SigLIP2: Optimized for Qualcomm Devices
SigLIP2 (Sigmoid Loss for Language-Image Pre-training 2) is a vision-language model from Google that computes cosine-similarity scores between images and text prompts. It can be used for zero-shot image classification, image search, and content moderation without any task-specific fine-tuning.
This is based on the implementation of SigLIP2 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 2.45, ONNX Runtime 1.27.1 | Download |
| ONNX | w8a16 | Universal | QAIRT 2.45, ONNX Runtime 1.27.1 | Download |
| QNN_DLC | float | Universal | QAIRT 2.45 | Download |
| QNN_DLC | w8a16 | Universal | QAIRT 2.45 | Download |
| TFLITE | float | Universal | QAIRT 2.45 | Download |
For more device-specific assets and performance metrics, visit SigLIP2 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 SigLIP2 on GitHub for usage instructions.
Model Details
Model Type: Model_use_case.image_classification
Model Stats:
- Image input resolution: 224x224
- Model checkpoint: google/siglip2-base-patch16-224
- Model size (image_encoder) (float): 352 MB
- Model size (image_encoder) (w8a16): 92.8 MB
- Model size (text_encoder) (float): 1.05 GB
- Model size (text_encoder) (w8a16): 461 MB
- Text sequence length: 64
Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit |
|---|---|---|---|---|---|---|
| image_encoder | ONNX | float | Snapdragon® X2 Elite | 4.738 ms | 2 - 2 MB | NPU |
| image_encoder | ONNX | float | Snapdragon® X Elite | 8.828 ms | 182 - 182 MB | NPU |
| image_encoder | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 5.893 ms | 1 - 311 MB | NPU |
| image_encoder | ONNX | float | Snapdragon® 8 Gen 1 Mobile | 13.248 ms | 1 - 294 MB | NPU |
| image_encoder | ONNX | float | Qualcomm® Dragonwing™ IQ-8275 | 11.923 ms | 0 - 4 MB | NPU |
| image_encoder | ONNX | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 8.445 ms | 0 - 613 MB | NPU |
| image_encoder | ONNX | float | Qualcomm® QCS8450 | 13.248 ms | 1 - 294 MB | NPU |
| image_encoder | ONNX | float | Qualcomm® Dragonwing™ IQ-9075 | 11.472 ms | 0 - 4 MB | NPU |
| image_encoder | ONNX | float | Qualcomm® Dragonwing™ IQ-X7181 | 8.828 ms | 182 - 182 MB | NPU |
| image_encoder | ONNX | float | Qualcomm® Dragonwing™ Q-8750 | 4.241 ms | 0 - 176 MB | NPU |
| image_encoder | ONNX | float | Snapdragon® 8 Elite Mobile | 4.241 ms | 0 - 176 MB | NPU |
| image_encoder | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 3.262 ms | 1 - 167 MB | NPU |
| image_encoder | ONNX | w8a16 | Snapdragon® X2 Elite | 3.05 ms | 1 - 1 MB | NPU |
| image_encoder | ONNX | w8a16 | Snapdragon® X Elite | 5.544 ms | 91 - 91 MB | NPU |
| image_encoder | ONNX | w8a16 | Snapdragon® 8 Gen 3 Mobile | 3.675 ms | 0 - 290 MB | NPU |
| image_encoder | ONNX | w8a16 | Snapdragon® 8 Gen 1 Mobile | 7.195 ms | 0 - 280 MB | NPU |
| image_encoder | ONNX | w8a16 | Qualcomm® Dragonwing™ QCS6490 | 23.34 ms | 2 - 5 MB | NPU |
| image_encoder | ONNX | w8a16 | Qualcomm® Dragonwing™ IQ-8275 | 5.066 ms | 0 - 4 MB | NPU |
| image_encoder | ONNX | w8a16 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 5.282 ms | 0 - 98 MB | NPU |
| image_encoder | ONNX | w8a16 | Qualcomm® QCS8450 | 7.195 ms | 0 - 280 MB | NPU |
| image_encoder | ONNX | w8a16 | Qualcomm® Dragonwing™ IQ-9075 | 5.258 ms | 0 - 3 MB | NPU |
| image_encoder | ONNX | w8a16 | Qualcomm® Dragonwing™ IQ-X7181 | 5.544 ms | 91 - 91 MB | NPU |
| image_encoder | ONNX | w8a16 | Qualcomm® Dragonwing™ Q-6690 | 53.198 ms | 0 - 395 MB | NPU |
| image_encoder | ONNX | w8a16 | Qualcomm® Dragonwing™ Q-7790 | 6.998 ms | 0 - 270 MB | NPU |
| image_encoder | ONNX | w8a16 | Qualcomm® Dragonwing™ Q-8750 | 2.837 ms | 0 - 236 MB | NPU |
| image_encoder | ONNX | w8a16 | Snapdragon® 8 Elite Mobile | 2.837 ms | 0 - 236 MB | NPU |
| image_encoder | ONNX | w8a16 | Snapdragon® 8 Elite Gen 5 Mobile | 2.083 ms | 0 - 238 MB | NPU |
| image_encoder | ONNX | w8a16 | Snapdragon® 7 Gen 4 Mobile | 6.998 ms | 0 - 270 MB | NPU |
| image_encoder | QNN_DLC | float | Snapdragon® X2 Elite | 3.989 ms | 1 - 1 MB | NPU |
| image_encoder | QNN_DLC | float | Snapdragon® X Elite | 9.205 ms | 1 - 1 MB | NPU |
| image_encoder | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 6.047 ms | 0 - 264 MB | NPU |
| image_encoder | QNN_DLC | float | Snapdragon® 8 Gen 1 Mobile | 13.512 ms | 0 - 243 MB | NPU |
| image_encoder | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-8275 | 12.117 ms | 1 - 4 MB | NPU |
| image_encoder | QNN_DLC | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 8.574 ms | 1 - 3 MB | NPU |
| image_encoder | QNN_DLC | float | Qualcomm® SA8775P | 11.417 ms | 1 - 175 MB | NPU |
| image_encoder | QNN_DLC | float | Qualcomm® SA8650P | 11.417 ms | 1 - 175 MB | NPU |
| image_encoder | QNN_DLC | float | Qualcomm® SA8255P | 11.417 ms | 1 - 175 MB | NPU |
| image_encoder | QNN_DLC | float | Qualcomm® QCS8450 | 13.512 ms | 0 - 243 MB | NPU |
| image_encoder | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-9075 | 11.475 ms | 1 - 3 MB | NPU |
| image_encoder | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-X7181 | 9.205 ms | 1 - 1 MB | NPU |
| image_encoder | QNN_DLC | float | Qualcomm® Dragonwing™ Q-8750 | 4.242 ms | 0 - 176 MB | NPU |
| image_encoder | QNN_DLC | float | Qualcomm® SA7255P | 36.986 ms | 1 - 174 MB | NPU |
| image_encoder | QNN_DLC | float | Qualcomm® SA8295P | 14.146 ms | 1 - 150 MB | NPU |
| image_encoder | QNN_DLC | float | Snapdragon® 8 Elite Mobile | 4.242 ms | 0 - 176 MB | NPU |
| image_encoder | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 3.396 ms | 1 - 164 MB | NPU |
| image_encoder | QNN_DLC | w8a16 | Snapdragon® X2 Elite | 4.005 ms | 0 - 0 MB | NPU |
| image_encoder | QNN_DLC | w8a16 | Snapdragon® X Elite | 9.137 ms | 0 - 0 MB | NPU |
| image_encoder | QNN_DLC | w8a16 | Snapdragon® 8 Gen 3 Mobile | 5.545 ms | 0 - 311 MB | NPU |
| image_encoder | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ IQ-8275 | 7.5 ms | 0 - 3 MB | NPU |
| image_encoder | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 8.443 ms | 0 - 282 MB | NPU |
| image_encoder | QNN_DLC | w8a16 | Qualcomm® SA8775P | 8.664 ms | 0 - 255 MB | NPU |
| image_encoder | QNN_DLC | w8a16 | Qualcomm® SA8650P | 8.664 ms | 0 - 255 MB | NPU |
| image_encoder | QNN_DLC | w8a16 | Qualcomm® SA8255P | 8.664 ms | 0 - 255 MB | NPU |
| image_encoder | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ IQ-9075 | 8.445 ms | 2 - 4 MB | NPU |
| image_encoder | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ IQ-X7181 | 9.137 ms | 0 - 0 MB | NPU |
| image_encoder | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ Q-6690 | 92.12 ms | 0 - 419 MB | NPU |
| image_encoder | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ Q-7790 | 13.929 ms | 0 - 402 MB | NPU |
| image_encoder | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ Q-8750 | 4.478 ms | 0 - 254 MB | NPU |
| image_encoder | QNN_DLC | w8a16 | Qualcomm® SA7255P | 19.111 ms | 0 - 253 MB | NPU |
| image_encoder | QNN_DLC | w8a16 | Snapdragon® 8 Elite Mobile | 4.478 ms | 0 - 254 MB | NPU |
| image_encoder | QNN_DLC | w8a16 | Snapdragon® 8 Elite Gen 5 Mobile | 3.28 ms | 0 - 257 MB | NPU |
| image_encoder | QNN_DLC | w8a16 | Snapdragon® 7 Gen 4 Mobile | 13.929 ms | 0 - 402 MB | NPU |
| image_encoder | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 5.92 ms | 0 - 263 MB | NPU |
| image_encoder | TFLITE | float | Snapdragon® 8 Gen 1 Mobile | 13.235 ms | 0 - 241 MB | NPU |
| image_encoder | TFLITE | float | Qualcomm® Dragonwing™ IQ-8275 | 11.996 ms | 0 - 184 MB | NPU |
| image_encoder | TFLITE | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 8.398 ms | 0 - 2 MB | NPU |
| image_encoder | TFLITE | float | Qualcomm® SA8775P | 11.269 ms | 0 - 176 MB | NPU |
| image_encoder | TFLITE | float | Qualcomm® SA8650P | 11.269 ms | 0 - 176 MB | NPU |
| image_encoder | TFLITE | float | Qualcomm® SA8255P | 11.269 ms | 0 - 176 MB | NPU |
| image_encoder | TFLITE | float | Qualcomm® QCS8450 | 13.235 ms | 0 - 241 MB | NPU |
| image_encoder | TFLITE | float | Qualcomm® Dragonwing™ IQ-9075 | 11.364 ms | 0 - 184 MB | NPU |
| image_encoder | TFLITE | float | Qualcomm® Dragonwing™ Q-8750 | 4.16 ms | 0 - 180 MB | NPU |
| image_encoder | TFLITE | float | Qualcomm® SA7255P | 36.747 ms | 0 - 175 MB | NPU |
| image_encoder | TFLITE | float | Qualcomm® SA8295P | 13.923 ms | 0 - 161 MB | NPU |
| image_encoder | TFLITE | float | Snapdragon® 8 Elite Mobile | 4.16 ms | 0 - 180 MB | NPU |
| image_encoder | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 3.297 ms | 0 - 170 MB | NPU |
| text_encoder | ONNX | float | Snapdragon® X2 Elite | 1.39 ms | 1 - 1 MB | NPU |
| text_encoder | ONNX | float | Snapdragon® X Elite | 2.9 ms | 542 - 542 MB | NPU |
| text_encoder | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 2.752 ms | 0 - 251 MB | NPU |
| text_encoder | ONNX | float | Snapdragon® 8 Gen 1 Mobile | 4.465 ms | 0 - 244 MB | NPU |
| text_encoder | ONNX | float | Qualcomm® Dragonwing™ IQ-8275 | 3.753 ms | 0 - 4 MB | NPU |
| text_encoder | ONNX | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 3.513 ms | 0 - 643 MB | NPU |
| text_encoder | ONNX | float | Qualcomm® QCS8450 | 4.465 ms | 0 - 244 MB | NPU |
| text_encoder | ONNX | float | Qualcomm® Dragonwing™ IQ-9075 | 3.519 ms | 0 - 3 MB | NPU |
| text_encoder | ONNX | float | Qualcomm® Dragonwing™ IQ-X7181 | 2.9 ms | 542 - 542 MB | NPU |
| text_encoder | ONNX | float | Qualcomm® Dragonwing™ Q-8750 | 2.409 ms | 0 - 138 MB | NPU |
| text_encoder | ONNX | float | Snapdragon® 8 Elite Mobile | 2.409 ms | 0 - 138 MB | NPU |
| text_encoder | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 2.405 ms | 0 - 140 MB | NPU |
| text_encoder | ONNX | w8a16 | Snapdragon® X2 Elite | 0.834 ms | 1 - 1 MB | NPU |
| text_encoder | ONNX | w8a16 | Snapdragon® X Elite | 1.722 ms | 460 - 460 MB | NPU |
| text_encoder | ONNX | w8a16 | Snapdragon® 8 Gen 3 Mobile | 1.453 ms | 0 - 272 MB | NPU |
| text_encoder | ONNX | w8a16 | Snapdragon® 8 Gen 1 Mobile | 2.832 ms | 0 - 276 MB | NPU |
| text_encoder | ONNX | w8a16 | Qualcomm® Dragonwing™ QCS6490 | 8.105 ms | 0 - 3 MB | NPU |
| text_encoder | ONNX | w8a16 | Qualcomm® Dragonwing™ IQ-8275 | 2.051 ms | 0 - 3 MB | NPU |
| text_encoder | ONNX | w8a16 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 1.866 ms | 0 - 514 MB | NPU |
| text_encoder | ONNX | w8a16 | Qualcomm® QCS8450 | 2.832 ms | 0 - 276 MB | NPU |
| text_encoder | ONNX | w8a16 | Qualcomm® Dragonwing™ IQ-9075 | 2.189 ms | 0 - 3 MB | NPU |
| text_encoder | ONNX | w8a16 | Qualcomm® Dragonwing™ IQ-X7181 | 1.722 ms | 460 - 460 MB | NPU |
| text_encoder | ONNX | w8a16 | Qualcomm® Dragonwing™ Q-6690 | 13.234 ms | 0 - 258 MB | NPU |
| text_encoder | ONNX | w8a16 | Qualcomm® Dragonwing™ Q-7790 | 2.847 ms | 0 - 239 MB | NPU |
| text_encoder | ONNX | w8a16 | Qualcomm® Dragonwing™ Q-8750 | 1.268 ms | 0 - 231 MB | NPU |
| text_encoder | ONNX | w8a16 | Snapdragon® 8 Elite Mobile | 1.268 ms | 0 - 231 MB | NPU |
| text_encoder | ONNX | w8a16 | Snapdragon® 8 Elite Gen 5 Mobile | 1.245 ms | 0 - 234 MB | NPU |
| text_encoder | ONNX | w8a16 | Snapdragon® 7 Gen 4 Mobile | 2.847 ms | 0 - 239 MB | NPU |
| text_encoder | QNN_DLC | float | Snapdragon® X2 Elite | 1.757 ms | 0 - 0 MB | NPU |
| text_encoder | QNN_DLC | float | Snapdragon® X Elite | 3.062 ms | 0 - 0 MB | NPU |
| text_encoder | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 2.686 ms | 0 - 249 MB | NPU |
| text_encoder | QNN_DLC | float | Snapdragon® 8 Gen 1 Mobile | 4.467 ms | 0 - 237 MB | NPU |
| text_encoder | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-8275 | 3.743 ms | 0 - 2 MB | NPU |
| text_encoder | QNN_DLC | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 3.499 ms | 0 - 2 MB | NPU |
| text_encoder | QNN_DLC | float | Qualcomm® SA8775P | 3.655 ms | 0 - 133 MB | NPU |
| text_encoder | QNN_DLC | float | Qualcomm® SA8650P | 3.655 ms | 0 - 133 MB | NPU |
| text_encoder | QNN_DLC | float | Qualcomm® SA8255P | 3.655 ms | 0 - 133 MB | NPU |
| text_encoder | QNN_DLC | float | Qualcomm® QCS8450 | 4.467 ms | 0 - 237 MB | NPU |
| text_encoder | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-9075 | 3.505 ms | 2 - 4 MB | NPU |
| text_encoder | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-X7181 | 3.062 ms | 0 - 0 MB | NPU |
| text_encoder | QNN_DLC | float | Qualcomm® Dragonwing™ Q-8750 | 2.401 ms | 0 - 134 MB | NPU |
| text_encoder | QNN_DLC | float | Qualcomm® SA7255P | 10.029 ms | 0 - 134 MB | NPU |
| text_encoder | QNN_DLC | float | Qualcomm® SA8295P | 4.649 ms | 0 - 123 MB | NPU |
| text_encoder | QNN_DLC | float | Snapdragon® 8 Elite Mobile | 2.401 ms | 0 - 134 MB | NPU |
| text_encoder | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 2.388 ms | 0 - 137 MB | NPU |
| text_encoder | QNN_DLC | w8a16 | Snapdragon® X2 Elite | 1.052 ms | 1 - 1 MB | NPU |
| text_encoder | QNN_DLC | w8a16 | Snapdragon® X Elite | 1.945 ms | 0 - 0 MB | NPU |
| text_encoder | QNN_DLC | w8a16 | Snapdragon® 8 Gen 3 Mobile | 1.456 ms | 0 - 261 MB | NPU |
| text_encoder | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ IQ-8275 | 2.067 ms | 0 - 2 MB | NPU |
| text_encoder | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 1.909 ms | 0 - 955 MB | NPU |
| text_encoder | QNN_DLC | w8a16 | Qualcomm® SA8775P | 2.216 ms | 0 - 226 MB | NPU |
| text_encoder | QNN_DLC | w8a16 | Qualcomm® SA8650P | 2.216 ms | 0 - 226 MB | NPU |
| text_encoder | QNN_DLC | w8a16 | Qualcomm® SA8255P | 2.216 ms | 0 - 226 MB | NPU |
| text_encoder | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ IQ-9075 | 2.078 ms | 0 - 2 MB | NPU |
| text_encoder | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ IQ-X7181 | 1.945 ms | 0 - 0 MB | NPU |
| text_encoder | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ Q-6690 | 13.635 ms | 0 - 261 MB | NPU |
| text_encoder | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ Q-7790 | 2.825 ms | 0 - 237 MB | NPU |
| text_encoder | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ Q-8750 | 1.281 ms | 0 - 227 MB | NPU |
| text_encoder | QNN_DLC | w8a16 | Qualcomm® SA7255P | 4.02 ms | 0 - 226 MB | NPU |
| text_encoder | QNN_DLC | w8a16 | Snapdragon® 8 Elite Mobile | 1.281 ms | 0 - 227 MB | NPU |
| text_encoder | QNN_DLC | w8a16 | Snapdragon® 8 Elite Gen 5 Mobile | 1.254 ms | 0 - 229 MB | NPU |
| text_encoder | QNN_DLC | w8a16 | Snapdragon® 7 Gen 4 Mobile | 2.825 ms | 0 - 237 MB | NPU |
| text_encoder | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 2.737 ms | 0 - 253 MB | NPU |
| text_encoder | TFLITE | float | Snapdragon® 8 Gen 1 Mobile | 4.461 ms | 0 - 244 MB | NPU |
| text_encoder | TFLITE | float | Qualcomm® Dragonwing™ IQ-8275 | 3.779 ms | 0 - 543 MB | NPU |
| text_encoder | TFLITE | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 3.502 ms | 0 - 3 MB | NPU |
| text_encoder | TFLITE | float | Qualcomm® SA8775P | 3.699 ms | 0 - 142 MB | NPU |
| text_encoder | TFLITE | float | Qualcomm® SA8650P | 3.699 ms | 0 - 142 MB | NPU |
| text_encoder | TFLITE | float | Qualcomm® SA8255P | 3.699 ms | 0 - 142 MB | NPU |
| text_encoder | TFLITE | float | Qualcomm® QCS8450 | 4.461 ms | 0 - 244 MB | NPU |
| text_encoder | TFLITE | float | Qualcomm® Dragonwing™ IQ-9075 | 3.526 ms | 0 - 543 MB | NPU |
| text_encoder | TFLITE | float | Qualcomm® Dragonwing™ Q-8750 | 2.412 ms | 0 - 146 MB | NPU |
| text_encoder | TFLITE | float | Qualcomm® SA7255P | 10.064 ms | 0 - 143 MB | NPU |
| text_encoder | TFLITE | float | Qualcomm® SA8295P | 4.671 ms | 0 - 129 MB | NPU |
| text_encoder | TFLITE | float | Snapdragon® 8 Elite Mobile | 2.412 ms | 0 - 146 MB | NPU |
| text_encoder | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 2.403 ms | 0 - 146 MB | NPU |
License
- The license for the original implementation of SigLIP2 can be found here.
References
- SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features
- 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.
