qualcomm

Real-ESRGAN-x4plus: Optimized for Mobile Deployment Upscale images and remove image noise Real-ESRGAN is a machine learning model that upscales an image with minimal loss in quality. The implementation is a derivative of the Real-ESRGAN-x4plus architecture, a larger and more powerful version compared to the Real-ESRGAN-general-x4v3 architecture. This model is an implementation of Real-ESRGAN-x4plus found here. This repository provides scripts to run Real-ESRGAN-x4plus on Qualcomm® devices. More details on model performance across various devices, can be found here. - Model Type: Modelusecase.superresolution - Model Stats: - Model checkpoint: RealESRGANx4plus - Input resolution: 128x128 - Number of parameters: 16.7M - Model size (float): 63.9 MB - Model size (w8a8): 16.7 MB | Model | Precision | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit | Target Model |---|---|---|---|---|---|---|---|---| | Real-ESRGAN-x4plus | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 454.901 ms | 0 - 190 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNNDLC | 448.882 ms | 0 - 143 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 143.893 ms | 3 - 160 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNNDLC | 114.996 ms | 0 - 170 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 72.958 ms | 3 - 48 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNNDLC | 70.37 ms | 0 - 44 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | ONNX | 69.886 ms | 6 - 51 MB | NPU | Real-ESRGAN-x4plus.onnx.zip | | Real-ESRGAN-x4plus | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 109.238 ms | 3 - 194 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNNDLC | 105.288 ms | 0 - 147 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | float | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 454.901 ms | 0 - 190 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | float | SA7255P ADP | Qualcomm® SA7255P | QNNDLC | 448.882 ms | 0 - 143 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 71.881 ms | 3 - 49 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNNDLC | 65.396 ms | 0 - 45 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | float | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 113.967 ms | 3 - 158 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | float | SA8295P ADP | Qualcomm® SA8295P | QNNDLC | 110.382 ms | 0 - 158 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 70.405 ms | 3 - 49 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNNDLC | 63.697 ms | 0 - 35 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | float | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 109.238 ms | 3 - 194 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | float | SA8775P ADP | Qualcomm® SA8775P | QNNDLC | 105.288 ms | 0 - 147 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 52.043 ms | 0 - 198 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNNDLC | 50.358 ms | 0 - 154 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 51.204 ms | 8 - 191 MB | NPU | Real-ESRGAN-x4plus.onnx.zip | | Real-ESRGAN-x4plus | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | TFLITE | 41.181 ms | 3 - 198 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNNDLC | 38.981 ms | 0 - 139 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | ONNX | 39.939 ms | 8 - 154 MB | NPU | Real-ESRGAN-x4plus.onnx.zip | | Real-ESRGAN-x4plus | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | TFLITE | 33.773 ms | 3 - 199 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | QNNDLC | 27.546 ms | 0 - 142 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | ONNX | 29.888 ms | 6 - 154 MB | NPU | Real-ESRGAN-x4plus.onnx.zip | | Real-ESRGAN-x4plus | float | Snapdragon X Elite CRD | Snapdragon® X Elite | QNNDLC | 64.985 ms | 121 - 121 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | float | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 65.641 ms | 37 - 37 MB | NPU | Real-ESRGAN-x4plus.onnx.zip | | Real-ESRGAN-x4plus | w8a8 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 73.781 ms | 1 - 176 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | w8a8 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNNDLC | 66.556 ms | 0 - 196 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | w8a8 | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 36.692 ms | 1 - 176 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | w8a8 | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNNDLC | 40.874 ms | 0 - 202 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 24.994 ms | 0 - 34 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNNDLC | 23.357 ms | 0 - 48 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | w8a8 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 23.57 ms | 1 - 175 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | w8a8 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNNDLC | 21.273 ms | 0 - 194 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | w8a8 | RB3 Gen 2 (Proxy) | Qualcomm® QCS6490 (Proxy) | TFLITE | 115.528 ms | 1 - 171 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | w8a8 | RB5 (Proxy) | Qualcomm® QCS8250 (Proxy) | TFLITE | 1373.118 ms | 0 - 96 MB | GPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | w8a8 | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 73.781 ms | 1 - 176 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | w8a8 | SA7255P ADP | Qualcomm® SA7255P | QNNDLC | 66.556 ms | 0 - 196 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | w8a8 | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 23.093 ms | 0 - 39 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | w8a8 | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNNDLC | 23.347 ms | 0 - 58 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | w8a8 | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 39.446 ms | 1 - 172 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | w8a8 | SA8295P ADP | Qualcomm® SA8295P | QNNDLC | 34.932 ms | 0 - 199 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | w8a8 | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 24.767 ms | 0 - 47 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | w8a8 | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNNDLC | 23.349 ms | 0 - 52 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | w8a8 | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 23.57 ms | 1 - 175 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | w8a8 | SA8775P ADP | Qualcomm® SA8775P | QNNDLC | 21.273 ms | 0 - 194 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 17.589 ms | 1 - 179 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNNDLC | 16.054 ms | 0 - 192 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | w8a8 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | TFLITE | 15.027 ms | 1 - 175 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | w8a8 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNNDLC | 11.972 ms | 0 - 177 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | w8a8 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | TFLITE | 11.63 ms | 1 - 176 MB | NPU | Real-ESRGAN-x4plus.tflite | | Real-ESRGAN-x4plus | w8a8 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | QNNDLC | 9.404 ms | 0 - 184 MB | NPU | Real-ESRGAN-x4plus.dlc | | Real-ESRGAN-x4plus | w8a8 | Snapdragon X Elite CRD | Snapdragon® X Elite | QNNDLC | 24.398 ms | 108 - 108 MB | NPU | Real-ESRGAN-x4plus.dlc | Install the package via pip: bash qai-hub configure --apitoken APITOKEN bash python -m qaihubmodels.models.realesrganx4plus.demo %run -m qaihubmodels.models.realesrganx4plus.demo bash python -m qaihubmodels.models.realesrganx4plus.export python import torch import qaihub as hub from qaihubmodels.models.realesrganx4plus import Model Trace model inputshape = torchmodel.getinputspec() sampleinputs = torchmodel.sampleinputs() ptmodel = torch.jit.trace(torchmodel, [torch.tensor(data[0]) for , data in sampleinputs.items()]) Compile model on a specific device compilejob = hub.submitcompilejob( model=ptmodel, device=device, inputspecs=torchmodel.getinputspec(), ) Get target model to run on-device targetmodel = compilejob.gettargetmodel() python profilejob = hub.submitprofilejob( model=targetmodel, device=device, ) python inputdata = torchmodel.sampleinputs() inferencejob = hub.submitinferencejob( model=targetmodel, device=device, inputs=inputdata, ) ondeviceoutput = inferencejob.downloadoutputdata() bash python -m qaihubmodels.models.realesrganx4plus.demo --eval-mode on-device %run -m qaihubmodels.models.realesrganx4plus.demo -- --eval-mode on-device ``` The models can be deployed using multiple runtimes: - TensorFlow Lite (`.tflite` export): This tutorial provides a guide to deploy the .tflite model in an Android application. - QNN (`.so` export ): This sample app provides instructions on how to use the `.so` shared library in an Android application. View on Qualcomm® AI Hub Get more details on Real-ESRGAN-x4plus's performance across various devices here. Explore all available models on Qualcomm® AI Hub License The license for the original implementation of Real-ESRGAN-x4plus can be found here. The license for the compiled assets for on-device deployment can be found here References Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data 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.

YamNet: Optimized for Mobile Deployment Audio Event classification Model An audio event classifier trained on the AudioSet dataset to predict audio events from the AudioSet ontology employing the Mobilenetv1 depthwise-separable convolution architecture. This model is an implementation of YamNet found here. This repository provides scripts to run YamNet on Qualcomm® devices. More details on model performance across various devices, can be found here. - Model Type: Modelusecase.audioclassification - Model Stats: - Model checkpoint: yamnet.pth - Input resolution: 1x1x96x64 - Number of parameters: 3.73M - Model size (float): 14.2 MB | Model | Precision | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit | Target Model |---|---|---|---|---|---|---|---|---| | YamNet | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 0.664 ms | 0 - 23 MB | NPU | YamNet.tflite | | YamNet | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNNDLC | 0.644 ms | 0 - 16 MB | NPU | YamNet.dlc | | YamNet | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 0.323 ms | 0 - 38 MB | NPU | YamNet.tflite | | YamNet | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNNDLC | 0.355 ms | 0 - 25 MB | NPU | YamNet.dlc | | YamNet | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 0.213 ms | 0 - 72 MB | NPU | YamNet.tflite | | YamNet | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNNDLC | 0.216 ms | 0 - 48 MB | NPU | YamNet.dlc | | YamNet | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | ONNX | 0.393 ms | 0 - 51 MB | NPU | YamNet.onnx.zip | | YamNet | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 0.368 ms | 0 - 23 MB | NPU | YamNet.tflite | | YamNet | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNNDLC | 1.096 ms | 0 - 16 MB | NPU | YamNet.dlc | | YamNet | float | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 0.664 ms | 0 - 23 MB | NPU | YamNet.tflite | | YamNet | float | SA7255P ADP | Qualcomm® SA7255P | QNNDLC | 0.644 ms | 0 - 16 MB | NPU | YamNet.dlc | | YamNet | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 0.22 ms | 0 - 71 MB | NPU | YamNet.tflite | | YamNet | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNNDLC | 0.223 ms | 0 - 50 MB | NPU | YamNet.dlc | | YamNet | float | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 0.539 ms | 0 - 29 MB | NPU | YamNet.tflite | | YamNet | float | SA8295P ADP | Qualcomm® SA8295P | QNNDLC | 0.528 ms | 0 - 22 MB | NPU | YamNet.dlc | | YamNet | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 0.221 ms | 0 - 70 MB | NPU | YamNet.tflite | | YamNet | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNNDLC | 0.222 ms | 0 - 4 MB | NPU | YamNet.dlc | | YamNet | float | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 0.368 ms | 0 - 23 MB | NPU | YamNet.tflite | | YamNet | float | SA8775P ADP | Qualcomm® SA8775P | QNNDLC | 1.096 ms | 0 - 16 MB | NPU | YamNet.dlc | | YamNet | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 0.173 ms | 10 - 48 MB | NPU | YamNet.tflite | | YamNet | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNNDLC | 0.176 ms | 0 - 29 MB | NPU | YamNet.dlc | | YamNet | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 0.259 ms | 0 - 26 MB | NPU | YamNet.onnx.zip | | YamNet | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | TFLITE | 0.153 ms | 0 - 31 MB | NPU | YamNet.tflite | | YamNet | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNNDLC | 0.15 ms | 0 - 24 MB | NPU | YamNet.dlc | | YamNet | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | ONNX | 0.259 ms | 0 - 23 MB | NPU | YamNet.onnx.zip | | YamNet | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | TFLITE | 0.151 ms | 0 - 30 MB | NPU | YamNet.tflite | | YamNet | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | QNNDLC | 0.153 ms | 0 - 22 MB | NPU | YamNet.dlc | | YamNet | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | ONNX | 0.263 ms | 0 - 19 MB | NPU | YamNet.onnx.zip | | YamNet | float | Snapdragon X Elite CRD | Snapdragon® X Elite | QNNDLC | 0.28 ms | 56 - 56 MB | NPU | YamNet.dlc | | YamNet | float | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 0.286 ms | 8 - 8 MB | NPU | YamNet.onnx.zip | Install the package via pip: bash qai-hub configure --apitoken APITOKEN bash python -m qaihubmodels.models.yamnet.demo %run -m qaihubmodels.models.yamnet.demo bash python -m qaihubmodels.models.yamnet.export python import torch import qaihub as hub from qaihubmodels.models.yamnet import Model Trace model inputshape = torchmodel.getinputspec() sampleinputs = torchmodel.sampleinputs() ptmodel = torch.jit.trace(torchmodel, [torch.tensor(data[0]) for , data in sampleinputs.items()]) Compile model on a specific device compilejob = hub.submitcompilejob( model=ptmodel, device=device, inputspecs=torchmodel.getinputspec(), ) Get target model to run on-device targetmodel = compilejob.gettargetmodel() python profilejob = hub.submitprofilejob( model=targetmodel, device=device, ) python inputdata = torchmodel.sampleinputs() inferencejob = hub.submitinferencejob( model=targetmodel, device=device, inputs=inputdata, ) ondeviceoutput = inferencejob.downloadoutputdata() bash python -m qaihubmodels.models.yamnet.demo --eval-mode on-device %run -m qaihubmodels.models.yamnet.demo -- --eval-mode on-device ``` The models can be deployed using multiple runtimes: - TensorFlow Lite (`.tflite` export): This tutorial provides a guide to deploy the .tflite model in an Android application. - QNN (`.so` export ): This sample app provides instructions on how to use the `.so` shared library in an Android application. View on Qualcomm® AI Hub Get more details on YamNet's performance across various devices here. Explore all available models on Qualcomm® AI Hub License The license for the original implementation of YamNet can be found here. The license for the compiled assets for on-device deployment can be found here References MobileNets Efficient Convolutional Neural Networks for Mobile Vision Applications 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.

XLSR: Optimized for Mobile Deployment Upscale images in real time XLSR is designed for lightweight real-time upscaling of images. This model is an implementation of XLSR found here. This repository provides scripts to run XLSR on Qualcomm® devices. More details on model performance across various devices, can be found here. - Model Type: Modelusecase.superresolution - Model Stats: - Model checkpoint: xlsr3xcheckpoint - Input resolution: 128x128 - Number of parameters: 28.0K - Model size (float): 115 KB - Model size (w8a8): 45.6 KB | Model | Precision | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit | Target Model |---|---|---|---|---|---|---|---|---| | XLSR | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 4.924 ms | 0 - 15 MB | NPU | XLSR.tflite | | XLSR | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNNDLC | 2.147 ms | 0 - 15 MB | NPU | XLSR.dlc | | XLSR | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 2.493 ms | 3 - 33 MB | NPU | XLSR.tflite | | XLSR | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNNDLC | 1.059 ms | 0 - 29 MB | NPU | XLSR.dlc | | XLSR | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 2.302 ms | 0 - 7 MB | NPU | XLSR.tflite | | XLSR | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNNDLC | 0.765 ms | 0 - 6 MB | NPU | XLSR.dlc | | XLSR | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | ONNX | 1.173 ms | 0 - 8 MB | NPU | XLSR.onnx.zip | | XLSR | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 2.769 ms | 0 - 15 MB | NPU | XLSR.tflite | | XLSR | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNNDLC | 1.144 ms | 0 - 15 MB | NPU | XLSR.dlc | | XLSR | float | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 4.924 ms | 0 - 15 MB | NPU | XLSR.tflite | | XLSR | float | SA7255P ADP | Qualcomm® SA7255P | QNNDLC | 2.147 ms | 0 - 15 MB | NPU | XLSR.dlc | | XLSR | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 2.302 ms | 0 - 7 MB | NPU | XLSR.tflite | | XLSR | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNNDLC | 0.766 ms | 0 - 4 MB | NPU | XLSR.dlc | | XLSR | float | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 3.163 ms | 0 - 22 MB | NPU | XLSR.tflite | | XLSR | float | SA8295P ADP | Qualcomm® SA8295P | QNNDLC | 1.357 ms | 0 - 21 MB | NPU | XLSR.dlc | | XLSR | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 2.305 ms | 0 - 7 MB | NPU | XLSR.tflite | | XLSR | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNNDLC | 0.784 ms | 0 - 4 MB | NPU | XLSR.dlc | | XLSR | float | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 2.769 ms | 0 - 15 MB | NPU | XLSR.tflite | | XLSR | float | SA8775P ADP | Qualcomm® SA8775P | QNNDLC | 1.144 ms | 0 - 15 MB | NPU | XLSR.dlc | | XLSR | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 1.461 ms | 0 - 25 MB | NPU | XLSR.tflite | | XLSR | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNNDLC | 0.468 ms | 0 - 25 MB | NPU | XLSR.dlc | | XLSR | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 0.761 ms | 0 - 27 MB | NPU | XLSR.onnx.zip | | XLSR | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | TFLITE | 1.109 ms | 0 - 22 MB | NPU | XLSR.tflite | | XLSR | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNNDLC | 0.35 ms | 0 - 24 MB | NPU | XLSR.dlc | | XLSR | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | ONNX | 0.583 ms | 0 - 21 MB | NPU | XLSR.onnx.zip | | XLSR | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | TFLITE | 0.871 ms | 0 - 19 MB | NPU | XLSR.tflite | | XLSR | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | QNNDLC | 0.349 ms | 0 - 19 MB | NPU | XLSR.dlc | | XLSR | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | ONNX | 0.531 ms | 3 - 20 MB | NPU | XLSR.onnx.zip | | XLSR | float | Snapdragon X Elite CRD | Snapdragon® X Elite | QNNDLC | 0.914 ms | 0 - 0 MB | NPU | XLSR.dlc | | XLSR | float | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 1.093 ms | 8 - 8 MB | NPU | XLSR.onnx.zip | | XLSR | w8a8 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 1.036 ms | 1 - 17 MB | NPU | XLSR.tflite | | XLSR | w8a8 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNNDLC | 0.915 ms | 0 - 15 MB | NPU | XLSR.dlc | | XLSR | w8a8 | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 0.529 ms | 1 - 30 MB | NPU | XLSR.tflite | | XLSR | w8a8 | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNNDLC | 0.568 ms | 0 - 29 MB | NPU | XLSR.dlc | | XLSR | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 0.436 ms | 0 - 9 MB | NPU | XLSR.tflite | | XLSR | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNNDLC | 0.368 ms | 0 - 3 MB | NPU | XLSR.dlc | | XLSR | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | ONNX | 1.525 ms | 0 - 17 MB | NPU | XLSR.onnx.zip | | XLSR | w8a8 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 0.662 ms | 0 - 16 MB | NPU | XLSR.tflite | | XLSR | w8a8 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNNDLC | 2.101 ms | 0 - 15 MB | NPU | XLSR.dlc | | XLSR | w8a8 | RB3 Gen 2 (Proxy) | Qualcomm® QCS6490 (Proxy) | TFLITE | 9.01 ms | 4 - 33 MB | GPU | XLSR.tflite | | XLSR | w8a8 | RB3 Gen 2 (Proxy) | Qualcomm® QCS6490 (Proxy) | QNNDLC | 1.052 ms | 0 - 20 MB | NPU | XLSR.dlc | | XLSR | w8a8 | RB3 Gen 2 (Proxy) | Qualcomm® QCS6490 (Proxy) | ONNX | 9.311 ms | 18 - 29 MB | CPU | XLSR.onnx.zip | | XLSR | w8a8 | RB5 (Proxy) | Qualcomm® QCS8250 (Proxy) | TFLITE | 5.294 ms | 4 - 11 MB | GPU | XLSR.tflite | | XLSR | w8a8 | RB5 (Proxy) | Qualcomm® QCS8250 (Proxy) | ONNX | 7.126 ms | 18 - 21 MB | CPU | XLSR.onnx.zip | | XLSR | w8a8 | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 1.036 ms | 1 - 17 MB | NPU | XLSR.tflite | | XLSR | w8a8 | SA7255P ADP | Qualcomm® SA7255P | QNNDLC | 0.915 ms | 0 - 15 MB | NPU | XLSR.dlc | | XLSR | w8a8 | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 0.438 ms | 0 - 4 MB | NPU | XLSR.tflite | | XLSR | w8a8 | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNNDLC | 0.372 ms | 0 - 11 MB | NPU | XLSR.dlc | | XLSR | w8a8 | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 0.863 ms | 0 - 21 MB | NPU | XLSR.tflite | | XLSR | w8a8 | SA8295P ADP | Qualcomm® SA8295P | QNNDLC | 0.772 ms | 0 - 21 MB | NPU | XLSR.dlc | | XLSR | w8a8 | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 0.439 ms | 0 - 10 MB | NPU | XLSR.tflite | | XLSR | w8a8 | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNNDLC | 0.37 ms | 0 - 4 MB | NPU | XLSR.dlc | | XLSR | w8a8 | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 0.662 ms | 0 - 16 MB | NPU | XLSR.tflite | | XLSR | w8a8 | SA8775P ADP | Qualcomm® SA8775P | QNNDLC | 2.101 ms | 0 - 15 MB | NPU | XLSR.dlc | | XLSR | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 0.276 ms | 0 - 23 MB | NPU | XLSR.tflite | | XLSR | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNNDLC | 0.241 ms | 0 - 23 MB | NPU | XLSR.dlc | | XLSR | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 0.982 ms | 0 - 25 MB | NPU | XLSR.onnx.zip | | XLSR | w8a8 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | TFLITE | 0.272 ms | 0 - 22 MB | NPU | XLSR.tflite | | XLSR | w8a8 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNNDLC | 0.187 ms | 0 - 18 MB | NPU | XLSR.dlc | | XLSR | w8a8 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | ONNX | 0.775 ms | 0 - 19 MB | NPU | XLSR.onnx.zip | | XLSR | w8a8 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | TFLITE | 0.267 ms | 0 - 18 MB | NPU | XLSR.tflite | | XLSR | w8a8 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | QNNDLC | 0.17 ms | 0 - 17 MB | NPU | XLSR.dlc | | XLSR | w8a8 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | ONNX | 0.712 ms | 3 - 22 MB | NPU | XLSR.onnx.zip | | XLSR | w8a8 | Snapdragon X Elite CRD | Snapdragon® X Elite | QNNDLC | 0.475 ms | 4 - 4 MB | NPU | XLSR.dlc | | XLSR | w8a8 | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 1.525 ms | 9 - 9 MB | NPU | XLSR.onnx.zip | Configure Qualcomm® AI Hub to run this model on a cloud-hosted device Sign-in to Qualcomm® AI Hub with your Qualcomm® ID. Once signed in navigate to `Account -> Settings -> API Token`. With this API token, you can configure your client to run models on the cloud hosted devices. The package contains a simple end-to-end demo that downloads pre-trained weights and runs this model on a sample input. The above demo runs a reference implementation of pre-processing, model inference, and post processing. NOTE: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above). In addition to the demo, you can also run the model on a cloud-hosted Qualcomm® device. This script does the following: Performance check on-device on a cloud-hosted device Downloads compiled assets that can be deployed on-device for Android. Accuracy check between PyTorch and on-device outputs. This export script leverages Qualcomm® AI Hub to optimize, validate, and deploy this model on-device. Lets go through each step below in detail: To compile a PyTorch model for on-device deployment, we first trace the model in memory using the `jit.trace` and then call the `submitcompilejob` API. Step 2: Performance profiling on cloud-hosted device After compiling models from step 1. Models can be profiled model on-device using the `targetmodel`. Note that this scripts runs the model on a device automatically provisioned in the cloud. Once the job is submitted, you can navigate to a provided job URL to view a variety of on-device performance metrics. To verify the accuracy of the model on-device, you can run on-device inference on sample input data on the same cloud hosted device. With the output of the model, you can compute like PSNR, relative errors or spot check the output with expected output. Note: This on-device profiling and inference requires access to Qualcomm® AI Hub. Sign up for access. NOTE: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above). The models can be deployed using multiple runtimes: - TensorFlow Lite (`.tflite` export): This tutorial provides a guide to deploy the .tflite model in an Android application. - QNN (`.so` export ): This sample app provides instructions on how to use the `.so` shared library in an Android application. View on Qualcomm® AI Hub Get more details on XLSR's performance across various devices here. Explore all available models on Qualcomm® AI Hub License The license for the original implementation of XLSR can be found here. The license for the compiled assets for on-device deployment can be found here References Extremely Lightweight Quantization Robust Real-Time Single-Image Super Resolution for Mobile Devices 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.

YOLOv8-Detection: Optimized for Mobile Deployment Real-time object detection optimized for mobile and edge by Ultralytics Ultralytics YOLOv8 is a machine learning model that predicts bounding boxes and classes of objects in an image. This model is an implementation of YOLOv8-Detection found here. This repository provides scripts to run YOLOv8-Detection on Qualcomm® devices. More details on model performance across various devices, can be found here. WARNING: The model assets are not readily available for download due to licensing restrictions. - Model Type: Modelusecase.objectdetection - Model Stats: - Model checkpoint: YOLOv8-N - Input resolution: 640x640 - Number of parameters: 3.18M - Model size (float): 12.2 MB - Model size (w8a8): 3.25 MB - Model size (w8a16): 3.60 MB | Model | Precision | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit | Target Model |---|---|---|---|---|---|---|---|---| | YOLOv8-Detection | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 13.467 ms | 0 - 81 MB | NPU | -- | | YOLOv8-Detection | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNNDLC | 13.009 ms | 0 - 120 MB | NPU | -- | | YOLOv8-Detection | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 6.131 ms | 0 - 40 MB | NPU | -- | | YOLOv8-Detection | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNNDLC | 7.56 ms | 5 - 41 MB | NPU | -- | | YOLOv8-Detection | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 3.473 ms | 0 - 82 MB | NPU | -- | | YOLOv8-Detection | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNNDLC | 3.301 ms | 0 - 66 MB | NPU | -- | | YOLOv8-Detection | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | ONNX | 5.338 ms | 0 - 59 MB | NPU | -- | | YOLOv8-Detection | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 20.509 ms | 0 - 81 MB | NPU | -- | | YOLOv8-Detection | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNNDLC | 4.904 ms | 1 - 105 MB | NPU | -- | | YOLOv8-Detection | float | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 13.467 ms | 0 - 81 MB | NPU | -- | | YOLOv8-Detection | float | SA7255P ADP | Qualcomm® SA7255P | QNNDLC | 13.009 ms | 0 - 120 MB | NPU | -- | | YOLOv8-Detection | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 3.488 ms | 0 - 88 MB | NPU | -- | | YOLOv8-Detection | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNNDLC | 3.287 ms | 5 - 83 MB | NPU | -- | | YOLOv8-Detection | float | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 7.219 ms | 0 - 34 MB | NPU | -- | | YOLOv8-Detection | float | SA8295P ADP | Qualcomm® SA8295P | QNNDLC | 7.175 ms | 0 - 32 MB | NPU | -- | | YOLOv8-Detection | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 3.484 ms | 0 - 87 MB | NPU | -- | | YOLOv8-Detection | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNNDLC | 3.296 ms | 5 - 79 MB | NPU | -- | | YOLOv8-Detection | float | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 20.509 ms | 0 - 81 MB | NPU | -- | | YOLOv8-Detection | float | SA8775P ADP | Qualcomm® SA8775P | QNNDLC | 4.904 ms | 1 - 105 MB | NPU | -- | | YOLOv8-Detection | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 2.586 ms | 0 - 162 MB | NPU | -- | | YOLOv8-Detection | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNNDLC | 2.453 ms | 5 - 261 MB | NPU | -- | | YOLOv8-Detection | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 3.454 ms | 0 - 102 MB | NPU | -- | | YOLOv8-Detection | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | TFLITE | 2.039 ms | 0 - 87 MB | NPU | -- | | YOLOv8-Detection | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNNDLC | 1.862 ms | 5 - 107 MB | NPU | -- | | YOLOv8-Detection | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | ONNX | 2.905 ms | 2 - 85 MB | NPU | -- | | YOLOv8-Detection | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | TFLITE | 1.565 ms | 0 - 82 MB | NPU | -- | | YOLOv8-Detection | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | QNNDLC | 1.461 ms | 5 - 128 MB | NPU | -- | | YOLOv8-Detection | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | ONNX | 2.422 ms | 1 - 69 MB | NPU | -- | | YOLOv8-Detection | float | Snapdragon X Elite CRD | Snapdragon® X Elite | QNNDLC | 3.702 ms | 141 - 141 MB | NPU | -- | | YOLOv8-Detection | float | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 5.607 ms | 5 - 5 MB | NPU | -- | | YOLOv8-Detection | w8a16 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNNDLC | 6.289 ms | 2 - 32 MB | NPU | -- | | YOLOv8-Detection | w8a16 | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNNDLC | 3.948 ms | 2 - 43 MB | NPU | -- | | YOLOv8-Detection | w8a16 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNNDLC | 3.173 ms | 2 - 13 MB | NPU | -- | | YOLOv8-Detection | w8a16 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | ONNX | 5.112 ms | 1 - 56 MB | NPU | -- | | YOLOv8-Detection | w8a16 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNNDLC | 3.762 ms | 2 - 31 MB | NPU | -- | | YOLOv8-Detection | w8a16 | RB3 Gen 2 (Proxy) | Qualcomm® QCS6490 (Proxy) | ONNX | 164.771 ms | 65 - 79 MB | CPU | -- | | YOLOv8-Detection | w8a16 | RB5 (Proxy) | Qualcomm® QCS8250 (Proxy) | ONNX | 136.502 ms | 62 - 66 MB | CPU | -- | | YOLOv8-Detection | w8a16 | SA7255P ADP | Qualcomm® SA7255P | QNNDLC | 6.289 ms | 2 - 32 MB | NPU | -- | | YOLOv8-Detection | w8a16 | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNNDLC | 3.18 ms | 1 - 13 MB | NPU | -- | | YOLOv8-Detection | w8a16 | SA8295P ADP | Qualcomm® SA8295P | QNNDLC | 4.345 ms | 2 - 41 MB | NPU | -- | | YOLOv8-Detection | w8a16 | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNNDLC | 3.185 ms | 2 - 13 MB | NPU | -- | | YOLOv8-Detection | w8a16 | SA8775P ADP | Qualcomm® SA8775P | QNNDLC | 3.762 ms | 2 - 31 MB | NPU | -- | | YOLOv8-Detection | w8a16 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNNDLC | 2.105 ms | 2 - 38 MB | NPU | -- | | YOLOv8-Detection | w8a16 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 3.068 ms | 0 - 106 MB | NPU | -- | | YOLOv8-Detection | w8a16 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNNDLC | 1.476 ms | 2 - 38 MB | NPU | -- | | YOLOv8-Detection | w8a16 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | ONNX | 2.392 ms | 0 - 76 MB | NPU | -- | | YOLOv8-Detection | w8a16 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | QNNDLC | 1.201 ms | 2 - 40 MB | NPU | -- | | YOLOv8-Detection | w8a16 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | ONNX | 2.147 ms | 1 - 78 MB | NPU | -- | | YOLOv8-Detection | w8a16 | Snapdragon X Elite CRD | Snapdragon® X Elite | QNNDLC | 3.568 ms | 5 - 5 MB | NPU | -- | | YOLOv8-Detection | w8a16 | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 5.257 ms | 2 - 2 MB | NPU | -- | | YOLOv8-Detection | w8a8 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 3.182 ms | 0 - 24 MB | NPU | -- | | YOLOv8-Detection | w8a8 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNNDLC | 3.047 ms | 1 - 25 MB | NPU | -- | | YOLOv8-Detection | w8a8 | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 1.585 ms | 0 - 35 MB | NPU | -- | | YOLOv8-Detection | w8a8 | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNNDLC | 1.595 ms | 1 - 33 MB | NPU | -- | | YOLOv8-Detection | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 1.416 ms | 0 - 14 MB | NPU | -- | | YOLOv8-Detection | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNNDLC | 1.308 ms | 1 - 15 MB | NPU | -- | | YOLOv8-Detection | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | ONNX | 2.196 ms | 0 - 45 MB | NPU | -- | | YOLOv8-Detection | w8a8 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 1.865 ms | 0 - 24 MB | NPU | -- | | YOLOv8-Detection | w8a8 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNNDLC | 1.678 ms | 1 - 25 MB | NPU | -- | | YOLOv8-Detection | w8a8 | RB3 Gen 2 (Proxy) | Qualcomm® QCS6490 (Proxy) | TFLITE | 3.697 ms | 0 - 32 MB | NPU | -- | | YOLOv8-Detection | w8a8 | RB3 Gen 2 (Proxy) | Qualcomm® QCS6490 (Proxy) | QNNDLC | 4.409 ms | 0 - 33 MB | NPU | -- | | YOLOv8-Detection | w8a8 | RB3 Gen 2 (Proxy) | Qualcomm® QCS6490 (Proxy) | ONNX | 40.749 ms | 22 - 38 MB | CPU | -- | | YOLOv8-Detection | w8a8 | RB5 (Proxy) | Qualcomm® QCS8250 (Proxy) | TFLITE | 43.901 ms | 3 - 11 MB | NPU | -- | | YOLOv8-Detection | w8a8 | RB5 (Proxy) | Qualcomm® QCS8250 (Proxy) | ONNX | 36.395 ms | 21 - 27 MB | CPU | -- | | YOLOv8-Detection | w8a8 | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 3.182 ms | 0 - 24 MB | NPU | -- | | YOLOv8-Detection | w8a8 | SA7255P ADP | Qualcomm® SA7255P | QNNDLC | 3.047 ms | 1 - 25 MB | NPU | -- | | YOLOv8-Detection | w8a8 | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 1.41 ms | 0 - 15 MB | NPU | -- | | YOLOv8-Detection | w8a8 | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNNDLC | 1.311 ms | 1 - 16 MB | NPU | -- | | YOLOv8-Detection | w8a8 | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 2.245 ms | 0 - 31 MB | NPU | -- | | YOLOv8-Detection | w8a8 | SA8295P ADP | Qualcomm® SA8295P | QNNDLC | 2.073 ms | 1 - 31 MB | NPU | -- | | YOLOv8-Detection | w8a8 | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 1.412 ms | 0 - 15 MB | NPU | -- | | YOLOv8-Detection | w8a8 | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNNDLC | 1.311 ms | 1 - 16 MB | NPU | -- | | YOLOv8-Detection | w8a8 | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 1.865 ms | 0 - 24 MB | NPU | -- | | YOLOv8-Detection | w8a8 | SA8775P ADP | Qualcomm® SA8775P | QNNDLC | 1.678 ms | 1 - 25 MB | NPU | -- | | YOLOv8-Detection | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 0.935 ms | 0 - 36 MB | NPU | -- | | YOLOv8-Detection | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNNDLC | 0.899 ms | 1 - 32 MB | NPU | -- | | YOLOv8-Detection | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 1.465 ms | 0 - 147 MB | NPU | -- | | YOLOv8-Detection | w8a8 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | TFLITE | 0.729 ms | 0 - 28 MB | NPU | -- | | YOLOv8-Detection | w8a8 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNNDLC | 0.644 ms | 1 - 29 MB | NPU | -- | | YOLOv8-Detection | w8a8 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | ONNX | 1.09 ms | 0 - 81 MB | NPU | -- | | YOLOv8-Detection | w8a8 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | TFLITE | 0.652 ms | 0 - 29 MB | NPU | -- | | YOLOv8-Detection | w8a8 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | QNNDLC | 0.583 ms | 1 - 35 MB | NPU | -- | | YOLOv8-Detection | w8a8 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | ONNX | 0.956 ms | 0 - 85 MB | NPU | -- | | YOLOv8-Detection | w8a8 | Snapdragon X Elite CRD | Snapdragon® X Elite | QNNDLC | 1.546 ms | 2 - 2 MB | NPU | -- | | YOLOv8-Detection | w8a8 | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 2.192 ms | 2 - 2 MB | NPU | -- | | YOLOv8-Detection | w8a8mixedint16 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNNDLC | 4.367 ms | 1 - 25 MB | NPU | -- | | YOLOv8-Detection | w8a8mixedint16 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNNDLC | 2.035 ms | 2 - 12 MB | NPU | -- | | YOLOv8-Detection | w8a8mixedint16 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNNDLC | 2.544 ms | 1 - 25 MB | NPU | -- | | YOLOv8-Detection | w8a8mixedint16 | SA7255P ADP | Qualcomm® SA7255P | QNNDLC | 4.367 ms | 1 - 25 MB | NPU | -- | | YOLOv8-Detection | w8a8mixedint16 | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNNDLC | 2.037 ms | 2 - 12 MB | NPU | -- | | YOLOv8-Detection | w8a8mixedint16 | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNNDLC | 2.04 ms | 2 - 12 MB | NPU | -- | | YOLOv8-Detection | w8a8mixedint16 | SA8775P ADP | Qualcomm® SA8775P | QNNDLC | 2.544 ms | 1 - 25 MB | NPU | -- | | YOLOv8-Detection | w8a8mixedint16 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNNDLC | 1.369 ms | 2 - 38 MB | NPU | -- | | YOLOv8-Detection | w8a8mixedint16 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNNDLC | 0.984 ms | 2 - 36 MB | NPU | -- | | YOLOv8-Detection | w8a8mixedint16 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | QNNDLC | 0.819 ms | 2 - 32 MB | NPU | -- | | YOLOv8-Detection | w8a8mixedint16 | Snapdragon X Elite CRD | Snapdragon® X Elite | QNNDLC | 2.318 ms | 2 - 2 MB | NPU | -- | Install the package via pip: bash qai-hub configure --apitoken APITOKEN bash python -m qaihubmodels.models.yolov8det.demo %run -m qaihubmodels.models.yolov8det.demo bash python -m qaihubmodels.models.yolov8det.export python import torch import qaihub as hub from qaihubmodels.models.yolov8det import Model Trace model inputshape = torchmodel.getinputspec() sampleinputs = torchmodel.sampleinputs() ptmodel = torch.jit.trace(torchmodel, [torch.tensor(data[0]) for , data in sampleinputs.items()]) Compile model on a specific device compilejob = hub.submitcompilejob( model=ptmodel, device=device, inputspecs=torchmodel.getinputspec(), ) Get target model to run on-device targetmodel = compilejob.gettargetmodel() python profilejob = hub.submitprofilejob( model=targetmodel, device=device, ) python inputdata = torchmodel.sampleinputs() inferencejob = hub.submitinferencejob( model=targetmodel, device=device, inputs=inputdata, ) ondeviceoutput = inferencejob.downloadoutputdata() bash python -m qaihubmodels.models.yolov8det.demo --eval-mode on-device %run -m qaihubmodels.models.yolov8det.demo -- --eval-mode on-device ``` The models can be deployed using multiple runtimes: - TensorFlow Lite (`.tflite` export): This tutorial provides a guide to deploy the .tflite model in an Android application. - QNN (`.so` export ): This sample app provides instructions on how to use the `.so` shared library in an Android application. View on Qualcomm® AI Hub Get more details on YOLOv8-Detection's performance across various devices here. Explore all available models on Qualcomm® AI Hub License The license for the original implementation of YOLOv8-Detection can be found here. The license for the compiled assets for on-device deployment can be found here References Ultralytics YOLOv8 Docs: Object Detection 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.

MediaPipe-Hand-Detection: Optimized for Mobile Deployment Real-time hand detection optimized for mobile and edge The MediaPipe Hand Landmark Detector is a machine learning pipeline that predicts bounding boxes and pose skeletons of hands in an image. This model is an implementation of MediaPipe-Hand-Detection found here. This repository provides scripts to run MediaPipe-Hand-Detection on Qualcomm® devices. More details on model performance across various devices, can be found here. - Model Type: Modelusecase.objectdetection - Model Stats: - Input resolution: 256x256 - Number of parameters (HandDetector): 1.76M - Model size (HandDetector) (float): 6.75 MB - Number of parameters (HandLandmarkDetector): 2.01M - Model size (HandLandmarkDetector) (float): 7.70 MB | Model | Precision | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit | Target Model |---|---|---|---|---|---|---|---|---| | HandDetector | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 3.823 ms | 0 - 26 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandDetector | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNNDLC | 3.748 ms | 1 - 28 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandDetector | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 1.31 ms | 0 - 30 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandDetector | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNNDLC | 1.374 ms | 1 - 34 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandDetector | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 0.726 ms | 0 - 38 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandDetector | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNNDLC | 0.727 ms | 0 - 34 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandDetector | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | ONNX | 1.113 ms | 0 - 30 MB | NPU | MediaPipe-Hand-Detection.onnx.zip | | HandDetector | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 5.181 ms | 0 - 27 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandDetector | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNNDLC | 5.115 ms | 1 - 28 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandDetector | float | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 3.823 ms | 0 - 26 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandDetector | float | SA7255P ADP | Qualcomm® SA7255P | QNNDLC | 3.748 ms | 1 - 28 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandDetector | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 0.73 ms | 0 - 37 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandDetector | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNNDLC | 0.723 ms | 0 - 34 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandDetector | float | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 1.763 ms | 0 - 28 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandDetector | float | SA8295P ADP | Qualcomm® SA8295P | QNNDLC | 1.708 ms | 0 - 29 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandDetector | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 0.73 ms | 0 - 38 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandDetector | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNNDLC | 0.726 ms | 0 - 35 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandDetector | float | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 5.181 ms | 0 - 27 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandDetector | float | SA8775P ADP | Qualcomm® SA8775P | QNNDLC | 5.115 ms | 1 - 28 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandDetector | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 0.528 ms | 0 - 35 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandDetector | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNNDLC | 0.517 ms | 0 - 38 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandDetector | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 0.736 ms | 0 - 43 MB | NPU | MediaPipe-Hand-Detection.onnx.zip | | HandDetector | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | TFLITE | 0.438 ms | 0 - 34 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandDetector | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNNDLC | 0.431 ms | 1 - 32 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandDetector | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | ONNX | 0.628 ms | 0 - 36 MB | NPU | MediaPipe-Hand-Detection.onnx.zip | | HandDetector | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | TFLITE | 0.373 ms | 0 - 31 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandDetector | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | QNNDLC | 0.377 ms | 21 - 52 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandDetector | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | ONNX | 0.588 ms | 1 - 32 MB | NPU | MediaPipe-Hand-Detection.onnx.zip | | HandDetector | float | Snapdragon X Elite CRD | Snapdragon® X Elite | QNNDLC | 0.884 ms | 26 - 26 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandDetector | float | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 0.987 ms | 3 - 3 MB | NPU | MediaPipe-Hand-Detection.onnx.zip | | HandLandmarkDetector | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 5.403 ms | 0 - 27 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandLandmarkDetector | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNNDLC | 5.292 ms | 1 - 23 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandLandmarkDetector | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 1.796 ms | 0 - 37 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandLandmarkDetector | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNNDLC | 1.94 ms | 1 - 33 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandLandmarkDetector | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 1.022 ms | 0 - 70 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandLandmarkDetector | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNNDLC | 1.0 ms | 1 - 57 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandLandmarkDetector | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | ONNX | 1.421 ms | 8 - 64 MB | NPU | MediaPipe-Hand-Detection.onnx.zip | | HandLandmarkDetector | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 1.911 ms | 0 - 27 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandLandmarkDetector | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNNDLC | 1.873 ms | 1 - 23 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandLandmarkDetector | float | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 5.403 ms | 0 - 27 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandLandmarkDetector | float | SA7255P ADP | Qualcomm® SA7255P | QNNDLC | 5.292 ms | 1 - 23 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandLandmarkDetector | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 1.019 ms | 0 - 68 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandLandmarkDetector | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNNDLC | 0.995 ms | 1 - 59 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandLandmarkDetector | float | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 2.307 ms | 0 - 31 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandLandmarkDetector | float | SA8295P ADP | Qualcomm® SA8295P | QNNDLC | 2.25 ms | 0 - 26 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandLandmarkDetector | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 1.03 ms | 0 - 69 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandLandmarkDetector | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNNDLC | 0.996 ms | 0 - 59 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandLandmarkDetector | float | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 1.911 ms | 0 - 27 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandLandmarkDetector | float | SA8775P ADP | Qualcomm® SA8775P | QNNDLC | 1.873 ms | 1 - 23 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandLandmarkDetector | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 0.751 ms | 0 - 42 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandLandmarkDetector | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNNDLC | 0.748 ms | 1 - 36 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandLandmarkDetector | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 0.961 ms | 0 - 40 MB | NPU | MediaPipe-Hand-Detection.onnx.zip | | HandLandmarkDetector | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | TFLITE | 0.586 ms | 0 - 34 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandLandmarkDetector | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNNDLC | 0.592 ms | 1 - 25 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandLandmarkDetector | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | ONNX | 0.794 ms | 0 - 26 MB | NPU | MediaPipe-Hand-Detection.onnx.zip | | HandLandmarkDetector | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | TFLITE | 0.502 ms | 0 - 27 MB | NPU | MediaPipe-Hand-Detection.tflite | | HandLandmarkDetector | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | QNNDLC | 0.511 ms | 1 - 26 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandLandmarkDetector | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | ONNX | 0.741 ms | 1 - 32 MB | NPU | MediaPipe-Hand-Detection.onnx.zip | | HandLandmarkDetector | float | Snapdragon X Elite CRD | Snapdragon® X Elite | QNNDLC | 1.272 ms | 46 - 46 MB | NPU | MediaPipe-Hand-Detection.dlc | | HandLandmarkDetector | float | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 1.36 ms | 6 - 6 MB | NPU | MediaPipe-Hand-Detection.onnx.zip | Configure Qualcomm® AI Hub to run this model on a cloud-hosted device Sign-in to Qualcomm® AI Hub with your Qualcomm® ID. Once signed in navigate to `Account -> Settings -> API Token`. With this API token, you can configure your client to run models on the cloud hosted devices. The package contains a simple end-to-end demo that downloads pre-trained weights and runs this model on a sample input. The above demo runs a reference implementation of pre-processing, model inference, and post processing. NOTE: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above). In addition to the demo, you can also run the model on a cloud-hosted Qualcomm® device. This script does the following: Performance check on-device on a cloud-hosted device Downloads compiled assets that can be deployed on-device for Android. Accuracy check between PyTorch and on-device outputs. This export script leverages Qualcomm® AI Hub to optimize, validate, and deploy this model on-device. Lets go through each step below in detail: To compile a PyTorch model for on-device deployment, we first trace the model in memory using the `jit.trace` and then call the `submitcompilejob` API. Step 2: Performance profiling on cloud-hosted device After compiling models from step 1. Models can be profiled model on-device using the `targetmodel`. Note that this scripts runs the model on a device automatically provisioned in the cloud. Once the job is submitted, you can navigate to a provided job URL to view a variety of on-device performance metrics. To verify the accuracy of the model on-device, you can run on-device inference on sample input data on the same cloud hosted device. With the output of the model, you can compute like PSNR, relative errors or spot check the output with expected output. Note: This on-device profiling and inference requires access to Qualcomm® AI Hub. Sign up for access. The models can be deployed using multiple runtimes: - TensorFlow Lite (`.tflite` export): This tutorial provides a guide to deploy the .tflite model in an Android application. - QNN (`.so` export ): This sample app provides instructions on how to use the `.so` shared library in an Android application. View on Qualcomm® AI Hub Get more details on MediaPipe-Hand-Detection's performance across various devices here. Explore all available models on Qualcomm® AI Hub License The license for the original implementation of MediaPipe-Hand-Detection can be found here. The license for the compiled assets for on-device deployment can be found here References MediaPipe Hands: On-device Real-time Hand Tracking 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.

FastSam-S: Optimized for Mobile Deployment Generate high quality segmentation mask on device The Fast Segment Anything Model (FastSAM) is a novel, real-time CNN-based solution for the Segment Anything task. This task is designed to segment any object within an image based on various possible user interaction prompts. The model performs competitively despite significantly reduced computation, making it a practical choice for a variety of vision tasks. This model is an implementation of FastSam-S found here. This repository provides scripts to run FastSam-S on Qualcomm® devices. More details on model performance across various devices, can be found here. - Model Type: Modelusecase.semanticsegmentation - Model Stats: - Model checkpoint: fastsam-s.pt - Inference latency: RealTime - Input resolution: 640x640 - Number of parameters: 11.8M - Model size (float): 45.1 MB | Model | Precision | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit | Target Model |---|---|---|---|---|---|---|---|---| | FastSam-S | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 38.513 ms | 4 - 89 MB | NPU | FastSam-S.tflite | | FastSam-S | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNNDLC | 37.74 ms | 5 - 121 MB | NPU | FastSam-S.dlc | | FastSam-S | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 12.826 ms | 4 - 65 MB | NPU | FastSam-S.tflite | | FastSam-S | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNNDLC | 15.997 ms | 4 - 46 MB | NPU | FastSam-S.dlc | | FastSam-S | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 7.194 ms | 0 - 59 MB | NPU | FastSam-S.tflite | | FastSam-S | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNNDLC | 6.827 ms | 5 - 46 MB | NPU | FastSam-S.dlc | | FastSam-S | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | ONNX | 8.33 ms | 0 - 88 MB | NPU | FastSam-S.onnx.zip | | FastSam-S | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 11.07 ms | 4 - 88 MB | NPU | FastSam-S.tflite | | FastSam-S | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNNDLC | 10.755 ms | 1 - 122 MB | NPU | FastSam-S.dlc | | FastSam-S | float | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 38.513 ms | 4 - 89 MB | NPU | FastSam-S.tflite | | FastSam-S | float | SA7255P ADP | Qualcomm® SA7255P | QNNDLC | 37.74 ms | 5 - 121 MB | NPU | FastSam-S.dlc | | FastSam-S | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 7.178 ms | 4 - 86 MB | NPU | FastSam-S.tflite | | FastSam-S | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNNDLC | 6.817 ms | 5 - 65 MB | NPU | FastSam-S.dlc | | FastSam-S | float | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 14.242 ms | 4 - 53 MB | NPU | FastSam-S.tflite | | FastSam-S | float | SA8295P ADP | Qualcomm® SA8295P | QNNDLC | 12.917 ms | 0 - 40 MB | NPU | FastSam-S.dlc | | FastSam-S | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 7.166 ms | 4 - 86 MB | NPU | FastSam-S.tflite | | FastSam-S | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNNDLC | 6.805 ms | 5 - 64 MB | NPU | FastSam-S.dlc | | FastSam-S | float | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 11.07 ms | 4 - 88 MB | NPU | FastSam-S.tflite | | FastSam-S | float | SA8775P ADP | Qualcomm® SA8775P | QNNDLC | 10.755 ms | 1 - 122 MB | NPU | FastSam-S.dlc | | FastSam-S | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 5.422 ms | 169 - 315 MB | NPU | FastSam-S.tflite | | FastSam-S | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNNDLC | 5.174 ms | 5 - 249 MB | NPU | FastSam-S.dlc | | FastSam-S | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 6.03 ms | 16 - 107 MB | NPU | FastSam-S.onnx.zip | | FastSam-S | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | TFLITE | 4.061 ms | 0 - 97 MB | NPU | FastSam-S.tflite | | FastSam-S | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNNDLC | 3.91 ms | 5 - 98 MB | NPU | FastSam-S.dlc | | FastSam-S | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | ONNX | 4.806 ms | 0 - 86 MB | NPU | FastSam-S.onnx.zip | | FastSam-S | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | TFLITE | 3.085 ms | 0 - 91 MB | NPU | FastSam-S.tflite | | FastSam-S | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | QNNDLC | 2.975 ms | 4 - 90 MB | NPU | FastSam-S.dlc | | FastSam-S | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | ONNX | 3.554 ms | 4 - 72 MB | NPU | FastSam-S.onnx.zip | | FastSam-S | float | Snapdragon X Elite CRD | Snapdragon® X Elite | QNNDLC | 7.426 ms | 114 - 114 MB | NPU | FastSam-S.dlc | | FastSam-S | float | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 8.558 ms | 20 - 20 MB | NPU | FastSam-S.onnx.zip | Install the package via pip: bash qai-hub configure --apitoken APITOKEN bash python -m qaihubmodels.models.fastsams.demo %run -m qaihubmodels.models.fastsams.demo bash python -m qaihubmodels.models.fastsams.export python import torch import qaihub as hub from qaihubmodels.models.fastsams import Model Trace model inputshape = torchmodel.getinputspec() sampleinputs = torchmodel.sampleinputs() ptmodel = torch.jit.trace(torchmodel, [torch.tensor(data[0]) for , data in sampleinputs.items()]) Compile model on a specific device compilejob = hub.submitcompilejob( model=ptmodel, device=device, inputspecs=torchmodel.getinputspec(), ) Get target model to run on-device targetmodel = compilejob.gettargetmodel() python profilejob = hub.submitprofilejob( model=targetmodel, device=device, ) python inputdata = torchmodel.sampleinputs() inferencejob = hub.submitinferencejob( model=targetmodel, device=device, inputs=inputdata, ) ondeviceoutput = inferencejob.downloadoutputdata() bash python -m qaihubmodels.models.fastsams.demo --eval-mode on-device %run -m qaihubmodels.models.fastsams.demo -- --eval-mode on-device ``` The models can be deployed using multiple runtimes: - TensorFlow Lite (`.tflite` export): This tutorial provides a guide to deploy the .tflite model in an Android application. - QNN (`.so` export ): This sample app provides instructions on how to use the `.so` shared library in an Android application. View on Qualcomm® AI Hub Get more details on FastSam-S's performance across various devices here. Explore all available models on Qualcomm® AI Hub License The license for the original implementation of FastSam-S can be found here. The license for the compiled assets for on-device deployment can be found here References Fast Segment Anything 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.

YOLOv8-Segmentation: Optimized for Mobile Deployment Real-time object segmentation optimized for mobile and edge by Ultralytics Ultralytics YOLOv8 is a machine learning model that predicts bounding boxes, segmentation masks and classes of objects in an image. This model is an implementation of YOLOv8-Segmentation found here. This repository provides scripts to run YOLOv8-Segmentation on Qualcomm® devices. More details on model performance across various devices, can be found here. WARNING: The model assets are not readily available for download due to licensing restrictions. - Model Type: Modelusecase.semanticsegmentation - Model Stats: - Model checkpoint: YOLOv8N-Seg - Input resolution: 640x640 - Number of output classes: 80 - Number of parameters: 3.43M - Model size (float): 13.2 MB - Model size (w8a16): 3.91 MB | Model | Precision | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit | Target Model |---|---|---|---|---|---|---|---|---| | YOLOv8-Segmentation | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 16.67 ms | 4 - 85 MB | NPU | -- | | YOLOv8-Segmentation | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNNDLC | 16.271 ms | 0 - 127 MB | NPU | -- | | YOLOv8-Segmentation | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 8.219 ms | 4 - 50 MB | NPU | -- | | YOLOv8-Segmentation | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNNDLC | 12.934 ms | 5 - 42 MB | NPU | -- | | YOLOv8-Segmentation | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 4.175 ms | 0 - 80 MB | NPU | -- | | YOLOv8-Segmentation | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNNDLC | 3.986 ms | 0 - 65 MB | NPU | -- | | YOLOv8-Segmentation | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | ONNX | 5.857 ms | 10 - 93 MB | NPU | -- | | YOLOv8-Segmentation | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 5.997 ms | 4 - 85 MB | NPU | -- | | YOLOv8-Segmentation | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNNDLC | 23.517 ms | 1 - 93 MB | NPU | -- | | YOLOv8-Segmentation | float | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 16.67 ms | 4 - 85 MB | NPU | -- | | YOLOv8-Segmentation | float | SA7255P ADP | Qualcomm® SA7255P | QNNDLC | 16.271 ms | 0 - 127 MB | NPU | -- | | YOLOv8-Segmentation | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 4.187 ms | 0 - 71 MB | NPU | -- | | YOLOv8-Segmentation | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNNDLC | 3.988 ms | 4 - 90 MB | NPU | -- | | YOLOv8-Segmentation | float | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 9.428 ms | 4 - 41 MB | NPU | -- | | YOLOv8-Segmentation | float | SA8295P ADP | Qualcomm® SA8295P | QNNDLC | 8.354 ms | 4 - 38 MB | NPU | -- | | YOLOv8-Segmentation | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 4.173 ms | 0 - 73 MB | NPU | -- | | YOLOv8-Segmentation | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNNDLC | 3.981 ms | 2 - 86 MB | NPU | -- | | YOLOv8-Segmentation | float | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 5.997 ms | 4 - 85 MB | NPU | -- | | YOLOv8-Segmentation | float | SA8775P ADP | Qualcomm® SA8775P | QNNDLC | 23.517 ms | 1 - 93 MB | NPU | -- | | YOLOv8-Segmentation | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 3.109 ms | 0 - 155 MB | NPU | -- | | YOLOv8-Segmentation | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNNDLC | 2.978 ms | 5 - 254 MB | NPU | -- | | YOLOv8-Segmentation | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 3.989 ms | 0 - 109 MB | NPU | -- | | YOLOv8-Segmentation | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | TFLITE | 2.46 ms | 0 - 89 MB | NPU | -- | | YOLOv8-Segmentation | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNNDLC | 2.27 ms | 5 - 108 MB | NPU | -- | | YOLOv8-Segmentation | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | ONNX | 3.33 ms | 1 - 89 MB | NPU | -- | | YOLOv8-Segmentation | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | TFLITE | 1.948 ms | 0 - 86 MB | NPU | -- | | YOLOv8-Segmentation | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | QNNDLC | 1.807 ms | 0 - 121 MB | NPU | -- | | YOLOv8-Segmentation | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | ONNX | 2.799 ms | 3 - 76 MB | NPU | -- | | YOLOv8-Segmentation | float | Snapdragon X Elite CRD | Snapdragon® X Elite | QNNDLC | 4.319 ms | 129 - 129 MB | NPU | -- | | YOLOv8-Segmentation | float | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 6.064 ms | 17 - 17 MB | NPU | -- | | YOLOv8-Segmentation | w8a16 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNNDLC | 7.321 ms | 2 - 33 MB | NPU | -- | | YOLOv8-Segmentation | w8a16 | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNNDLC | 4.72 ms | 2 - 47 MB | NPU | -- | | YOLOv8-Segmentation | w8a16 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNNDLC | 3.704 ms | 2 - 13 MB | NPU | -- | | YOLOv8-Segmentation | w8a16 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNNDLC | 4.319 ms | 1 - 33 MB | NPU | -- | | YOLOv8-Segmentation | w8a16 | SA7255P ADP | Qualcomm® SA7255P | QNNDLC | 7.321 ms | 2 - 33 MB | NPU | -- | | YOLOv8-Segmentation | w8a16 | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNNDLC | 3.704 ms | 4 - 13 MB | NPU | -- | | YOLOv8-Segmentation | w8a16 | SA8295P ADP | Qualcomm® SA8295P | QNNDLC | 5.032 ms | 1 - 39 MB | NPU | -- | | YOLOv8-Segmentation | w8a16 | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNNDLC | 3.688 ms | 2 - 12 MB | NPU | -- | | YOLOv8-Segmentation | w8a16 | SA8775P ADP | Qualcomm® SA8775P | QNNDLC | 4.319 ms | 1 - 33 MB | NPU | -- | | YOLOv8-Segmentation | w8a16 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNNDLC | 2.459 ms | 2 - 42 MB | NPU | -- | | YOLOv8-Segmentation | w8a16 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNNDLC | 1.671 ms | 2 - 41 MB | NPU | -- | | YOLOv8-Segmentation | w8a16 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | QNNDLC | 1.383 ms | 2 - 42 MB | NPU | -- | | YOLOv8-Segmentation | w8a16 | Snapdragon X Elite CRD | Snapdragon® X Elite | QNNDLC | 4.159 ms | 5 - 5 MB | NPU | -- | Install the package via pip: bash qai-hub configure --apitoken APITOKEN bash python -m qaihubmodels.models.yolov8seg.demo %run -m qaihubmodels.models.yolov8seg.demo bash python -m qaihubmodels.models.yolov8seg.export python import torch import qaihub as hub from qaihubmodels.models.yolov8seg import Model Trace model inputshape = torchmodel.getinputspec() sampleinputs = torchmodel.sampleinputs() ptmodel = torch.jit.trace(torchmodel, [torch.tensor(data[0]) for , data in sampleinputs.items()]) Compile model on a specific device compilejob = hub.submitcompilejob( model=ptmodel, device=device, inputspecs=torchmodel.getinputspec(), ) Get target model to run on-device targetmodel = compilejob.gettargetmodel() python profilejob = hub.submitprofilejob( model=targetmodel, device=device, ) python inputdata = torchmodel.sampleinputs() inferencejob = hub.submitinferencejob( model=targetmodel, device=device, inputs=inputdata, ) ondeviceoutput = inferencejob.downloadoutputdata() bash python -m qaihubmodels.models.yolov8seg.demo --eval-mode on-device %run -m qaihubmodels.models.yolov8seg.demo -- --eval-mode on-device ``` The models can be deployed using multiple runtimes: - TensorFlow Lite (`.tflite` export): This tutorial provides a guide to deploy the .tflite model in an Android application. - QNN (`.so` export ): This sample app provides instructions on how to use the `.so` shared library in an Android application. View on Qualcomm® AI Hub Get more details on YOLOv8-Segmentation's performance across various devices here. Explore all available models on Qualcomm® AI Hub License The license for the original implementation of YOLOv8-Segmentation can be found here. The license for the compiled assets for on-device deployment can be found here References Ultralytics YOLOv8 Docs: Instance 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.