Diffusion Degradation Oriented Model

Environment

Please clone this repo and run following command locally for install the environment:

conda create --name ddom
pip install -r requirements.txt

Datasets

Please put the training dataset into datasets/ folder and test dataset into exp/datasets/. The restoration results will be saved in exp/image_samples/. Datasets can be downloaded at link1 and link2.

Quick Start

We have prepared training and testing scripts for various image restoration tasks in the exp.sh file. For example, if we want to train DDOM for SR, we can run below command:

CUDA_VISIBLE_DEVICES=0 python train.py --A_type adapter --epochs 100 --nums_rb 20 --add_temb True --res_adap True \
--lr 1e-5 --ni --config celeba_hq.yml --path_y celeba_hq --eta 0.85 --batch_size 2 --init_x Apy \
--deg "sr_averagepooling" --deg_scale 4 --sigma_y 0. -i main_sr4 --train_size 2 \
--save save/main_sr4

where nums_rb is the number of layers in DO-Adapter, deg is degradation task, config is the name of the config file (see configs/ folder), path_y is the path of test images (please save in exp/datasets/).

This implementation is extended based on DDNM, thanks!