FFNet-78S: Optimized for Qualcomm Devices

FFNet-78S is a "fuss-free network" that segments street scene images with per-pixel classes like road, sidewalk, and pedestrian. Trained on the Cityscapes dataset.

This is based on the implementation of FFNet-78S 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 2.37, ONNX Runtime 1.23.0	Download
ONNX	w8a8	Universal	QAIRT 2.37, ONNX Runtime 1.23.0	Download
QNN_DLC	float	Universal	QAIRT 2.42	Download
QNN_DLC	w8a8	Universal	QAIRT 2.42	Download
TFLITE	float	Universal	QAIRT 2.42, TFLite 2.17.0	Download
TFLITE	w8a8	Universal	QAIRT 2.42, TFLite 2.17.0	Download

For more device-specific assets and performance metrics, visit FFNet-78S 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 FFNet-78S on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.semantic_segmentation

Model Stats:

• Model checkpoint: ffnet78S_dBBB_cityscapes_state_dict_quarts
• Input resolution: 2048x1024
• Number of output classes: 19
• Number of parameters: 27.5M
• Model size (float): 105 MB
• Model size (w8a8): 26.7 MB

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
FFNet-78S	ONNX	float	Snapdragon® X Elite	33.983 ms	31 - 31 MB	NPU
FFNet-78S	ONNX	float	Snapdragon® 8 Gen 3 Mobile	24.542 ms	31 - 300 MB	NPU
FFNet-78S	ONNX	float	Qualcomm® QCS8550 (Proxy)	34.536 ms	24 - 55 MB	NPU
FFNet-78S	ONNX	float	Qualcomm® QCS9075	59.887 ms	24 - 51 MB	NPU
FFNet-78S	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	26.338 ms	6 - 160 MB	NPU
FFNet-78S	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	13.372 ms	13 - 212 MB	NPU
FFNet-78S	ONNX	w8a8	Snapdragon® X Elite	12.19 ms	22 - 22 MB	NPU
FFNet-78S	ONNX	w8a8	Snapdragon® 8 Gen 3 Mobile	7.941 ms	7 - 239 MB	NPU
FFNet-78S	ONNX	w8a8	Qualcomm® QCS6490	488.488 ms	162 - 231 MB	CPU
FFNet-78S	ONNX	w8a8	Qualcomm® QCS8550 (Proxy)	11.597 ms	0 - 24 MB	NPU
FFNet-78S	ONNX	w8a8	Qualcomm® QCS9075	14.86 ms	6 - 9 MB	NPU
FFNet-78S	ONNX	w8a8	Qualcomm® QCM6690	536.419 ms	167 - 177 MB	CPU
FFNet-78S	ONNX	w8a8	Snapdragon® 8 Elite For Galaxy Mobile	6.021 ms	2 - 160 MB	NPU
FFNet-78S	ONNX	w8a8	Snapdragon® 7 Gen 4 Mobile	532.054 ms	172 - 182 MB	CPU
FFNet-78S	ONNX	w8a8	Snapdragon® 8 Elite Gen 5 Mobile	4.503 ms	2 - 168 MB	NPU
FFNet-78S	QNN_DLC	float	Snapdragon® X Elite	44.274 ms	24 - 24 MB	NPU
FFNet-78S	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	29.913 ms	24 - 333 MB	NPU
FFNet-78S	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	188.209 ms	24 - 234 MB	NPU
FFNet-78S	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	42.829 ms	24 - 26 MB	NPU
FFNet-78S	QNN_DLC	float	Qualcomm® SA8775P	60.916 ms	24 - 235 MB	NPU
FFNet-78S	QNN_DLC	float	Qualcomm® QCS9075	73.536 ms	24 - 52 MB	NPU
FFNet-78S	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	84.656 ms	2 - 294 MB	NPU
FFNet-78S	QNN_DLC	float	Qualcomm® SA7255P	188.209 ms	24 - 234 MB	NPU
FFNet-78S	QNN_DLC	float	Qualcomm® SA8295P	66.245 ms	24 - 230 MB	NPU
FFNet-78S	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	21.872 ms	23 - 259 MB	NPU
FFNet-78S	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	15.403 ms	24 - 275 MB	NPU
FFNet-78S	QNN_DLC	w8a8	Snapdragon® X Elite	17.601 ms	6 - 6 MB	NPU
FFNet-78S	QNN_DLC	w8a8	Snapdragon® 8 Gen 3 Mobile	11.715 ms	6 - 289 MB	NPU
FFNet-78S	QNN_DLC	w8a8	Qualcomm® QCS6490	72.702 ms	6 - 14 MB	NPU
FFNet-78S	QNN_DLC	w8a8	Qualcomm® QCS8275 (Proxy)	39.342 ms	6 - 217 MB	NPU
FFNet-78S	QNN_DLC	w8a8	Qualcomm® QCS8550 (Proxy)	16.762 ms	6 - 8 MB	NPU
FFNet-78S	QNN_DLC	w8a8	Qualcomm® SA8775P	17.256 ms	6 - 218 MB	NPU
FFNet-78S	QNN_DLC	w8a8	Qualcomm® QCS9075	19.692 ms	6 - 14 MB	NPU
FFNet-78S	QNN_DLC	w8a8	Qualcomm® QCM6690	163.978 ms	6 - 254 MB	NPU
FFNet-78S	QNN_DLC	w8a8	Qualcomm® QCS8450 (Proxy)	23.855 ms	6 - 286 MB	NPU
FFNet-78S	QNN_DLC	w8a8	Qualcomm® SA7255P	39.342 ms	6 - 217 MB	NPU
FFNet-78S	QNN_DLC	w8a8	Qualcomm® SA8295P	23.444 ms	6 - 220 MB	NPU
FFNet-78S	QNN_DLC	w8a8	Snapdragon® 8 Elite For Galaxy Mobile	8.024 ms	6 - 232 MB	NPU
FFNet-78S	QNN_DLC	w8a8	Snapdragon® 7 Gen 4 Mobile	22.55 ms	6 - 236 MB	NPU
FFNet-78S	QNN_DLC	w8a8	Snapdragon® 8 Elite Gen 5 Mobile	5.878 ms	6 - 256 MB	NPU
FFNet-78S	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	29.599 ms	2 - 394 MB	NPU
FFNet-78S	TFLITE	float	Qualcomm® QCS8275 (Proxy)	188.5 ms	3 - 248 MB	NPU
FFNet-78S	TFLITE	float	Qualcomm® QCS8550 (Proxy)	42.497 ms	2 - 62 MB	NPU
FFNet-78S	TFLITE	float	Qualcomm® SA8775P	283.205 ms	3 - 249 MB	NPU
FFNet-78S	TFLITE	float	Qualcomm® QCS9075	73.289 ms	0 - 82 MB	NPU
FFNet-78S	TFLITE	float	Qualcomm® QCS8450 (Proxy)	85.796 ms	3 - 377 MB	NPU
FFNet-78S	TFLITE	float	Qualcomm® SA7255P	188.5 ms	3 - 248 MB	NPU
FFNet-78S	TFLITE	float	Qualcomm® SA8295P	66.125 ms	2 - 244 MB	NPU
FFNet-78S	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	21.778 ms	1 - 265 MB	NPU
FFNet-78S	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	15.421 ms	2 - 288 MB	NPU
FFNet-78S	TFLITE	w8a8	Snapdragon® 8 Gen 3 Mobile	6.474 ms	1 - 284 MB	NPU
FFNet-78S	TFLITE	w8a8	Qualcomm® QCS6490	57.237 ms	1 - 36 MB	NPU
FFNet-78S	TFLITE	w8a8	Qualcomm® QCS8275 (Proxy)	26.423 ms	1 - 210 MB	NPU
FFNet-78S	TFLITE	w8a8	Qualcomm® QCS8550 (Proxy)	9.08 ms	1 - 3 MB	NPU
FFNet-78S	TFLITE	w8a8	Qualcomm® SA8775P	9.716 ms	1 - 212 MB	NPU
FFNet-78S	TFLITE	w8a8	Qualcomm® QCS9075	10.947 ms	1 - 36 MB	NPU
FFNet-78S	TFLITE	w8a8	Qualcomm® QCM6690	139.439 ms	1 - 246 MB	NPU
FFNet-78S	TFLITE	w8a8	Qualcomm® QCS8450 (Proxy)	18.38 ms	1 - 282 MB	NPU
FFNet-78S	TFLITE	w8a8	Qualcomm® SA7255P	26.423 ms	1 - 210 MB	NPU
FFNet-78S	TFLITE	w8a8	Qualcomm® SA8295P	14.967 ms	0 - 213 MB	NPU
FFNet-78S	TFLITE	w8a8	Snapdragon® 8 Elite For Galaxy Mobile	4.984 ms	1 - 227 MB	NPU
FFNet-78S	TFLITE	w8a8	Snapdragon® 7 Gen 4 Mobile	14.654 ms	0 - 229 MB	NPU
FFNet-78S	TFLITE	w8a8	Snapdragon® 8 Elite Gen 5 Mobile	3.61 ms	0 - 247 MB	NPU

License

• The license for the original implementation of FFNet-78S can be found here.

References

• Simple and Efficient Architectures for Semantic Segmentation
• 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.