MGDM Algorithm#

— A Mixture-Based Framework for Guiding Diffusion Models —

Audio separation on Slakh2100 dataset#

MGDM can be used to separate sources given a mixture audio. We showcase that on Slakh2100 test dataset.

Here, we provide several samples tracks, where

  • Input is the mixture sound that is given to the algorithm

  • Bass, Drums, Guitar, and Piano are the outputs of the algorithm

Track01876 (00:11 - 00:44)#

Input

Bass

Drums

Guitar

Piano
Track01877 (01:17 - 01:50)#

Input

Bass

Drums

Guitar

Piano
Track01884 (02:56 - 03:29)#

Input

Bass

Drums

Guitar

Piano
Track02005 (01:06 - 01:39)#

Input

Bass

Drums

Guitar

Piano
Track02084 (02:12 - 02:45)#

Input

Bass

Drums

Guitar

Piano

Cite#

MGDM is part of a series of publications that explore training-free approach to guide pre-trained diffusion models.

@article{janati2025mgdm,
    title  ={A Mixture-Based Framework for Guiding Diffusion Models},
    author ={Janati, Yazid and Moufad, Badr and Abou El Qassim, Mehdi and
             Durmus, Alain and Moulines, Eric and Olsson, Jimmy},
    journal={preprint},
    year   ={2025}
}

@article{moufad2025mgps,
    title  ={Variational Diffusion Posterior Sampling with Midpoint Guidance},
    author ={Moufad, Badr and Janati, Yazid and Bedin, Lisa and
             Durmus, Alain and Douc, Randal and Moulines, Eric and Olsson, Jimmy},
    journal={ICLR},
    year   ={2025}
}

@inproceedings{janati2024dcps,
    title    ={Divide-and-Conquer Posterior Sampling for Denoising Diffusion Priors},
    author   ={Janati, Yazid and Moufad, Badr and Durmus, Alain and
               Moulines, Eric and Olsson, Jimmy},
    booktitle={Neurips},
    year     ={2024}
}