Denoising Diffusion Bridge Models (ICLR 2024)

Implementation of Denoising Diffusion Bridge Models.

Goal of this project is to create the core for future research projects based on bridge models.

Code is based on the official implementation: ddbm and has simplified interface in training and sampling using torchrun commands instead of mpi, and also enables training and sampling on custom pairwise datasets.

Dependencies

1. conda create -n ddbm python=3.12
2. conda activate ddbm
3. pip install -r requirements.txt

Datasets

Code is reorganized to work with custom datasets. As an example, pix2pix datasets are utilized for trainings and samplings.

Dataset preparations are at pix2pix_utils.py

Train & sample

To run training (you can also use one device):

torchrun --nproc_per_node=2 --nnodes=1 ddbm_train_dist.py --devices 0 1

To run sampling:

torchrun --nproc_per_node=1 --nnodes=1 ddbm_sample_dist.py --devices 0

Example of retriving sampling results you can find at ddbm.ipynb.

Citation

@article{zhou2023denoising,
  title={Denoising diffusion bridge models},
  author={Zhou, Linqi and Lou, Aaron and Khanna, Samar and Ermon, Stefano},
  journal={arXiv preprint arXiv:2309.16948},
  year={2023}
}