LaMa-Dilated: Optimized for Qualcomm Devices

LaMa-Dilated is a machine learning model that allows to erase and in-paint part of given input image.

This is based on the implementation of LaMa-Dilated 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 1.24.1	Download
ONNX	w8a16	Universal	QAIRT 2.42, ONNX Runtime 1.24.1	Download
QNN_DLC	float	Universal	QAIRT 2.43	Download
QNN_DLC	w8a16	Universal	QAIRT 2.43	Download
TFLITE	float	Universal	QAIRT 2.43, TFLite 2.17.0	Download

For more device-specific assets and performance metrics, visit LaMa-Dilated 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 LaMa-Dilated on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.image_editing

Model Stats:

• Model checkpoint: Dilated CelebAHQ
• Input resolution: 512x512
• Number of parameters: 45.6M
• Model size (float): 174 MB

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
LaMa-Dilated	ONNX	float	Snapdragon® X2 Elite	36.363 ms	90 - 90 MB	NPU
LaMa-Dilated	ONNX	float	Snapdragon® X Elite	78.251 ms	89 - 89 MB	NPU
LaMa-Dilated	ONNX	float	Snapdragon® 8 Gen 3 Mobile	53.896 ms	0 - 489 MB	NPU
LaMa-Dilated	ONNX	float	Qualcomm® QCS8550 (Proxy)	79.295 ms	10 - 14 MB	NPU
LaMa-Dilated	ONNX	float	Qualcomm® QCS9075	117.484 ms	6 - 13 MB	NPU
LaMa-Dilated	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	41.915 ms	0 - 409 MB	NPU
LaMa-Dilated	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	34.395 ms	10 - 398 MB	NPU
LaMa-Dilated	ONNX	w8a16	Snapdragon® X2 Elite	28.023 ms	50 - 50 MB	NPU
LaMa-Dilated	ONNX	w8a16	Snapdragon® X Elite	66.547 ms	50 - 50 MB	NPU
LaMa-Dilated	ONNX	w8a16	Snapdragon® 8 Gen 3 Mobile	44.935 ms	0 - 719 MB	NPU
LaMa-Dilated	ONNX	w8a16	Qualcomm® QCS6490	9901.901 ms	423 - 433 MB	CPU
LaMa-Dilated	ONNX	w8a16	Qualcomm® QCS8550 (Proxy)	64.794 ms	0 - 57 MB	NPU
LaMa-Dilated	ONNX	w8a16	Qualcomm® QCS9075	71.715 ms	3 - 8 MB	NPU
LaMa-Dilated	ONNX	w8a16	Qualcomm® QCM6690	4498.097 ms	371 - 381 MB	CPU
LaMa-Dilated	ONNX	w8a16	Snapdragon® 8 Elite For Galaxy Mobile	32.936 ms	0 - 464 MB	NPU
LaMa-Dilated	ONNX	w8a16	Snapdragon® 7 Gen 4 Mobile	4516.794 ms	359 - 370 MB	CPU
LaMa-Dilated	ONNX	w8a16	Snapdragon® 8 Elite Gen 5 Mobile	26.484 ms	1 - 537 MB	NPU
LaMa-Dilated	QNN_DLC	float	Snapdragon® X2 Elite	34.456 ms	4 - 4 MB	NPU
LaMa-Dilated	QNN_DLC	float	Snapdragon® X Elite	75.599 ms	4 - 4 MB	NPU
LaMa-Dilated	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	53.053 ms	6 - 474 MB	NPU
LaMa-Dilated	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	403.503 ms	1 - 360 MB	NPU
LaMa-Dilated	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	75.546 ms	4 - 7 MB	NPU
LaMa-Dilated	QNN_DLC	float	Qualcomm® SA8775P	107.107 ms	1 - 360 MB	NPU
LaMa-Dilated	QNN_DLC	float	Qualcomm® QCS9075	119.79 ms	4 - 13 MB	NPU
LaMa-Dilated	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	134.892 ms	3 - 416 MB	NPU
LaMa-Dilated	QNN_DLC	float	Qualcomm® SA7255P	403.503 ms	1 - 360 MB	NPU
LaMa-Dilated	QNN_DLC	float	Qualcomm® SA8295P	113.807 ms	1 - 342 MB	NPU
LaMa-Dilated	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	40.444 ms	1 - 363 MB	NPU
LaMa-Dilated	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	31.962 ms	4 - 390 MB	NPU
LaMa-Dilated	QNN_DLC	w8a16	Snapdragon® X2 Elite	30.619 ms	2 - 2 MB	NPU
LaMa-Dilated	QNN_DLC	w8a16	Snapdragon® X Elite	66.701 ms	2 - 2 MB	NPU
LaMa-Dilated	QNN_DLC	w8a16	Snapdragon® 8 Gen 3 Mobile	44.212 ms	0 - 759 MB	NPU
LaMa-Dilated	QNN_DLC	w8a16	Qualcomm® QCS8275 (Proxy)	132.812 ms	2 - 614 MB	NPU
LaMa-Dilated	QNN_DLC	w8a16	Qualcomm® QCS8550 (Proxy)	64.722 ms	2 - 608 MB	NPU
LaMa-Dilated	QNN_DLC	w8a16	Qualcomm® SA8775P	65.408 ms	2 - 614 MB	NPU
LaMa-Dilated	QNN_DLC	w8a16	Qualcomm® QCS9075	79.522 ms	1 - 6 MB	NPU
LaMa-Dilated	QNN_DLC	w8a16	Qualcomm® QCS8450 (Proxy)	118.248 ms	1 - 580 MB	NPU
LaMa-Dilated	QNN_DLC	w8a16	Qualcomm® SA7255P	132.812 ms	2 - 614 MB	NPU
LaMa-Dilated	QNN_DLC	w8a16	Qualcomm® SA8295P	90.411 ms	2 - 434 MB	NPU
LaMa-Dilated	QNN_DLC	w8a16	Snapdragon® 8 Elite For Galaxy Mobile	32.534 ms	2 - 616 MB	NPU
LaMa-Dilated	QNN_DLC	w8a16	Snapdragon® 8 Elite Gen 5 Mobile	27.787 ms	2 - 765 MB	NPU
LaMa-Dilated	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	53.129 ms	2 - 556 MB	NPU
LaMa-Dilated	TFLITE	float	Qualcomm® QCS8275 (Proxy)	404.052 ms	3 - 375 MB	NPU
LaMa-Dilated	TFLITE	float	Qualcomm® QCS8550 (Proxy)	74.284 ms	0 - 49 MB	NPU
LaMa-Dilated	TFLITE	float	Qualcomm® SA8775P	495.76 ms	3 - 375 MB	NPU
LaMa-Dilated	TFLITE	float	Qualcomm® QCS9075	119.423 ms	2 - 104 MB	NPU
LaMa-Dilated	TFLITE	float	Qualcomm® QCS8450 (Proxy)	136.832 ms	4 - 498 MB	NPU
LaMa-Dilated	TFLITE	float	Qualcomm® SA7255P	404.052 ms	3 - 375 MB	NPU
LaMa-Dilated	TFLITE	float	Qualcomm® SA8295P	113.848 ms	3 - 357 MB	NPU
LaMa-Dilated	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	40.722 ms	2 - 379 MB	NPU
LaMa-Dilated	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	32.155 ms	3 - 401 MB	NPU

License

• The license for the original implementation of LaMa-Dilated can be found here.

References

• Resolution-robust Large Mask Inpainting with Fourier Convolutions
• 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.