This repository contains the implementation and pre-trained weights for the paper Exploring the usage of diffusion models for thermal image super-resolution: a generic, uncertainty-aware approach for guided and non-guided schemes.
First download the checkpoints from the releases tab
Place the LR images in a folder, lets call it LR (in the case of guided super-resolution, also place all the visible image in another folder, lets call it visible, be sure that matching images have the same names).
python inference.py checkpoint.pth --bs 8 -i LR visible -o rough_results
Then, since this a two-model approach use the refiner model to further enhance the results
python inference.py refiner_checkpoint.pth --refiner --bs 8 -i LR visible rough_results -o results
If you want to train the model in another dataset you may start by finetuning the one pretrained in imagenet (available in the releases tab). Multiple configuration files are provided in the config folder, adjust accordingly and run using
python diffusion_resshift.py fit --config my_config.yaml
If you instead want to train/finetune the refiner model, you may need to generate rough_predictions from the diffusion model using the code described in the inference section and adjust its configuration file accordingly.
python unet_refiner.py fit --config my_refiner_config.yaml
Distributed under the MIT License. See LICENSE for more information.