Segment-Anything-Model-2: Optimized for Qualcomm Devices

SAM 2, the successor to Meta's Segment Anything Model (SAM), is a cutting-edge tool designed for comprehensive object segmentation in both images and videos. It excels in handling complex visual data through a unified, promptable model architecture that supports real-time processing and zero-shot generalization.

This is based on the implementation of Segment-Anything-Model-2 found here. This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the Qualcomm® AI Hub Models library to export with custom configurations. More details on model performance across various devices, can be found here.

Qualcomm AI Hub Models uses Qualcomm AI Hub Workbench to compile, profile, and evaluate this model. Sign up to run these models on a hosted Qualcomm® device.

Getting Started

There are two ways to deploy this model on your device:

Option 1: Download Pre-Exported Models

Below are pre-exported model assets ready for deployment.

Runtime	Precision	Chipset	SDK Versions	Download
ONNX	float	Universal	QAIRT 2.42, ONNX Runtime 1.24.3	Download
QNN_DLC	float	Universal	QAIRT 2.45	Download
QNN_DLC	w8a8	Universal	QAIRT 2.45	Download

For more device-specific assets and performance metrics, visit Segment-Anything-Model-2 on Qualcomm® AI Hub.

Option 2: Export with Custom Configurations

Use the Qualcomm® AI Hub Models Python library to compile and export the model with your own:

• Custom weights (e.g., fine-tuned checkpoints)
• Custom input shapes
• Target device and runtime configurations

This option is ideal if you need to customize the model beyond the default configuration provided here.

See our repository for Segment-Anything-Model-2 on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.semantic_segmentation

Model Stats:

• Model checkpoint: sam2.1_hiera_t
• Input resolution: 720p (720x1280)
• Number of parameters (encoder): 33.5M
• Model size (encoder) (float): 128 MB
• Number of parameters (decoder): 6.22M
• Model size (decoder) (float): 23.7 MB

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
decoder	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	3.139 ms	17 - 242 MB	NPU
decoder	ONNX	float	Snapdragon® X2 Elite	3.216 ms	15 - 15 MB	NPU
decoder	ONNX	float	Snapdragon® 8 Gen 3 Mobile	4.73 ms	16 - 339 MB	NPU
decoder	ONNX	float	Qualcomm® QCS8550 (Proxy)	6.804 ms	16 - 23 MB	NPU
decoder	ONNX	float	Qualcomm® QCS9075	8.467 ms	16 - 19 MB	NPU
decoder	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	3.476 ms	1 - 233 MB	NPU
decoder	QNN_DLC	float	Snapdragon® X2 Elite	3.112 ms	16 - 16 MB	NPU
decoder	QNN_DLC	float	Snapdragon® X Elite	5.516 ms	16 - 16 MB	NPU
decoder	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	3.678 ms	4 - 214 MB	NPU
decoder	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	5.381 ms	16 - 125 MB	NPU
decoder	QNN_DLC	float	Qualcomm® QCS9075	6.811 ms	16 - 34 MB	NPU
decoder	QNN_DLC	float	Qualcomm® SA8295P	8.166 ms	0 - 193 MB	NPU
decoder	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	2.899 ms	4 - 189 MB	NPU
decoder	QNN_DLC	w8a8	Snapdragon® 8 Elite Gen 5 Mobile	0.877 ms	1 - 109 MB	NPU
decoder	QNN_DLC	w8a8	Snapdragon® X2 Elite	1.259 ms	4 - 4 MB	NPU
decoder	QNN_DLC	w8a8	Snapdragon® X Elite	2.319 ms	4 - 4 MB	NPU
decoder	QNN_DLC	w8a8	Snapdragon® 8 Gen 3 Mobile	1.297 ms	0 - 110 MB	NPU
decoder	QNN_DLC	w8a8	Qualcomm® QCS6490	8.252 ms	1 - 7 MB	NPU
decoder	QNN_DLC	w8a8	Qualcomm® QCS8275 (Proxy)	3.891 ms	4 - 88 MB	NPU
decoder	QNN_DLC	w8a8	Qualcomm® QCS8550 (Proxy)	2.025 ms	4 - 113 MB	NPU
decoder	QNN_DLC	w8a8	Qualcomm® SA8775P	2.634 ms	4 - 89 MB	NPU
decoder	QNN_DLC	w8a8	Qualcomm® QCS9075	2.41 ms	0 - 6 MB	NPU
decoder	QNN_DLC	w8a8	Qualcomm® QCM6690	10.205 ms	4 - 275 MB	NPU
decoder	QNN_DLC	w8a8	Qualcomm® QCS8450 (Proxy)	2.556 ms	0 - 109 MB	NPU
decoder	QNN_DLC	w8a8	Qualcomm® SA7255P	3.891 ms	4 - 88 MB	NPU
decoder	QNN_DLC	w8a8	Qualcomm® SA8295P	2.959 ms	4 - 108 MB	NPU
decoder	QNN_DLC	w8a8	Snapdragon® 8 Elite For Galaxy Mobile	1.029 ms	1 - 89 MB	NPU
decoder	QNN_DLC	w8a8	Snapdragon® 7 Gen 4 Mobile	2.663 ms	4 - 215 MB	NPU
encoder	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	97.534 ms	44 - 2275 MB	NPU
encoder	ONNX	float	Snapdragon® X2 Elite	99.01 ms	71 - 71 MB	NPU
encoder	ONNX	float	Snapdragon® 8 Gen 3 Mobile	146.245 ms	1 - 3808 MB	NPU
encoder	ONNX	float	Qualcomm® QCS8550 (Proxy)	199.079 ms	0 - 88 MB	NPU
encoder	ONNX	float	Qualcomm® QCS9075	273.038 ms	12 - 26 MB	NPU
encoder	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	113.051 ms	33 - 1669 MB	NPU
encoder	QNN_DLC	float	Snapdragon® X2 Elite	107.977 ms	12 - 12 MB	NPU
encoder	QNN_DLC	float	Snapdragon® X Elite	206.229 ms	12 - 12 MB	NPU
encoder	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	155.813 ms	12 - 1967 MB	NPU
encoder	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	215.599 ms	12 - 16 MB	NPU
encoder	QNN_DLC	float	Qualcomm® QCS9075	258.641 ms	12 - 42 MB	NPU
encoder	QNN_DLC	float	Qualcomm® SA8295P	309.446 ms	0 - 1671 MB	NPU
encoder	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	119.576 ms	0 - 1735 MB	NPU
encoder	QNN_DLC	w8a8	Snapdragon® 8 Elite Gen 5 Mobile	38.315 ms	3 - 1810 MB	NPU
encoder	QNN_DLC	w8a8	Snapdragon® X2 Elite	40.925 ms	3 - 3 MB	NPU
encoder	QNN_DLC	w8a8	Snapdragon® X Elite	84.916 ms	3 - 3 MB	NPU
encoder	QNN_DLC	w8a8	Snapdragon® 8 Gen 3 Mobile	57.396 ms	0 - 1978 MB	NPU
encoder	QNN_DLC	w8a8	Qualcomm® QCS6490	666.244 ms	3 - 12 MB	NPU
encoder	QNN_DLC	w8a8	Qualcomm® QCS8275 (Proxy)	144.841 ms	3 - 1664 MB	NPU
encoder	QNN_DLC	w8a8	Qualcomm® QCS8550 (Proxy)	82.328 ms	3 - 6 MB	NPU
encoder	QNN_DLC	w8a8	Qualcomm® SA8775P	83.916 ms	3 - 1664 MB	NPU
encoder	QNN_DLC	w8a8	Qualcomm® QCS9075	102.775 ms	3 - 12 MB	NPU
encoder	QNN_DLC	w8a8	Qualcomm® QCM6690	1283.67 ms	3 - 1971 MB	NPU
encoder	QNN_DLC	w8a8	Qualcomm® QCS8450 (Proxy)	141.394 ms	3 - 2036 MB	NPU
encoder	QNN_DLC	w8a8	Qualcomm® SA7255P	144.841 ms	3 - 1664 MB	NPU
encoder	QNN_DLC	w8a8	Qualcomm® SA8295P	96.974 ms	3 - 1672 MB	NPU
encoder	QNN_DLC	w8a8	Snapdragon® 8 Elite For Galaxy Mobile	42.794 ms	3 - 1734 MB	NPU
encoder	QNN_DLC	w8a8	Snapdragon® 7 Gen 4 Mobile	110.547 ms	3 - 1774 MB	NPU

License

• The license for the original implementation of Segment-Anything-Model-2 can be found here.

References

• SAM 2 Segment Anything in Images and Videos
• Source Model Implementation

Community

• Join our AI Hub Slack community to collaborate, post questions and learn more about on-device AI.
• For questions or feedback please reach out to us.