---
library_name: pytorch
license: other
tags:
- android
pipeline_tag: image-to-image

---

![](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/nafnet_denoise/web-assets/model_demo.png)

# NAFNet-DeNoise: Optimized for Qualcomm Devices

NAFNET is designed for lightweight real-time denoising of images.

This is based on the implementation of NAFNet-DeNoise found [here](https://github.com/megvii-research/NAFNet.git).
This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/v0.58.0/src/qai_hub_models/models/nafnet_denoise) library to export with custom configurations. More details on model performance across various devices, can be found [here](#performance-summary).

Qualcomm AI Hub Models uses [Qualcomm AI Hub Workbench](https://workbench.aihub.qualcomm.com) to compile, profile, and evaluate this model. [Sign up](https://myaccount.qualcomm.com/signup) 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 1.25.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/nafnet_denoise/releases/v0.58.0/nafnet_denoise-onnx-float.zip)
| ONNX | w8a16 | Universal | QAIRT 2.45, ONNX Runtime 1.25.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/nafnet_denoise/releases/v0.58.0/nafnet_denoise-onnx-w8a16.zip)
| QNN_DLC | float | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/nafnet_denoise/releases/v0.58.0/nafnet_denoise-qnn_dlc-float.zip)
| QNN_DLC | w8a16 | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/nafnet_denoise/releases/v0.58.0/nafnet_denoise-qnn_dlc-w8a16.zip)
| TFLITE | float | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/nafnet_denoise/releases/v0.58.0/nafnet_denoise-tflite-float.zip)

For more device-specific assets and performance metrics, visit **[NAFNet-DeNoise on Qualcomm® AI Hub](https://aihub.qualcomm.com/models/nafnet_denoise)**.


### Option 2: Export with Custom Configurations

Use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/v0.58.0/src/qai_hub_models/models/nafnet_denoise) 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 [NAFNet-DeNoise on GitHub](https://github.com/qualcomm/ai-hub-models/blob/v0.58.0/src/qai_hub_models/models/nafnet_denoise) for usage instructions.

## Model Details

**Model Type:** Model_use_case.image_editing

**Model Stats:**
- Model checkpoint: NAFNet-SIDD-width64
- Input resolution: 256x256
- Number of parameters: 115.98M
- Model size (float): 463.93 MB

## Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit
|---|---|---|---|---|---|---
| NAFNet-DeNoise | ONNX | float | Snapdragon® X2 Elite | 14.863 ms | 4 - 4 MB | NPU
| NAFNet-DeNoise | ONNX | float | Snapdragon® X Elite | 33.829 ms | 228 - 228 MB | NPU
| NAFNet-DeNoise | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 27.109 ms | 2 - 849 MB | NPU
| NAFNet-DeNoise | ONNX | float | Snapdragon® 8 Gen 1 Mobile | 62.276 ms | 3 - 722 MB | NPU
| NAFNet-DeNoise | ONNX | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 35.366 ms | 0 - 259 MB | NPU
| NAFNet-DeNoise | ONNX | float | Qualcomm® QCS8450 | 62.276 ms | 3 - 722 MB | NPU
| NAFNet-DeNoise | ONNX | float | Qualcomm® Dragonwing™ IQ-9075 | 42.82 ms | 2 - 5 MB | NPU
| NAFNet-DeNoise | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 14.462 ms | 2 - 538 MB | NPU
| NAFNet-DeNoise | ONNX | float | Snapdragon® 8 Elite Mobile | 19.823 ms | 2 - 498 MB | NPU
| NAFNet-DeNoise | ONNX | float | Qualcomm® Dragonwing™ Q-8750 | 19.823 ms | 2 - 498 MB | NPU
| NAFNet-DeNoise | ONNX | float | Qualcomm® Dragonwing™ IQ-X7181 | 33.829 ms | 228 - 228 MB | NPU
| NAFNet-DeNoise | ONNX | w8a16 | Snapdragon® X2 Elite | 14.349 ms | 1 - 1 MB | NPU
| NAFNet-DeNoise | ONNX | w8a16 | Snapdragon® X Elite | 33.215 ms | 119 - 119 MB | NPU
| NAFNet-DeNoise | ONNX | w8a16 | Snapdragon® 8 Gen 3 Mobile | 24.074 ms | 0 - 693 MB | NPU
| NAFNet-DeNoise | ONNX | w8a16 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 33.012 ms | 1 - 5 MB | NPU
| NAFNet-DeNoise | ONNX | w8a16 | Qualcomm® Dragonwing™ IQ-9075 | 38.669 ms | 1 - 4 MB | NPU
| NAFNet-DeNoise | ONNX | w8a16 | Snapdragon® 7 Gen 4 Mobile | 38.486 ms | 1 - 725 MB | NPU
| NAFNet-DeNoise | ONNX | w8a16 | Snapdragon® 8 Elite Gen 5 Mobile | 13.044 ms | 0 - 684 MB | NPU
| NAFNet-DeNoise | ONNX | w8a16 | Qualcomm® Dragonwing™ Q-6690 | 296.179 ms | 0 - 916 MB | NPU
| NAFNet-DeNoise | ONNX | w8a16 | Snapdragon® 8 Elite Mobile | 19.191 ms | 1 - 563 MB | NPU
| NAFNet-DeNoise | ONNX | w8a16 | Qualcomm® Dragonwing™ Q-7790 | 38.486 ms | 1 - 725 MB | NPU
| NAFNet-DeNoise | ONNX | w8a16 | Qualcomm® Dragonwing™ Q-8750 | 19.191 ms | 1 - 563 MB | NPU
| NAFNet-DeNoise | ONNX | w8a16 | Qualcomm® Dragonwing™ IQ-X7181 | 33.215 ms | 119 - 119 MB | NPU
| NAFNet-DeNoise | QNN_DLC | float | Snapdragon® X2 Elite | 15.724 ms | 1 - 1 MB | NPU
| NAFNet-DeNoise | QNN_DLC | float | Snapdragon® X Elite | 35.781 ms | 1 - 1 MB | NPU
| NAFNet-DeNoise | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 28.339 ms | 0 - 890 MB | NPU
| NAFNet-DeNoise | QNN_DLC | float | Snapdragon® 8 Gen 1 Mobile | 60.784 ms | 0 - 729 MB | NPU
| NAFNet-DeNoise | QNN_DLC | float | Qualcomm® QCS8275 | 136.111 ms | 1 - 545 MB | NPU
| NAFNet-DeNoise | QNN_DLC | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 37.933 ms | 1 - 3 MB | NPU
| NAFNet-DeNoise | QNN_DLC | float | Qualcomm® SA8775P | 43.449 ms | 1 - 557 MB | NPU
| NAFNet-DeNoise | QNN_DLC | float | Qualcomm® SA8650P | 43.449 ms | 1 - 557 MB | NPU
| NAFNet-DeNoise | QNN_DLC | float | Qualcomm® SA8255P | 43.449 ms | 1 - 557 MB | NPU
| NAFNet-DeNoise | QNN_DLC | float | Qualcomm® QCS8450 | 60.784 ms | 0 - 729 MB | NPU
| NAFNet-DeNoise | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-9075 | 43.535 ms | 1 - 4 MB | NPU
| NAFNet-DeNoise | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 14.715 ms | 0 - 622 MB | NPU
| NAFNet-DeNoise | QNN_DLC | float | Qualcomm® SA7255P | 136.111 ms | 1 - 545 MB | NPU
| NAFNet-DeNoise | QNN_DLC | float | Snapdragon® 8 Elite Mobile | 20.425 ms | 1 - 542 MB | NPU
| NAFNet-DeNoise | QNN_DLC | float | Qualcomm® SA8295P | 47.984 ms | 1 - 399 MB | NPU
| NAFNet-DeNoise | QNN_DLC | float | Qualcomm® Dragonwing™ Q-8750 | 20.425 ms | 1 - 542 MB | NPU
| NAFNet-DeNoise | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-X7181 | 35.781 ms | 1 - 1 MB | NPU
| NAFNet-DeNoise | QNN_DLC | w8a16 | Snapdragon® X2 Elite | 15.127 ms | 0 - 0 MB | NPU
| NAFNet-DeNoise | QNN_DLC | w8a16 | Snapdragon® X Elite | 33.151 ms | 0 - 0 MB | NPU
| NAFNet-DeNoise | QNN_DLC | w8a16 | Snapdragon® 8 Gen 3 Mobile | 22.533 ms | 0 - 688 MB | NPU
| NAFNet-DeNoise | QNN_DLC | w8a16 | Qualcomm® QCS8275 | 65.065 ms | 1 - 577 MB | NPU
| NAFNet-DeNoise | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 31.673 ms | 0 - 553 MB | NPU
| NAFNet-DeNoise | QNN_DLC | w8a16 | Qualcomm® SA8775P | 31.612 ms | 1 - 577 MB | NPU
| NAFNet-DeNoise | QNN_DLC | w8a16 | Qualcomm® SA8650P | 31.612 ms | 1 - 577 MB | NPU
| NAFNet-DeNoise | QNN_DLC | w8a16 | Qualcomm® SA8255P | 31.612 ms | 1 - 577 MB | NPU
| NAFNet-DeNoise | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ IQ-9075 | 31.793 ms | 0 - 3 MB | NPU
| NAFNet-DeNoise | QNN_DLC | w8a16 | Snapdragon® 7 Gen 4 Mobile | 38.515 ms | 0 - 812 MB | NPU
| NAFNet-DeNoise | QNN_DLC | w8a16 | Snapdragon® 8 Elite Gen 5 Mobile | 13.277 ms | 0 - 706 MB | NPU
| NAFNet-DeNoise | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ Q-6690 | 322.373 ms | 0 - 954 MB | NPU
| NAFNet-DeNoise | QNN_DLC | w8a16 | Qualcomm® SA7255P | 65.065 ms | 1 - 577 MB | NPU
| NAFNet-DeNoise | QNN_DLC | w8a16 | Snapdragon® 8 Elite Mobile | 21.31 ms | 0 - 586 MB | NPU
| NAFNet-DeNoise | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ Q-7790 | 38.515 ms | 0 - 812 MB | NPU
| NAFNet-DeNoise | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ Q-8750 | 21.31 ms | 0 - 586 MB | NPU
| NAFNet-DeNoise | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ IQ-X7181 | 33.151 ms | 0 - 0 MB | NPU
| NAFNet-DeNoise | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 28.399 ms | 1 - 1122 MB | NPU
| NAFNet-DeNoise | TFLITE | float | Snapdragon® 8 Gen 1 Mobile | 63.491 ms | 1 - 961 MB | NPU
| NAFNet-DeNoise | TFLITE | float | Qualcomm® QCS8275 | 136.649 ms | 1 - 765 MB | NPU
| NAFNet-DeNoise | TFLITE | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 36.98 ms | 1 - 4 MB | NPU
| NAFNet-DeNoise | TFLITE | float | Qualcomm® SA8775P | 43.67 ms | 1 - 778 MB | NPU
| NAFNet-DeNoise | TFLITE | float | Qualcomm® SA8650P | 43.67 ms | 1 - 778 MB | NPU
| NAFNet-DeNoise | TFLITE | float | Qualcomm® SA8255P | 43.67 ms | 1 - 778 MB | NPU
| NAFNet-DeNoise | TFLITE | float | Qualcomm® QCS8450 | 63.491 ms | 1 - 961 MB | NPU
| NAFNet-DeNoise | TFLITE | float | Qualcomm® Dragonwing™ IQ-9075 | 43.821 ms | 1 - 233 MB | NPU
| NAFNet-DeNoise | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 14.757 ms | 1 - 849 MB | NPU
| NAFNet-DeNoise | TFLITE | float | Qualcomm® SA7255P | 136.649 ms | 1 - 765 MB | NPU
| NAFNet-DeNoise | TFLITE | float | Snapdragon® 8 Elite Mobile | 19.997 ms | 1 - 759 MB | NPU
| NAFNet-DeNoise | TFLITE | float | Qualcomm® SA8295P | 50.755 ms | 1 - 630 MB | NPU
| NAFNet-DeNoise | TFLITE | float | Qualcomm® Dragonwing™ Q-8750 | 19.997 ms | 1 - 759 MB | NPU

## License
* The license for the original implementation of NAFNet-DeNoise can be found
  [here](https://github.com/megvii-research/NAFNet/blob/main/LICENSE).

## References
* [Simple Baselines for Image Restoration](https://arxiv.org/abs/2204.04676)
* [Source Model Implementation](https://github.com/megvii-research/NAFNet.git)

## Community
* Join [our AI Hub Slack community](https://aihub.qualcomm.com/community/slack) to collaborate, post questions and learn more about on-device AI.
* For questions or feedback please [reach out to us](mailto:ai-hub-support@qti.qualcomm.com).