Meet Mustango, an exciting addition to the vibrant landscape of Multimodal Large Language Models designed for controlled music generation. Mustango leverages Latent Diffusion Model (LDM), Flan-T5, and musical features to do the magic!

🔥 Live demo available on Replicate and HuggingFace.

Quickstart Guide

Generate music from a text prompt:

import IPython
import soundfile as sf
from mustango import Mustango

model = Mustango("declare-lab/mustango")

prompt = "This is a new age piece. There is a flute playing the main melody with a lot of staccato notes. The rhythmic background consists of a medium tempo electronic drum beat with percussive elements all over the spectrum. There is a playful atmosphere to the piece. This piece can be used in the soundtrack of a children's TV show or an advertisement jingle."

music = model.generate(prompt)
sf.write(f"{prompt}.wav", music, samplerate=16000)
IPython.display.Audio(data=music, rate=16000)

Installation

git clone https://github.com/AMAAI-Lab/mustango
cd mustango
pip install -r requirements.txt
cd diffusers
pip install -e .

Datasets

The MusicBench dataset contains 52k music fragments with a rich music-specific text caption.

Subjective Evaluation by Expert Listeners

Training

We use the accelerate package from Hugging Face for multi-gpu training. Run accelerate config from terminal and set up your run configuration by the answering the questions asked.

You can now train Mustango on the MusicBench dataset using:

accelerate launch train.py \
--text_encoder_name="google/flan-t5-large" \
--scheduler_name="stabilityai/stable-diffusion-2-1" \
--unet_model_config="configs/diffusion_model_config_munet.json" \
--model_type Mustango --freeze_text_encoder --uncondition_all --uncondition_single \
--drop_sentences --random_pick_text_column --snr_gamma 5 \

The --model_type flag allows to choose either Mustango, or Tango to be trained with the same code. However, do note that you also need to change --unet_model_config to the relevant config: diffusion_model_config_munet for Mustango; diffusion_model_config for Tango.

The arguments --uncondition_all, --uncondition_single, --drop_sentences control the dropout functions as per Section 5.2 in our paper. The argument of --random_pick_text_column allows to randomly pick between two input text prompts - in the case of MusicBench, we pick between ChatGPT rephrased captions and original enhanced MusicCaps prompts, as depicted in Figure 1 in our paper.

Recommended training time from scratch on MusicBench is at least 40 epochs.

Model Zoo

We have released the following models:

Mustango Pretrained: https://huggingface.co/declare-lab/mustango-pretrained

Mustango: https://huggingface.co/declare-lab/mustango

Citation

Please consider citing the following article if you found our work useful:

@inproceedings{melechovsky2024mustango,
  title={Mustango: Toward Controllable Text-to-Music Generation},
  author={Melechovsky, Jan and Guo, Zixun and Ghosal, Deepanway and Majumder, Navonil and Herremans, Dorien and Poria, Soujanya},
  booktitle={Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)},
  pages={8286--8309},
  year={2024}
}