TangoFlux: Super Fast and Faithful Text to Audio Generation with Flow Matching and Clap-Ranked Preference Optimization

Model Overview

TangoFlux consists of FluxTransformer blocks which are Diffusion Transformer (DiT) and Multimodal Diffusion Transformer (MMDiT), conditioned on textual prompt and duration embedding to generate audio at 44.1kHz up to 30 seconds. TangoFlux learns a rectified flow trajectory from audio latent representation encoded by a variational autoencoder (VAE). The TangoFlux training pipeline consists of three stages: pre-training, fine-tuning, and preference optimization. TangoFlux is aligned via CRPO which iteratively generates new synthetic data and constructs preference pairs to perform preference optimization.

Getting Started

Get TangoFlux from our GitHub repo https://github.com/declare-lab/TangoFlux with

Bash

pip install git+https://github.com/declare-lab/TangoFlux

The model will be automatically downloaded and saved in a cache. The subsequent runs will load the model directly from the cache.

The generate function uses 25 steps by default to sample from the flow model. We recommend using 50 steps for generating better quality audios. This comes at the cost of increased run-time.

Python

import torchaudio
from tangoflux import TangoFluxInference
from IPython.display import Audio

model = TangoFluxInference(name='declare-lab/TangoFlux')
audio = model.generate('Hammer slowly hitting the wooden table', steps=50, duration=10)

Audio(data=audio, rate=44100)

License

The TangoFlux checkpoints are for non-commercial research use only. They are subject to the Stable Audio Open’s license, WavCap’s license, and the original licenses accompanying each training dataset.

Citation

https://arxiv.org/abs/2412.21037

Bibtex

@misc{hung2024tangofluxsuperfastfaithful,
      title={TangoFlux: Super Fast and Faithful Text to Audio Generation with Flow Matching and Clap-Ranked Preference Optimization}, 
      author={Chia-Yu Hung and Navonil Majumder and Zhifeng Kong and Ambuj Mehrish and Rafael Valle and Bryan Catanzaro and Soujanya Poria},
      year={2024},
      eprint={2412.21037},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2412.21037}, 
}