Stable diffusion for real-time music generation
Go to file
Veeral Patel 70fb80f565 Pin diffusers version 2023-11-30 13:09:40 -08:00
.github/workflows Add to support pip install -e . 2023-03-26 15:27:27 -07:00
integrations updating cog readme 2023-01-19 22:05:02 -08:00
riffusion Streamlit fixes and batch text to audio with multiple models 2023-04-11 21:36:09 -07:00
seed_images adding seed images 2022-12-13 23:33:52 -08:00
test Pull out black, ruff, mypy to individual CI actions 2023-03-26 14:40:54 -07:00
.gitignore Audio to audio handles interpolation within it 2023-01-14 11:45:45 -08:00
CITATION Update python project configuration files 2022-12-26 17:47:11 -08:00
LICENSE Add license 2022-12-13 02:47:48 +00:00 Make it possible to run the playground directly 2023-03-26 16:37:13 -07:00
VERSION Upgrade playground app to Streamlit 1.18+ 2023-03-26 16:25:47 -07:00
cog.yaml cog using common load fxn 2023-01-05 00:01:55 -08:00
pyproject.toml Pull out black, ruff, mypy to individual CI actions 2023-03-26 14:40:54 -07:00
requirements.txt Pin diffusers version 2023-11-30 13:09:40 -08:00
requirements_all.txt Pull out black, ruff, mypy to individual CI actions 2023-03-26 14:40:54 -07:00
requirements_dev.txt Add to support pip install -e . 2023-03-26 15:27:27 -07:00 Add py.typed for mypy use 2023-03-26 17:05:49 -07:00

🎸 Riffusion

CI status Python 3.9 | 3.10 MIT License

Riffusion is a library for real-time music and audio generation with stable diffusion.

Read about it at and try it at

This is the core repository for riffusion image and audio processing code.

  • Diffusion pipeline that performs prompt interpolation combined with image conditioning
  • Conversions between spectrogram images and audio clips
  • Command-line interface for common tasks
  • Interactive app using streamlit
  • Flask server to provide model inference via API
  • Various third party integrations

Related repositories:


If you build on this work, please cite it as follows:

  author = {Forsgren, Seth* and Martiros, Hayk*},
  title = {{Riffusion - Stable diffusion for real-time music generation}},
  url = {},
  year = {2022}


Tested in CI with Python 3.9 and 3.10.

It's highly recommended to set up a virtual Python environment with conda or virtualenv:

conda create --name riffusion python=3.9
conda activate riffusion

Install Python dependencies:

python -m pip install -r requirements.txt

In order to use audio formats other than WAV, ffmpeg is required.

sudo apt-get install ffmpeg          # linux
brew install ffmpeg                  # mac
conda install -c conda-forge ffmpeg  # conda

If torchaudio has no backend, you may need to install libsndfile. See this issue.

If you have an issue, try upgrading diffusers. Tested with 0.9 - 0.11.




cpu is supported but is quite slow.


cuda is the recommended and most performant backend.

To use with CUDA, make sure you have torch and torchaudio installed with CUDA support. See the install guide or stable wheels.

To generate audio in real-time, you need a GPU that can run stable diffusion with approximately 50 steps in under five seconds, such as a 3090 or A10G.

Test availability with:

import torch


The mps backend on Apple Silicon is supported for inference but some operations fall back to CPU, particularly for audio processing. You may need to set PYTORCH_ENABLE_MPS_FALLBACK=1.

In addition, this backend is not deterministic.

Test availability with:

import torch

Command-line interface

Riffusion comes with a command line interface for performing common tasks.

See available commands:

python -m riffusion.cli -h

Get help for a specific command:

python -m riffusion.cli image-to-audio -h


python -m riffusion.cli image-to-audio --image spectrogram_image.png --audio clip.wav

Riffusion Playground

Riffusion contains a streamlit app for interactive use and exploration.

Run with:

python -m riffusion.streamlit.playground

And access at

Riffusion Playground

Run the model server

Riffusion can be run as a flask server that provides inference via API. This server enables the web app to run locally.

Run with:

python -m riffusion.server --host --port 3013

You can specify --checkpoint with your own directory or huggingface ID in diffusers format.

Use the --device argument to specify the torch device to use.

The model endpoint is now available at via POST request.

Example input (see InferenceInput for the API):

  "alpha": 0.75,
  "num_inference_steps": 50,
  "seed_image_id": "og_beat",

  "start": {
    "prompt": "church bells on sunday",
    "seed": 42,
    "denoising": 0.75,
    "guidance": 7.0

  "end": {
    "prompt": "jazz with piano",
    "seed": 123,
    "denoising": 0.75,
    "guidance": 7.0

Example output (see InferenceOutput for the API):

  "image": "< base64 encoded JPEG image >",
  "audio": "< base64 encoded MP3 clip >"


Tests live in the test/ directory and are implemented with unittest.

To run all tests:

python -m unittest test/*

To run a single test:

python -m unittest test.audio_to_image_test

To preserve temporary outputs for debugging, set RIFFUSION_TEST_DEBUG:

RIFFUSION_TEST_DEBUG=1 python -m unittest test.audio_to_image_test

To run a single test case within a test:

python -m unittest test.audio_to_image_test -k AudioToImageTest.test_stereo

To run tests using a specific torch device, set RIFFUSION_TEST_DEVICE. Tests should pass with cpu, cuda, and mps backends.

Development Guide

Install additional packages for dev with python -m pip install -r requirements_dev.txt.

  • Linter: ruff
  • Formatter: black
  • Type checker: mypy

These are configured in pyproject.toml.

The results of mypy ., black ., and ruff . must be clean to accept a PR.

CI is run through GitHub Actions from .github/workflows/ci.yml.

Contributions are welcome through pull requests.