Stable-Diffusion-v2.1: Optimized for Mobile Deployment

State-of-the-art generative AI model used to generate detailed images conditioned on text descriptions

Generates high resolution images from text prompts using a latent diffusion model. This model uses CLIP ViT-L/14 as text encoder, U-Net based latent denoising, and VAE based decoder to generate the final image.

This model is an implementation of Stable-Diffusion-v2.1 found here.

This repository provides scripts to run Stable-Diffusion-v2.1 on Qualcomm® devices. More details on model performance across various devices, can be found here.

Model Details

Model Type: Image generation
Model Stats:
- Input: Text prompt to generate image
- Text Encoder Number of parameters: 340M
- UNet Number of parameters: 865M
- VAE Decoder Number of parameters: 83M
- Model size: 1GB

Model	Device	Chipset	Target Runtime	Inference Time (ms)	Peak Memory Range (MB)	Precision	Primary Compute Unit	Target Model
TextEncoderQuantizable	Samsung Galaxy S23	Snapdragon® 8 Gen 2	QNN	6.666 ms	0 - 3 MB	W8A16	NPU	Stable-Diffusion-v2.1.so
TextEncoderQuantizable	Samsung Galaxy S24	Snapdragon® 8 Gen 3	QNN	4.647 ms	0 - 20 MB	W8A16	NPU	Stable-Diffusion-v2.1.so
TextEncoderQuantizable	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	QNN	4.2 ms	0 - 15 MB	W8A16	NPU	Use Export Script
TextEncoderQuantizable	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN	6.84 ms	0 - 0 MB	W8A16	NPU	Use Export Script
TextEncoderQuantizable	SA7255P ADP	SA7255P	QNN	88.113 ms	0 - 9 MB	W8A16	NPU	Use Export Script
TextEncoderQuantizable	SA8255 (Proxy)	SA8255P Proxy	QNN	6.62 ms	0 - 3 MB	W8A16	NPU	Use Export Script
TextEncoderQuantizable	SA8650 (Proxy)	SA8650P Proxy	QNN	6.654 ms	0 - 2 MB	W8A16	NPU	Use Export Script
TextEncoderQuantizable	SA8775P ADP	SA8775P	QNN	7.869 ms	0 - 10 MB	W8A16	NPU	Use Export Script
TextEncoderQuantizable	QCS8275 (Proxy)	QCS8275 Proxy	QNN	88.113 ms	0 - 9 MB	W8A16	NPU	Use Export Script
TextEncoderQuantizable	QCS8550 (Proxy)	QCS8550 Proxy	QNN	6.636 ms	0 - 3 MB	W8A16	NPU	Use Export Script
TextEncoderQuantizable	QCS9075 (Proxy)	QCS9075 Proxy	QNN	7.869 ms	0 - 10 MB	W8A16	NPU	Use Export Script
UnetQuantizable	Samsung Galaxy S23	Snapdragon® 8 Gen 2	QNN	96.977 ms	0 - 3 MB	W8A16	NPU	Stable-Diffusion-v2.1.so
UnetQuantizable	Samsung Galaxy S24	Snapdragon® 8 Gen 3	QNN	69.178 ms	0 - 17 MB	W8A16	NPU	Stable-Diffusion-v2.1.so
UnetQuantizable	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	QNN	61.668 ms	0 - 14 MB	W8A16	NPU	Use Export Script
UnetQuantizable	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN	99.461 ms	0 - 0 MB	W8A16	NPU	Use Export Script
UnetQuantizable	SA7255P ADP	SA7255P	QNN	1467.935 ms	0 - 7 MB	W8A16	NPU	Use Export Script
UnetQuantizable	SA8255 (Proxy)	SA8255P Proxy	QNN	98.746 ms	0 - 2 MB	W8A16	NPU	Use Export Script
UnetQuantizable	SA8650 (Proxy)	SA8650P Proxy	QNN	97.177 ms	1 - 3 MB	W8A16	NPU	Use Export Script
UnetQuantizable	SA8775P ADP	SA8775P	QNN	110.665 ms	0 - 8 MB	W8A16	NPU	Use Export Script
UnetQuantizable	QCS8275 (Proxy)	QCS8275 Proxy	QNN	1467.935 ms	0 - 7 MB	W8A16	NPU	Use Export Script
UnetQuantizable	QCS8550 (Proxy)	QCS8550 Proxy	QNN	97.457 ms	0 - 3 MB	W8A16	NPU	Use Export Script
UnetQuantizable	QCS9075 (Proxy)	QCS9075 Proxy	QNN	110.665 ms	0 - 8 MB	W8A16	NPU	Use Export Script
VaeDecoderQuantizable	Samsung Galaxy S23	Snapdragon® 8 Gen 2	QNN	295.307 ms	0 - 71 MB	W8A16	NPU	Stable-Diffusion-v2.1.so
VaeDecoderQuantizable	Samsung Galaxy S24	Snapdragon® 8 Gen 3	QNN	223.33 ms	0 - 312 MB	W8A16	NPU	Stable-Diffusion-v2.1.so
VaeDecoderQuantizable	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	QNN	189.418 ms	0 - 356 MB	W8A16	NPU	Use Export Script
VaeDecoderQuantizable	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN	267.095 ms	0 - 0 MB	W8A16	NPU	Use Export Script
VaeDecoderQuantizable	SA7255P ADP	SA7255P	QNN	4460.526 ms	0 - 10 MB	W8A16	NPU	Use Export Script
VaeDecoderQuantizable	SA8255 (Proxy)	SA8255P Proxy	QNN	274.71 ms	0 - 2 MB	W8A16	NPU	Use Export Script
VaeDecoderQuantizable	SA8650 (Proxy)	SA8650P Proxy	QNN	269.652 ms	0 - 2 MB	W8A16	NPU	Use Export Script
VaeDecoderQuantizable	SA8775P ADP	SA8775P	QNN	301.141 ms	0 - 10 MB	W8A16	NPU	Use Export Script
VaeDecoderQuantizable	QCS8275 (Proxy)	QCS8275 Proxy	QNN	4460.526 ms	0 - 10 MB	W8A16	NPU	Use Export Script
VaeDecoderQuantizable	QCS8550 (Proxy)	QCS8550 Proxy	QNN	271.222 ms	0 - 3 MB	W8A16	NPU	Use Export Script
VaeDecoderQuantizable	QCS9075 (Proxy)	QCS9075 Proxy	QNN	301.141 ms	0 - 10 MB	W8A16	NPU	Use Export Script

Installation

Install the package via pip:

pip install "qai-hub-models[stable-diffusion-v2-1-quantized]" -f https://qaihub-public-python-wheels.s3.us-west-2.amazonaws.com/index.html

Configure Qualcomm® AI Hub to run this model on a cloud-hosted device

Sign-in to Qualcomm® AI Hub with your Qualcomm® ID. Once signed in navigate to Account -> Settings -> API Token.

With this API token, you can configure your client to run models on the cloud hosted devices.

qai-hub configure --api_token API_TOKEN

Navigate to docs for more information.

Demo off target

The package contains a simple end-to-end demo that downloads pre-trained weights and runs this model on a sample input.

python -m qai_hub_models.models.stable_diffusion_v2_1_quantized.demo

The above demo runs a reference implementation of pre-processing, model inference, and post processing.

NOTE: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above).

%run -m qai_hub_models.models.stable_diffusion_v2_1_quantized.demo

Run model on a cloud-hosted device

In addition to the demo, you can also run the model on a cloud-hosted Qualcomm® device. This script does the following:

Performance check on-device on a cloud-hosted device
Downloads compiled assets that can be deployed on-device for Android.
Accuracy check between PyTorch and on-device outputs.

python -m qai_hub_models.models.stable_diffusion_v2_1_quantized.export

Profiling Results
------------------------------------------------------------
TextEncoderQuantizable
Device                          : Samsung Galaxy S23 (13)
Runtime                         : QNN                    
Estimated inference time (ms)   : 6.7                    
Estimated peak memory usage (MB): [0, 3]                 
Total # Ops                     : 787                    
Compute Unit(s)                 : NPU (787 ops)          

------------------------------------------------------------
UnetQuantizable
Device                          : Samsung Galaxy S23 (13)
Runtime                         : QNN                    
Estimated inference time (ms)   : 97.0                   
Estimated peak memory usage (MB): [0, 3]                 
Total # Ops                     : 5891                   
Compute Unit(s)                 : NPU (5891 ops)         

------------------------------------------------------------
VaeDecoderQuantizable
Device                          : Samsung Galaxy S23 (13)
Runtime                         : QNN                    
Estimated inference time (ms)   : 295.3                  
Estimated peak memory usage (MB): [0, 71]                
Total # Ops                     : 189                    
Compute Unit(s)                 : NPU (189 ops)

Deploying compiled model to Android

The models can be deployed using multiple runtimes:

TensorFlow Lite (.tflite export): This tutorial provides a guide to deploy the .tflite model in an Android application.
QNN (.so export ): This sample app provides instructions on how to use the .so shared library in an Android application.

View on Qualcomm® AI Hub

Get more details on Stable-Diffusion-v2.1's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

The license for the original implementation of Stable-Diffusion-v2.1 can be found here.
The license for the compiled assets for on-device deployment can be found here

References

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.