Official PyTorch implementation of the NeurIPS 2023 paper Direct Diffusion Bridge using Data Consistency for Inverse Problems. Code modified from I2SB.

by Hyungjin Chung, Jeongsol Kim, and Jong Chul Ye

The pre-trained checkpoints and dependencies all follow I2SB. Please consult the original source. We list the steps to make the repository self-contained.

conda env create --file requirements.yaml python=3
conda activate cddb

One can download the checkpoints by simply running

bash scripts/download_checkpoint.sh $DEG_NAME

In this work, we consider $DEG_NAME: sr4x-pool, sr4x-bicubic, blur-uni, blur-gauss, jpeg-10, but others can be used. One can also manually download the model weights and place it under ./results/{$DEG_NAME}

Simply run

./scripts/sample.sh

Use use-cddb or use-cddb-deep flag to run either algorithm. When neither are used, DDB (I2SB) sampling will be performed.

Also, make sure that --dataset-dir specified matches the paths specified in ./dataset/val_faster_imagefolder_10k_fn.txt and ./dataset/val_Faster_imagefolder_10k_label.txt. If not, modify the txt file.

If you find this work interesting, please consider citing

@article{chung2023direct,
  title={Direct Diffusion Bridge using Data Consistency for Inverse Problems},
  author={Chung, Hyungjin and Kim, Jeongsol and Ye, Jong Chul},
  journal={Advances in Neural Information Processing Systems},
  year={2023}
}