# Implicit Diffusion Models for Continuous Super-Resolution
This repository is an offical implementation of the paper "Implicit Diffusion Models for Continuous Super-Resolution" from CVPR 2023.
This repository is still under development.
## Updates:
The pre-trained model for 8X face continuous SR has been updated in [link](https://drive.google.com/drive/folders/1VISy9fVWa9iOSr6F4oVtKVTOViWuKohQ?usp=drive_link).
## Environment configuration
The codes are based on python3.7+, CUDA version 11.0+. The specific configuration steps are as follows:
1. Create conda environment
```shell
conda create -n idm python=3.7.10
conda activate idm
```
2. Install pytorch
```shell
conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3
```
3. Installation profile
```shell
pip install -r requirements.txt
python setup.py develop
```
## Data preparation
Firstly, download the datasets used.
- [FFHQ](https://github.com/NVlabs/ffhq-dataset) | [CelebaHQ](https://www.kaggle.com/badasstechie/celebahq-resized-256x256)
- [DIV2K](https://data.vision.ee.ethz.ch/cvl/DIV2K/) ｜ [Flick2K](http://cv.snu.ac.kr/research/EDSR/Flickr2K.tar)
Then, resize to get LR_IMGS and HR_IMGS.
```
python data/prepare_data.py --path [dataset root] --out [output root] --size 16,128 -l
```
## Pre-trained checkpoints
The pre-trained checkpoints can be found at the following: [link](https://drive.google.com/drive/folders/1VISy9fVWa9iOSr6F4oVtKVTOViWuKohQ?usp=drive_link).
## Training and Validation
Run the following command for the training and validation:
```shell
sh run.sh
```
Add the command "-use_ddim" to implement DDIM sampling.
## Acknowledgements
This code is mainly built on [SR3](https://github.com/Janspiry/Image-Super-Resolution-via-Iterative-Refinement), [stylegan2-ada-pytorch](https://github.com/NVlabs/stylegan2-ada-pytorch), and [LIIF](https://github.com/yinboc/liif).