Riffusion: Optimized for Mobile Deployment

State-of-the-art generative AI model used to generate spectrogram images given any text input. These spectrograms can be converted into audio clips

Generates high resolution spectrograms 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 Riffusion found here.

This repository provides scripts to run Riffusion 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 spectrogram 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
TextEncoder_Quantized	Samsung Galaxy S23	Snapdragon® 8 Gen 2	QNN	7.045 ms	0 - 67 MB	INT8	NPU	Riffusion.bin
TextEncoder_Quantized	Samsung Galaxy S24	Snapdragon® 8 Gen 3	QNN	4.789 ms	0 - 161 MB	INT8	NPU	Riffusion.bin
TextEncoder_Quantized	QCS8550 (Proxy)	QCS8550 Proxy	QNN	6.715 ms	0 - 1 MB	UINT16	NPU	Use Export Script
TextEncoder_Quantized	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN	7.594 ms	0 - 0 MB	INT8	NPU	Use Export Script
VAEDecoder_Quantized	Samsung Galaxy S23	Snapdragon® 8 Gen 2	QNN	233.844 ms	0 - 46 MB	INT8	NPU	Riffusion.bin
VAEDecoder_Quantized	Samsung Galaxy S24	Snapdragon® 8 Gen 3	QNN	175.734 ms	0 - 64 MB	INT8	NPU	Riffusion.bin
VAEDecoder_Quantized	QCS8550 (Proxy)	QCS8550 Proxy	QNN	239.643 ms	0 - 1 MB	UINT16	NPU	Use Export Script
VAEDecoder_Quantized	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN	227.581 ms	0 - 0 MB	INT8	NPU	Use Export Script
UNet_Quantized	Samsung Galaxy S23	Snapdragon® 8 Gen 2	QNN	127.531 ms	0 - 13 MB	INT8	NPU	Riffusion.bin
UNet_Quantized	Samsung Galaxy S24	Snapdragon® 8 Gen 3	QNN	90.167 ms	0 - 1750 MB	INT8	NPU	Riffusion.bin
UNet_Quantized	QCS8550 (Proxy)	QCS8550 Proxy	QNN	128.206 ms	1 - 2 MB	UINT16	NPU	Use Export Script
UNet_Quantized	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN	129.856 ms	0 - 0 MB	INT8	NPU	Use Export Script

Installation

This model can be installed as a Python package via pip.

pip install "qai-hub-models[riffusion_quantized]"

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 on-device

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.riffusion_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.riffusion_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.riffusion_quantized.export

Profiling Results
------------------------------------------------------------
TextEncoder_Quantized
Device                          : Samsung Galaxy S23 (13)
Runtime                         : QNN                    
Estimated inference time (ms)   : 7.0                    
Estimated peak memory usage (MB): [0, 67]                
Total # Ops                     : 569                    
Compute Unit(s)                 : NPU (569 ops)          

------------------------------------------------------------
VAEDecoder_Quantized
Device                          : Samsung Galaxy S23 (13)
Runtime                         : QNN                    
Estimated inference time (ms)   : 233.8                  
Estimated peak memory usage (MB): [0, 46]                
Total # Ops                     : 170                    
Compute Unit(s)                 : NPU (170 ops)          

------------------------------------------------------------
UNet_Quantized
Device                          : Samsung Galaxy S23 (13)
Runtime                         : QNN                    
Estimated inference time (ms)   : 127.5                  
Estimated peak memory usage (MB): [0, 13]                
Total # Ops                     : 4933                   
Compute Unit(s)                 : NPU (4933 ops)

How does this work?

This export script leverages Qualcomm® AI Hub to optimize, validate, and deploy this model on-device. Lets go through each step below in detail:

Step 1: Compile model for on-device deployment

To compile a PyTorch model for on-device deployment, we first trace the model in memory using the jit.trace and then call the submit_compile_job API.

import torch

import qai_hub as hub
from qai_hub_models.models.riffusion_quantized import Model

# Load the model
model = Model.from_pretrained()
text_encoder_model = model.text_encoder
unet_model = model.unet
vae_decoder_model = model.vae_decoder

# Device
device = hub.Device("Samsung Galaxy S23")

# Trace model
text_encoder_input_shape = text_encoder_model.get_input_spec()
text_encoder_sample_inputs = text_encoder_model.sample_inputs()

traced_text_encoder_model = torch.jit.trace(text_encoder_model, [torch.tensor(data[0]) for _, data in text_encoder_sample_inputs.items()])

# Compile model on a specific device
text_encoder_compile_job = hub.submit_compile_job(
    model=traced_text_encoder_model ,
    device=device,
    input_specs=text_encoder_model.get_input_spec(),
)

# Get target model to run on-device
text_encoder_target_model = text_encoder_compile_job.get_target_model()
# Trace model
unet_input_shape = unet_model.get_input_spec()
unet_sample_inputs = unet_model.sample_inputs()

traced_unet_model = torch.jit.trace(unet_model, [torch.tensor(data[0]) for _, data in unet_sample_inputs.items()])

# Compile model on a specific device
unet_compile_job = hub.submit_compile_job(
    model=traced_unet_model ,
    device=device,
    input_specs=unet_model.get_input_spec(),
)

# Get target model to run on-device
unet_target_model = unet_compile_job.get_target_model()
# Trace model
vae_decoder_input_shape = vae_decoder_model.get_input_spec()
vae_decoder_sample_inputs = vae_decoder_model.sample_inputs()

traced_vae_decoder_model = torch.jit.trace(vae_decoder_model, [torch.tensor(data[0]) for _, data in vae_decoder_sample_inputs.items()])

# Compile model on a specific device
vae_decoder_compile_job = hub.submit_compile_job(
    model=traced_vae_decoder_model ,
    device=device,
    input_specs=vae_decoder_model.get_input_spec(),
)

# Get target model to run on-device
vae_decoder_target_model = vae_decoder_compile_job.get_target_model()

Step 2: Performance profiling on cloud-hosted device

After uploading compiled models from step 1. Models can be profiled model on-device using the target_model. Note that this scripts runs the model on a device automatically provisioned in the cloud. Once the job is submitted, you can navigate to a provided job URL to view a variety of on-device performance metrics.


# Device
device = hub.Device("Samsung Galaxy S23")
profile_job_textencoder_quantized = hub.submit_profile_job(
    model=model_textencoder_quantized,
    device=device,
)
profile_job_unet_quantized = hub.submit_profile_job(
    model=model_unet_quantized,
    device=device,
)
profile_job_vaedecoder_quantized = hub.submit_profile_job(
    model=model_vaedecoder_quantized,
    device=device,
)

Step 3: Verify on-device accuracy

To verify the accuracy of the model on-device, you can run on-device inference on sample input data on the same cloud hosted device.


input_data_textencoder_quantized = model.text_encoder.sample_inputs()
inference_job_textencoder_quantized = hub.submit_inference_job(
    model=model_textencoder_quantized,
    device=device,
    inputs=input_data_textencoder_quantized,
)
on_device_output_textencoder_quantized = inference_job_textencoder_quantized.download_output_data()

input_data_unet_quantized = model.unet.sample_inputs()
inference_job_unet_quantized = hub.submit_inference_job(
    model=model_unet_quantized,
    device=device,
    inputs=input_data_unet_quantized,
)
on_device_output_unet_quantized = inference_job_unet_quantized.download_output_data()

input_data_vaedecoder_quantized = model.vae_decoder.sample_inputs()
inference_job_vaedecoder_quantized = hub.submit_inference_job(
    model=model_vaedecoder_quantized,
    device=device,
    inputs=input_data_vaedecoder_quantized,
)
on_device_output_vaedecoder_quantized = inference_job_vaedecoder_quantized.download_output_data()

With the output of the model, you can compute like PSNR, relative errors or spot check the output with expected output.

Note: This on-device profiling and inference requires access to Qualcomm® AI Hub. Sign up for access.

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 / .bin export ): This sample app provides instructions on how to use the .so shared library or .bin context binary in an Android application.

View on Qualcomm® AI Hub

Get more details on Riffusion's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

The license for the original implementation of Riffusion 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.

Usage and Limitations

Model may not be used for or in connection with any of the following applications:

Accessing essential private and public services and benefits;
Administration of justice and democratic processes;
Assessing or recognizing the emotional state of a person;
Biometric and biometrics-based systems, including categorization of persons based on sensitive characteristics;
Education and vocational training;
Employment and workers management;
Exploitation of the vulnerabilities of persons resulting in harmful behavior;
General purpose social scoring;
Law enforcement;
Management and operation of critical infrastructure;
Migration, asylum and border control management;
Predictive policing;
Real-time remote biometric identification in public spaces;
Recommender systems of social media platforms;
Scraping of facial images (from the internet or otherwise); and/or
Subliminal manipulation

qualcomm
/

Riffusion