.export_sam2 \
    --checkpoint original_models/sam2_hiera_tiny.pt \
    --output_encoder output_models/sam2_hiera_tiny.encoder.onnx \
    --output_decoder output_models/sam2_hiera_tiny.decoder.onnx \
    --model_type sam2_hiera_tiny
 
# SAM2.1 Tiny
python -m samexporter.export_sam2 \
    --checkpoint original_models/sam2.1_hiera_tiny.pt \
    --output_encoder output_models/sam2.1_hiera_tiny.encoder.onnx \
    --output_decoder output_models/sam2.1_hiera_tiny.decoder.onnx \
    --model_type sam2.1_hiera_tiny

Batch convert all SAM2 / SAM2.1 variants:

bash convert_all_meta_sam2.sh

Inference

python -m samexporter.inference \
    --encoder_model output_models/sam2_hiera_tiny.encoder.onnx \
    --decoder_model output_models/sam2_hiera_tiny.decoder.onnx \
    --image images/truck.jpg \
    --prompt images/truck_prompt.json \
    --sam_variant sam2 \
    --output output_images/sam2_truck.png \
    --show

SAM2 truck inference result

Note: SAM2.1 uses the same inference command (--sam_variant sam2) as SAM2. The only difference is the model files.

SAM3 — Open-Vocabulary Segmentation

SAM3 extends the SAM family with text-driven, open-vocabulary segmentation. It accepts natural-language text prompts (e.g., "truck", "person on a bike") to detect and segment objects without any class-specific training.

SAM3 consists of three separate ONNX models that run in sequence:

Image → [Image Encoder] → image features ─┐
Text  → [Language Encoder] → text features ┼→ [Decoder] → boxes + scores + masks
Prompt (box/point) ────────────────────────┘

Pre-exported ONNX models

Pre-exported models are available on HuggingFace and are downloaded automatically when using AnyLabeling:

vietanhdev/segment-anything-3-onnx-models
├── sam3_image_encoder.onnx   + sam3_image_encoder.onnx.data   (~1.8 GB)
├── sam3_language_encoder.onnx + sam3_language_encoder.onnx.data (~1.6 GB)
└── sam3_decoder.onnx          + sam3_decoder.onnx.data         (~116 MB)

Export from PyTorch (optional)

Only needed if you want to re-export the models yourself. Requires the sam3 git submodule.

git submodule update --init sam3
pip install osam  # CLIP tokenizer for SAM3
python -m samexporter.export_sam3 --output_dir output_models/sam3 --opset 18

Inference

Always pass --text_prompt when running SAM3 inference. Without it the model defaults to a visual (non-text) token and may produce zero detections.

Text-only (finds all matching objects):

python -m samexporter.inference \
    --sam_variant sam3 \
    --encoder_model output_models/sam3/sam3_image_encoder.onnx \
    --decoder_model output_models/sam3/sam3_decoder.onnx \
    --language_encoder_model output_models/sam3/sam3_language_encoder.onnx \
    --image images/truck.jpg \
    --prompt images/truck_sam3.json \
    --text_prompt "truck" \
    --output output_images/truck_sam3.png \
    --show

Text + rectangle (text drives detection, rectangle refines region):

python -m samexporter.inference \
    --sam_variant sam3 \
    --encoder_model output_models/sam3/sam3_image_encoder.onnx \
    --decoder_model output_models/sam3/sam3_decoder.onnx \
    --language_encoder_model output_models/sam3/sam3_language_encoder.onnx \
    --image images/truck.jpg \
    --prompt images/truck_sam3_box.json \
    --text_prompt "truck" \
    --output output_images/truck_sam3_box.png \
    --show

Text + point:

python -m samexporter.inference \
    --sam_variant sam3 \
    --encoder_model output_models/sam3/sam3_image_encoder.onnx \
    --decoder_model output_models/sam3/sam3_decoder.onnx \
    --language_encoder_model output_models/sam3/sam3_language_encoder.onnx \
    --image images/truck.jpg \
    --prompt images/truck_sam3_point.json \
    --text_prompt "truck" \
    --output output_images/truck_sam3_point.png \
    --show

Prompt JSON format

Each prompt file is a JSON array of mark objects:

[
  {"type": "point",     "data": [x, y],           "label": 1},
  {"type": "rectangle", "data": [x1, y1, x2, y2]},
  {"type": "text",      "data": "object description"}
]

label: 1 — foreground point; label: 0 — background point (negative)
type: "text" is for SAM3; use --text_prompt on the CLI for convenience

Inference CLI reference

samexporter.inference

  --encoder_model PATH           Image encoder ONNX model
  --decoder_model PATH           Mask decoder ONNX model
  --language_encoder_model PATH  Language encoder (SAM3 only)
  --image PATH                   Input image (JPG, PNG, ...)
  --prompt PATH                  Prompt JSON file
  --output PATH                  Output image path
  --sam_variant {sam,sam2,sam3}  Model family (default: sam)
  --text_prompt TEXT             Text override for SAM3 (e.g. "truck")
  --show                         Display result window

Architecture notes

SAM / MobileSAM

Encoder: Takes the full image, outputs a fixed 256-dim embedding. Run once per image.
Decoder: Takes the embedding + user prompt (points/boxes), outputs masks in real time.

SAM2 / SAM2.1

Encoder: Outputs three feature levels (high_res_feats_0, high_res_feats_1, image_embedding).
Decoder: Takes multi-scale features + prompt, outputs the best mask.

SAM3

Image Encoder: Input is raw uint8 [3, 1008, 1008]. Outputs 6 tensors (vision_pos_enc_{0,1,2}, backbone_fpn_{0,1,2}). Normalization is baked in.
Language Encoder: Input is CLIP tokens [1, 32] int64 (max 32 tokens). Outputs text attention mask, text memory, text embeddings.
Decoder: Accepts image features + language features + geometric prompt. Outputs boxes (N,4), scores (N,), masks (N,1,H,W) (boolean). All N detected objects are returned.

Tips

Use quantized models (*.quant.onnx) for faster inference and smaller download size. Accuracy is only marginally reduced.
MobileSAM is the best choice for CPU-only environments with tight latency requirements.
SAM2 / SAM2.1 outperform SAM1 on most benchmarks and are recommended for new deployments.
SAM3 is uniquely suited for open-set detection tasks where you do not know the class list in advance.
The image encoder runs once per image. The lightweight decoder handles prompt changes interactively without re-encoding.

Running tests

pip install pytest
cd samexporter
pytest tests/

All 14 unit tests run without requiring ONNX model files (sessions are mocked).

Segment Anything Model (SAM)YOLO Models