Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data
Paper
•
2107.10833
•
Published
•
1
Real-ESRGAN-x4plus: Optimized for Qualcomm Devices
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 is based on the implementation of Real-ESRGAN-x4plus found here. This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the Qualcomm® AI Hub Models library to export with custom configurations. More details on model performance across various devices, can be found here.
Qualcomm AI Hub Models uses Qualcomm AI Hub Workbench to compile, profile, and evaluate this model. Sign up to run these models on a hosted Qualcomm® device.
Getting Started
There are two ways to deploy this model on your device:
Option 1: Download Pre-Exported Models
Below are pre-exported model assets ready for deployment.
| Runtime | Precision | Chipset | SDK Versions | Download |
|---|---|---|---|---|
| ONNX | float | Universal | QAIRT 2.37, ONNX Runtime 1.23.0 | Download |
| QNN_DLC | float | Universal | QAIRT 2.42 | Download |
| QNN_DLC | w8a8 | Universal | QAIRT 2.42 | Download |
| TFLITE | float | Universal | QAIRT 2.42, TFLite 2.17.0 | Download |
| TFLITE | w8a8 | Universal | QAIRT 2.42, TFLite 2.17.0 | Download |
For more device-specific assets and performance metrics, visit Real-ESRGAN-x4plus 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 Real-ESRGAN-x4plus on GitHub for usage instructions.
Model Details
Model Type: Model_use_case.super_resolution
Model Stats:
- Model checkpoint: RealESRGAN_x4plus
- Input resolution: 128x128
- Number of parameters: 16.7M
- Model size (float): 63.9 MB
- Model size (w8a8): 16.7 MB
Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit |
|---|---|---|---|---|---|---|
| Real-ESRGAN-x4plus | ONNX | float | Snapdragon® X Elite | 65.564 ms | 37 - 37 MB | NPU |
| Real-ESRGAN-x4plus | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 49.752 ms | 0 - 689 MB | NPU |
| Real-ESRGAN-x4plus | ONNX | float | Qualcomm® QCS8550 (Proxy) | 64.138 ms | 0 - 559 MB | NPU |
| Real-ESRGAN-x4plus | ONNX | float | Qualcomm® QCS9075 | 106.758 ms | 6 - 9 MB | NPU |
| Real-ESRGAN-x4plus | ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 38.817 ms | 7 - 258 MB | NPU |
| Real-ESRGAN-x4plus | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 27.445 ms | 6 - 266 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | float | Snapdragon® X Elite | 65.048 ms | 0 - 0 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 49.703 ms | 0 - 752 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | float | Qualcomm® QCS8275 (Proxy) | 452.116 ms | 0 - 348 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | float | Qualcomm® QCS8550 (Proxy) | 62.963 ms | 0 - 4 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | float | Qualcomm® SA8775P | 503.535 ms | 0 - 347 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | float | Qualcomm® QCS9075 | 108.847 ms | 0 - 5 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | float | Qualcomm® QCS8450 (Proxy) | 113.142 ms | 1 - 757 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | float | Qualcomm® SA7255P | 452.116 ms | 0 - 348 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | float | Qualcomm® SA8295P | 111.293 ms | 0 - 357 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | float | Snapdragon® 8 Elite For Galaxy Mobile | 37.922 ms | 0 - 332 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 25.742 ms | 0 - 331 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | w8a8 | Snapdragon® X Elite | 25.884 ms | 0 - 0 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | w8a8 | Snapdragon® 8 Gen 3 Mobile | 16.79 ms | 0 - 691 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | w8a8 | Qualcomm® QCS6490 | 109.873 ms | 2 - 4 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | w8a8 | Qualcomm® QCS8275 (Proxy) | 69.033 ms | 0 - 478 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | w8a8 | Qualcomm® QCS8550 (Proxy) | 24.619 ms | 0 - 3 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | w8a8 | Qualcomm® SA8775P | 22.985 ms | 0 - 477 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | w8a8 | Qualcomm® QCS9075 | 29.524 ms | 0 - 2 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | w8a8 | Qualcomm® QCM6690 | 493.304 ms | 0 - 465 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | w8a8 | Qualcomm® QCS8450 (Proxy) | 37.494 ms | 0 - 713 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | w8a8 | Qualcomm® SA7255P | 69.033 ms | 0 - 478 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | w8a8 | Qualcomm® SA8295P | 36.364 ms | 0 - 484 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 12.361 ms | 0 - 417 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | w8a8 | Snapdragon® 7 Gen 4 Mobile | 38.592 ms | 0 - 484 MB | NPU |
| Real-ESRGAN-x4plus | QNN_DLC | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 9.126 ms | 0 - 431 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 48.871 ms | 3 - 796 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | float | Qualcomm® QCS8275 (Proxy) | 452.332 ms | 3 - 395 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | float | Qualcomm® QCS8550 (Proxy) | 59.885 ms | 3 - 6 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | float | Qualcomm® SA8775P | 105.619 ms | 3 - 393 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | float | Qualcomm® QCS9075 | 109.28 ms | 1 - 45 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | float | Qualcomm® QCS8450 (Proxy) | 113.953 ms | 4 - 784 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | float | Qualcomm® SA7255P | 452.332 ms | 3 - 395 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | float | Qualcomm® SA8295P | 110.425 ms | 3 - 398 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | float | Snapdragon® 8 Elite For Galaxy Mobile | 38.056 ms | 3 - 373 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 28.131 ms | 3 - 377 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | w8a8 | Snapdragon® 8 Gen 3 Mobile | 13.828 ms | 1 - 674 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | w8a8 | Qualcomm® QCS6490 | 93.178 ms | 1 - 26 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | w8a8 | Qualcomm® QCS8275 (Proxy) | 64.674 ms | 1 - 450 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | w8a8 | Qualcomm® QCS8550 (Proxy) | 18.667 ms | 1 - 4 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | w8a8 | Qualcomm® SA8775P | 18.301 ms | 1 - 450 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | w8a8 | Qualcomm® QCS9075 | 25.716 ms | 1 - 27 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | w8a8 | Qualcomm® QCM6690 | 434.742 ms | 1 - 436 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | w8a8 | Qualcomm® QCS8450 (Proxy) | 37.707 ms | 1 - 684 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | w8a8 | Qualcomm® SA7255P | 64.674 ms | 1 - 450 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | w8a8 | Qualcomm® SA8295P | 33.678 ms | 1 - 450 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 11.599 ms | 1 - 479 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | w8a8 | Snapdragon® 7 Gen 4 Mobile | 31.163 ms | 1 - 442 MB | NPU |
| Real-ESRGAN-x4plus | TFLITE | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 7.474 ms | 1 - 508 MB | NPU |
License
- The license for the original implementation of Real-ESRGAN-x4plus 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.