Coarse-to-Fine Mechanisms Mitigate Diffusion Limitations on Image Restoration (C2F-DFT)

Liyan Wang, Qinyu Yang, Cong Wang, Wei Wang, and Zhixun Su*

[2024-08-11] Our paper is accepted to Computer Vision and Image Understanding (CVIU).

Coarse Training Pipeline of Diffusion Transformer Model (DFT)

Fine Training Pipeline and Sampling Phase

Requirements

CUDA 10.1 (or later)
Python 3.9 (or later)
Pytorch 1.8.1 (or later)
Torchvision 0.19
OpenCV 4.7.0
tensorboard, skimage, scipy, lmdb, tqdm, yaml, einops, natsort

Training and Evaluation

Training and testing instructions and visualization results for Image Deraining, Image Deblurring, and Real Image Denoising are provied in the links below.

Task	Training	Testing	C2F-DFT's Visual Results
Image Deraining	Link	Link	Download
Image Deblurring	Link	Link	Download
Real Image Denoising	Link	Link	Download

Experimental Results

Image Deraining (click to expand)

Image Deblurring (click to expand)

Real Image Denoising (click to expand)

Citation

@article{WANG2024104118,
title = {Coarse-to-fine mechanisms mitigate diffusion limitations on image restoration},
journal = {Computer Vision and Image Understanding},
volume = {248},
pages = {104118},
year = {2024},
issn = {1077-3142},
}

Acknowledgments

This code is based on the BasicSR toolbox and Restormer, WeatherDiffusion.

wlydlut/C2F-DFT