Export Meta AI's newest Segment Anything 3 (SAM-3) model to ONNX and then build a TensorRT engine you can deploy for real-time segmentation.

• Minimal, readable export script using Hugging Face transformers' Sam3Model and Sam3Processor
• Clean ONNX graph with instance masks (pred_masks) and semantic mask outputs
• TensorRT-ready: plug the produced ONNX into trtexec or your own build pipeline
• CPU-friendly export for maximum compatibility (flip device to use GPU if you prefer)

1. Request access to the gated model

   • Visit https://huggingface.co/facebook/sam3 and click “Access repository” (approval is required before downloads succeed). Make sure your HF_TOKEN has that access.

2. Install dependencies

pip install torch transformers pillow requests

# Optional: If after a pip install the ONNX export script complains that it cannot find SAM3, it's just because SAM-3 is very new and a formal release of `transformers` has not added support yet. In this case, please install `transformers` using git

git clone https://github.com/huggingface/transformers.git
cd transformers
pip install '.[torch]'
# or uv pip install '.[torch]'

Use a PyTorch wheel that matches your CUDA if you plan to export on GPU. By default a GPU is not required to export to ONNX.

1. Export to ONNX

export HF_TOKEN=<YOUR TOKEN>
python onnxexport.py

This downloads facebook/sam3, runs a sample prompt ("ear"), and writes onnx_weights/sam3_static.onnx plus associated weight shards. SAM-3 is ~3.2 GB, so the ONNX exporter will create external data files; if you build TensorRT on another machine you must copy the entire onnx_weights/ directory (not just the .onnx).

1. Build a TensorRT engine Use trtexec (or your favorite builder) on the generated ONNX:

trtexec --onnx=onnx_weights/sam3_static.onnx --saveEngine=sam3_fp16.plan --fp16 --verbose # fp16
trtexec --onnx=onnx_weights/sam3_static.onnx --saveEngine=sam3_int8.plan --int8 --verbose # int8
trtexec --onnx=onnx_weights/sam3_static.onnx --saveEngine=sam3_fp8.plan --fp8 --verbose # fp8
trtexec --onnx=onnx_weights/sam3_static.onnx --saveEngine=sam3_int4.plan --int4 --verbose # int4

1. Validate (optional)

• Run onnxruntime on onnx_weights/sam3_static.onnx to confirm outputs
• Benchmark sam3_fp16.plan with trtexec --loadEngine=sam3_fp16.plan

• Change the prompt/image in onnxexport.py to better reflect your production use case.
• Swap device to "cuda" for GPU-side export if your environment supports it.
• Add more outputs from Sam3Model (e.g., scores) by extending Sam3ONNXWrapper.

• onnxexport.py — single, documented script that pulls SAM-3, builds clean inputs, and exports ONNX
• LICENSE — MIT

If this saved you time, drop a ⭐ so others can find it and ship SAM-3 faster.