Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data
Paper • 2107.10833 • Published • 1
Real-ESRGAN-General-x4v3: Optimized for Qualcomm Devices
Real-ESRGAN is a machine learning model that upscales an image with minimal loss in quality.
This is based on the implementation of Real-ESRGAN-General-x4v3 found here. This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the Qualcomm® AI Hub Models library to export with custom configurations. More details on model performance across various devices, can be found here.
Qualcomm AI Hub Models uses Qualcomm AI Hub Workbench to compile, profile, and evaluate this model. Sign up to run these models on a hosted Qualcomm® device.
Getting Started
There are two ways to deploy this model on your device:
Option 1: Download Pre-Exported Models
Below are pre-exported model assets ready for deployment.
| Runtime | Precision | Chipset | SDK Versions | Download |
|---|---|---|---|---|
| ONNX | float | Universal | QAIRT 2.42, ONNX Runtime 1.24.3 | Download |
| ONNX | w8a8 | Universal | QAIRT 2.42, ONNX Runtime 1.24.3 | Download |
| QNN_DLC | float | Universal | QAIRT 2.43 | Download |
| QNN_DLC | w8a8 | Universal | QAIRT 2.43 | Download |
| TFLITE | float | Universal | QAIRT 2.43, TFLite 2.19.1 | Download |
| TFLITE | w8a8 | Universal | QAIRT 2.43, TFLite 2.19.1 | Download |
For more device-specific assets and performance metrics, visit Real-ESRGAN-General-x4v3 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-General-x4v3 on GitHub for usage instructions.
Model Details
Model Type: Model_use_case.super_resolution
Model Stats:
- Model checkpoint: realesr-general-x4v3
- Input resolution: 128x128
- Number of parameters: 1.21M
- Model size (float): 4.65 MB
- Model size (w8a8): 1.25 MB
Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit |
|---|---|---|---|---|---|---|
| Real-ESRGAN-General-x4v3 | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 2.655 ms | 0 - 152 MB | NPU |
| Real-ESRGAN-General-x4v3 | ONNX | float | Snapdragon® X2 Elite | 3.443 ms | 4 - 4 MB | NPU |
| Real-ESRGAN-General-x4v3 | ONNX | float | Snapdragon® X Elite | 6.439 ms | 8 - 8 MB | NPU |
| Real-ESRGAN-General-x4v3 | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 4.555 ms | 6 - 193 MB | NPU |
| Real-ESRGAN-General-x4v3 | ONNX | float | Qualcomm® QCS8550 (Proxy) | 5.984 ms | 0 - 27 MB | NPU |
| Real-ESRGAN-General-x4v3 | ONNX | float | Qualcomm® QCS9075 | 9.783 ms | 6 - 9 MB | NPU |
| Real-ESRGAN-General-x4v3 | ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 3.564 ms | 0 - 151 MB | NPU |
| Real-ESRGAN-General-x4v3 | ONNX | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.693 ms | 0 - 158 MB | NPU |
| Real-ESRGAN-General-x4v3 | ONNX | w8a8 | Snapdragon® X2 Elite | 0.87 ms | 0 - 0 MB | NPU |
| Real-ESRGAN-General-x4v3 | ONNX | w8a8 | Snapdragon® X Elite | 1.927 ms | 2 - 2 MB | NPU |
| Real-ESRGAN-General-x4v3 | ONNX | w8a8 | Snapdragon® 8 Gen 3 Mobile | 1.226 ms | 0 - 169 MB | NPU |
| Real-ESRGAN-General-x4v3 | ONNX | w8a8 | Qualcomm® QCS6490 | 153.435 ms | 25 - 31 MB | CPU |
| Real-ESRGAN-General-x4v3 | ONNX | w8a8 | Qualcomm® QCS8550 (Proxy) | 1.728 ms | 0 - 4 MB | NPU |
| Real-ESRGAN-General-x4v3 | ONNX | w8a8 | Qualcomm® QCS9075 | 1.947 ms | 0 - 3 MB | NPU |
| Real-ESRGAN-General-x4v3 | ONNX | w8a8 | Qualcomm® QCM6690 | 146.197 ms | 26 - 34 MB | CPU |
| Real-ESRGAN-General-x4v3 | ONNX | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 0.966 ms | 0 - 39 MB | NPU |
| Real-ESRGAN-General-x4v3 | ONNX | w8a8 | Snapdragon® 7 Gen 4 Mobile | 116.958 ms | 26 - 35 MB | CPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 2.304 ms | 0 - 159 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | float | Snapdragon® X2 Elite | 3.536 ms | 0 - 0 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | float | Snapdragon® X Elite | 6.346 ms | 0 - 0 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 4.281 ms | 0 - 186 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | float | Qualcomm® QCS8275 (Proxy) | 33.193 ms | 0 - 151 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | float | Qualcomm® QCS8550 (Proxy) | 5.838 ms | 0 - 2 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | float | Qualcomm® SA8775P | 9.415 ms | 0 - 154 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | float | Qualcomm® QCS9075 | 9.63 ms | 0 - 5 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | float | Qualcomm® QCS8450 (Proxy) | 8.836 ms | 0 - 92 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | float | Qualcomm® SA7255P | 33.193 ms | 0 - 151 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | float | Qualcomm® SA8295P | 9.759 ms | 0 - 50 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | float | Snapdragon® 8 Elite For Galaxy Mobile | 3.215 ms | 0 - 154 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.664 ms | 0 - 150 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | w8a8 | Snapdragon® X2 Elite | 0.991 ms | 0 - 0 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | w8a8 | Snapdragon® X Elite | 1.954 ms | 0 - 0 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | w8a8 | Snapdragon® 8 Gen 3 Mobile | 1.226 ms | 0 - 163 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | w8a8 | Qualcomm® QCS6490 | 6.145 ms | 0 - 2 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | w8a8 | Qualcomm® QCS8275 (Proxy) | 5.66 ms | 0 - 147 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | w8a8 | Qualcomm® QCS8550 (Proxy) | 1.745 ms | 0 - 2 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | w8a8 | Qualcomm® SA8775P | 2.03 ms | 0 - 152 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | w8a8 | Qualcomm® QCS9075 | 1.982 ms | 2 - 4 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | w8a8 | Qualcomm® QCM6690 | 25.633 ms | 0 - 147 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | w8a8 | Qualcomm® QCS8450 (Proxy) | 2.768 ms | 0 - 66 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | w8a8 | Qualcomm® SA7255P | 5.66 ms | 0 - 147 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | w8a8 | Qualcomm® SA8295P | 3.284 ms | 0 - 44 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 0.907 ms | 0 - 35 MB | NPU |
| Real-ESRGAN-General-x4v3 | QNN_DLC | w8a8 | Snapdragon® 7 Gen 4 Mobile | 2.413 ms | 0 - 147 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 2.3 ms | 0 - 162 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 4.28 ms | 0 - 188 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | float | Qualcomm® QCS8275 (Proxy) | 33.23 ms | 3 - 157 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | float | Qualcomm® QCS8550 (Proxy) | 5.841 ms | 3 - 5 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | float | Qualcomm® SA8775P | 9.45 ms | 3 - 160 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | float | Qualcomm® QCS9075 | 9.595 ms | 3 - 11 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | float | Qualcomm® QCS8450 (Proxy) | 8.841 ms | 3 - 99 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | float | Qualcomm® SA7255P | 33.23 ms | 3 - 157 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | float | Qualcomm® SA8295P | 9.744 ms | 3 - 57 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | float | Snapdragon® 8 Elite For Galaxy Mobile | 3.224 ms | 0 - 157 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.667 ms | 0 - 153 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | w8a8 | Snapdragon® 8 Gen 3 Mobile | 1.231 ms | 0 - 164 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | w8a8 | Qualcomm® QCS6490 | 6.246 ms | 1 - 5 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | w8a8 | Qualcomm® QCS8275 (Proxy) | 5.709 ms | 1 - 149 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | w8a8 | Qualcomm® QCS8550 (Proxy) | 1.763 ms | 0 - 2 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | w8a8 | Qualcomm® SA8775P | 2.077 ms | 0 - 153 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | w8a8 | Qualcomm® QCS9075 | 1.995 ms | 0 - 4 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | w8a8 | Qualcomm® QCM6690 | 25.457 ms | 1 - 149 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | w8a8 | Qualcomm® QCS8450 (Proxy) | 3.003 ms | 0 - 66 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | w8a8 | Qualcomm® SA7255P | 5.709 ms | 1 - 149 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | w8a8 | Qualcomm® SA8295P | 3.269 ms | 0 - 45 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 0.917 ms | 0 - 40 MB | NPU |
| Real-ESRGAN-General-x4v3 | TFLITE | w8a8 | Snapdragon® 7 Gen 4 Mobile | 2.421 ms | 0 - 148 MB | NPU |
License
- The license for the original implementation of Real-ESRGAN-General-x4v3 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.