README.md

---
library_name: pytorch
license: other
tags:
- real_time
- android
pipeline_tag: image-segmentation

---

![](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/ffnet_40s/web-assets/model_demo.png)

# FFNet-40S: Optimized for Qualcomm Devices

FFNet-40S is a "fuss-free network" that segments street scene images with per-pixel classes like road, sidewalk, and pedestrian. Trained on the Cityscapes dataset.

This is based on the implementation of FFNet-40S found [here](https://github.com/Qualcomm-AI-research/FFNet).
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/main/src/qai_hub_models/models/ffnet_40s) 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 2.42, ONNX Runtime 1.24.3 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/ffnet_40s/releases/v0.51.0/ffnet_40s-onnx-float.zip)
| ONNX | w8a8 | Universal | QAIRT 2.42, ONNX Runtime 1.24.3 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/ffnet_40s/releases/v0.51.0/ffnet_40s-onnx-w8a8.zip)
| QNN_DLC | float | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/ffnet_40s/releases/v0.51.0/ffnet_40s-qnn_dlc-float.zip)
| QNN_DLC | w8a8 | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/ffnet_40s/releases/v0.51.0/ffnet_40s-qnn_dlc-w8a8.zip)
| TFLITE | float | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/ffnet_40s/releases/v0.51.0/ffnet_40s-tflite-float.zip)
| TFLITE | w8a8 | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/ffnet_40s/releases/v0.51.0/ffnet_40s-tflite-w8a8.zip)

For more device-specific assets and performance metrics, visit **[FFNet-40S on Qualcomm® AI Hub](https://aihub.qualcomm.com/models/ffnet_40s)**.


### Option 2: Export with Custom Configurations

Use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/main/src/qai_hub_models/models/ffnet_40s) 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 [FFNet-40S on GitHub](https://github.com/qualcomm/ai-hub-models/blob/main/src/qai_hub_models/models/ffnet_40s) for usage instructions.

## Model Details

**Model Type:** Model_use_case.semantic_segmentation

**Model Stats:**
- Model checkpoint: ffnet40S_dBBB_cityscapes_state_dict_quarts
- Input resolution: 2048x1024
- Number of output classes: 19
- Number of parameters: 13.9M
- Model size (float): 53.1 MB
- Model size (w8a8): 13.5 MB

## Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit
|---|---|---|---|---|---|---
| FFNet-40S | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 12.367 ms | 13 - 234 MB | NPU
| FFNet-40S | ONNX | float | Snapdragon® X2 Elite | 13.428 ms | 22 - 22 MB | NPU
| FFNet-40S | ONNX | float | Snapdragon® X Elite | 31.615 ms | 24 - 24 MB | NPU
| FFNet-40S | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 22.552 ms | 29 - 304 MB | NPU
| FFNet-40S | ONNX | float | Qualcomm® QCS8550 (Proxy) | 32.116 ms | 24 - 27 MB | NPU
| FFNet-40S | ONNX | float | Qualcomm® QCS9075 | 47.946 ms | 24 - 27 MB | NPU
| FFNet-40S | ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 15.648 ms | 7 - 208 MB | NPU
| FFNet-40S | ONNX | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 6.541 ms | 2 - 197 MB | NPU
| FFNet-40S | ONNX | w8a8 | Snapdragon® X2 Elite | 6.9 ms | 9 - 9 MB | NPU
| FFNet-40S | ONNX | w8a8 | Snapdragon® X Elite | 10.478 ms | 8 - 8 MB | NPU
| FFNet-40S | ONNX | w8a8 | Snapdragon® 8 Gen 3 Mobile | 6.545 ms | 7 - 250 MB | NPU
| FFNet-40S | ONNX | w8a8 | Qualcomm® QCS6490 | 362.326 ms | 202 - 239 MB | CPU
| FFNet-40S | ONNX | w8a8 | Qualcomm® QCS8550 (Proxy) | 9.789 ms | 0 - 26 MB | NPU
| FFNet-40S | ONNX | w8a8 | Qualcomm® QCS9075 | 12.939 ms | 6 - 9 MB | NPU
| FFNet-40S | ONNX | w8a8 | Qualcomm® QCM6690 | 367.564 ms | 230 - 239 MB | CPU
| FFNet-40S | ONNX | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 7.725 ms | 1 - 194 MB | NPU
| FFNet-40S | ONNX | w8a8 | Snapdragon® 7 Gen 4 Mobile | 360.556 ms | 169 - 178 MB | CPU
| FFNet-40S | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 12.62 ms | 24 - 263 MB | NPU
| FFNet-40S | QNN_DLC | float | Snapdragon® X2 Elite | 14.624 ms | 24 - 24 MB | NPU
| FFNet-40S | QNN_DLC | float | Snapdragon® X Elite | 37.519 ms | 24 - 24 MB | NPU
| FFNet-40S | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 24.881 ms | 0 - 268 MB | NPU
| FFNet-40S | QNN_DLC | float | Qualcomm® QCS8275 (Proxy) | 135.577 ms | 24 - 219 MB | NPU
| FFNet-40S | QNN_DLC | float | Qualcomm® QCS8550 (Proxy) | 36.528 ms | 24 - 26 MB | NPU
| FFNet-40S | QNN_DLC | float | Qualcomm® SA8775P | 48.892 ms | 24 - 219 MB | NPU
| FFNet-40S | QNN_DLC | float | Qualcomm® QCS9075 | 62.187 ms | 24 - 52 MB | NPU
| FFNet-40S | QNN_DLC | float | Qualcomm® QCS8450 (Proxy) | 63.5 ms | 24 - 288 MB | NPU
| FFNet-40S | QNN_DLC | float | Qualcomm® SA7255P | 135.577 ms | 24 - 219 MB | NPU
| FFNet-40S | QNN_DLC | float | Qualcomm® SA8295P | 53.694 ms | 24 - 223 MB | NPU
| FFNet-40S | QNN_DLC | float | Snapdragon® 8 Elite For Galaxy Mobile | 18.242 ms | 22 - 245 MB | NPU
| FFNet-40S | QNN_DLC | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 5.238 ms | 6 - 232 MB | NPU
| FFNet-40S | QNN_DLC | w8a8 | Snapdragon® X2 Elite | 6.149 ms | 6 - 6 MB | NPU
| FFNet-40S | QNN_DLC | w8a8 | Snapdragon® X Elite | 15.897 ms | 6 - 6 MB | NPU
| FFNet-40S | QNN_DLC | w8a8 | Snapdragon® 8 Gen 3 Mobile | 10.494 ms | 6 - 251 MB | NPU
| FFNet-40S | QNN_DLC | w8a8 | Qualcomm® QCS6490 | 67.473 ms | 6 - 14 MB | NPU
| FFNet-40S | QNN_DLC | w8a8 | Qualcomm® QCS8275 (Proxy) | 32.925 ms | 6 - 203 MB | NPU
| FFNet-40S | QNN_DLC | w8a8 | Qualcomm® QCS8550 (Proxy) | 15.134 ms | 6 - 33 MB | NPU
| FFNet-40S | QNN_DLC | w8a8 | Qualcomm® SA8775P | 15.684 ms | 6 - 203 MB | NPU
| FFNet-40S | QNN_DLC | w8a8 | Qualcomm® QCS9075 | 18.528 ms | 6 - 14 MB | NPU
| FFNet-40S | QNN_DLC | w8a8 | Qualcomm® QCM6690 | 126.872 ms | 6 - 238 MB | NPU
| FFNet-40S | QNN_DLC | w8a8 | Qualcomm® QCS8450 (Proxy) | 21.473 ms | 6 - 252 MB | NPU
| FFNet-40S | QNN_DLC | w8a8 | Qualcomm® SA7255P | 32.925 ms | 6 - 203 MB | NPU
| FFNet-40S | QNN_DLC | w8a8 | Qualcomm® SA8295P | 20.051 ms | 6 - 203 MB | NPU
| FFNet-40S | QNN_DLC | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 7.086 ms | 6 - 217 MB | NPU
| FFNet-40S | QNN_DLC | w8a8 | Snapdragon® 7 Gen 4 Mobile | 19.337 ms | 6 - 223 MB | NPU
| FFNet-40S | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 12.623 ms | 2 - 252 MB | NPU
| FFNet-40S | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 24.979 ms | 2 - 303 MB | NPU
| FFNet-40S | TFLITE | float | Qualcomm® QCS8275 (Proxy) | 135.665 ms | 2 - 210 MB | NPU
| FFNet-40S | TFLITE | float | Qualcomm® QCS8550 (Proxy) | 36.208 ms | 2 - 5 MB | NPU
| FFNet-40S | TFLITE | float | Qualcomm® SA8775P | 48.985 ms | 2 - 210 MB | NPU
| FFNet-40S | TFLITE | float | Qualcomm® QCS9075 | 62.379 ms | 0 - 56 MB | NPU
| FFNet-40S | TFLITE | float | Qualcomm® QCS8450 (Proxy) | 63.678 ms | 3 - 303 MB | NPU
| FFNet-40S | TFLITE | float | Qualcomm® SA7255P | 135.665 ms | 2 - 210 MB | NPU
| FFNet-40S | TFLITE | float | Qualcomm® SA8295P | 53.691 ms | 2 - 217 MB | NPU
| FFNet-40S | TFLITE | float | Snapdragon® 8 Elite For Galaxy Mobile | 18.087 ms | 1 - 239 MB | NPU
| FFNet-40S | TFLITE | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 2.905 ms | 0 - 224 MB | NPU
| FFNet-40S | TFLITE | w8a8 | Snapdragon® 8 Gen 3 Mobile | 5.455 ms | 1 - 249 MB | NPU
| FFNet-40S | TFLITE | w8a8 | Qualcomm® QCS6490 | 52.01 ms | 1 - 23 MB | NPU
| FFNet-40S | TFLITE | w8a8 | Qualcomm® QCS8275 (Proxy) | 20.065 ms | 1 - 196 MB | NPU
| FFNet-40S | TFLITE | w8a8 | Qualcomm® QCS8550 (Proxy) | 7.509 ms | 1 - 4 MB | NPU
| FFNet-40S | TFLITE | w8a8 | Qualcomm® SA8775P | 8.219 ms | 1 - 197 MB | NPU
| FFNet-40S | TFLITE | w8a8 | Qualcomm® QCS9075 | 9.582 ms | 1 - 23 MB | NPU
| FFNet-40S | TFLITE | w8a8 | Qualcomm® QCM6690 | 98.38 ms | 1 - 231 MB | NPU
| FFNet-40S | TFLITE | w8a8 | Qualcomm® QCS8450 (Proxy) | 12.372 ms | 0 - 246 MB | NPU
| FFNet-40S | TFLITE | w8a8 | Qualcomm® SA7255P | 20.065 ms | 1 - 196 MB | NPU
| FFNet-40S | TFLITE | w8a8 | Qualcomm® SA8295P | 11.697 ms | 1 - 200 MB | NPU
| FFNet-40S | TFLITE | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 3.992 ms | 0 - 208 MB | NPU
| FFNet-40S | TFLITE | w8a8 | Snapdragon® 7 Gen 4 Mobile | 11.552 ms | 1 - 216 MB | NPU

## License
* The license for the original implementation of FFNet-40S can be found
  [here](https://github.com/Qualcomm-AI-research/FFNet/blob/master/LICENSE).

## References
* [Simple and Efficient Architectures for Semantic Segmentation](https://arxiv.org/abs/2206.08236)
* [Source Model Implementation](https://github.com/Qualcomm-AI-research/FFNet)

## 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).