INDIGO+: A Unified INN-Guided Probabilistic Diffusion Algorithm for Blind and Non-Blind Image Restoration (JSTSP 2024)

Di You | Pier Luigi Dragotti

Imperial College London

This repository is the official PyTorch implementation of the paper INDIGO+: A Unified INN-Guided Probabilistic Diffusion Algorithm for Blind and Non-Blind Image Restoration (JSTSP 2024)

Overview framework

Experimental Results

Dependencies and Installation

conda env create -f environment.yaml
conda activate indigo

Inference

Download the pretrained INN models and testing data from GoogleDrive. The pretrained models for Diffusion model and SwinIR will be downloaded automatically.

python inference.py \
-i [INPUT_DIR]  -o [RESULT_DIR]  --task restoration \
--eta 0.5 --aligned  --use_fp16 \
--config_indigo configs/sample/indigo_syn.yaml

python inference.py \
-i [INPUT_DIR]  -o [RESULT_DIR]  --task restoration \
--eta 0.5 --aligned  --use_fp16 \
--config_indigo configs/sample/indigo_real.yaml

Contact

If you have any questions, please feel free to contact: [email protected]

Citations

If our code helps your research or work, please consider citing our paper. The following are BibTeX references:

@ARTICLE{10670023,
  author={You, Di and Dragotti, Pier Luigi},
  journal={IEEE Journal of Selected Topics in Signal Processing}, 
  title={INDIGO+: A Unified INN-Guided Probabilistic Diffusion Algorithm for Blind and Non-Blind Image Restoration}, 
  year={2024},
  volume={},
  number={},
  pages={1-15},
  keywords={Degradation;Image restoration;Diffusion models;Noise;Training;Image reconstruction;Noise measurement;Blind image restoration;diffusion models;image restoration;invertible neural networks},
  doi={10.1109/JSTSP.2024.3454957}}