---
title: sam3-onnx
canonical_url: "https://www.modelscope.cn/models/sealofyou/sam3-onnx"
md_url: "https://www.modelscope.cn/models/sealofyou/sam3-onnx.md"
repository: sealofyou/sam3-onnx
chinese_name: "sam3的onnx量化"
last_updated: 2026-02-11
pipeline_tag: image-object-detection
tasks:
  - image-object-detection
library_name:
  - onnx
  - pytorch
frameworks:
  - Pytorch
downloads: 23
stars: 1
---

# sam3-onnx

> sam3-onnx - sealofyou 在 ModelScope 开源的模型。This repository contains an ONNX FP16 encoder model for SAM3 benchmarking.

- **Repository**: sealofyou/sam3-onnx
- **Tasks**: image-object-detection
- **Downloads**: 23
- **Stars**: 1
- **Last updated**: 2026-02-11

Source: https://www.modelscope.cn/models/sealofyou/sam3-onnx

---

# sam3-onnx

This repository contains an ONNX FP16 encoder model for SAM3 benchmarking.

## Model
- File: `sam3_image_encoder_norope_fp16_native.onnx`
- Input: `float16`, shape `[1, 3, 1008, 1008]`
- Output: `vision_features`, shape `[1, 256, 72, 72]`

## Notes
- This is an encoder-only ONNX export (`no_rope`) used for memory/latency benchmarking.
- It is not a full end-to-end SAM3 detector/tracker model.
- Runtime tested with ONNX Runtime CUDA EP.

## Quick Example
```python
import numpy as np
import onnxruntime as ort

so = ort.SessionOptions()
so.graph_optimization_level = ort.GraphOptimizationLevel.ORT_DISABLE_ALL
sess = ort.InferenceSession(
    "sam3_image_encoder_norope_fp16_native.onnx",
    sess_options=so,
    providers=[("CUDAExecutionProvider", {}), "CPUExecutionProvider"],
)

x = np.random.randn(1, 3, 1008, 1008).astype(np.float16)
y = sess.run(None, {sess.get_inputs()[0].name: x})[0]
print(y.shape)
```