This repository contains the implementation and pre-trained weights for my master's thesis Exploring the usage of diffusion models for image super-resolution. It contains implementation and trained weights for the ResShift method (Yue, Wang and Loy, 2023) using both UNet and Difusion Transformer architectures.
First follow torch installation instructions here to correctly install torch according to your cuda version. Used versions are included in requirements_torch.txt. Then, install the rest of the requirements.
pip install -r requirements.txt
First download the checkpoints from the releases tab
Place the LR images in a folder, lets call it LR.
python inference.py checkpoint.pth [unet | transformer] --bs 8 -i LR visible -o results --cuda
Multiple configuration files are provided in the config folder, adjust accordingly.
You will need to make 3 files containing the filenames of the train,valid and tests splits relative to the data_dir in the config file. A simple way of making them asuming you have them split in 3 folders inside data_dir is by using find train/ -type f > train_list.txt.
python diffusion_resshift_pixart.py fit --config my_config.yaml
or
python diffusion_resshift_unet.py fit --config my_config.yaml
In order to finetune from a provided checkpoint you can enter the path to checkpoint file in the ckpt_path field in the config file.
Distributed under the MIT License. See LICENSE for more information.
