VisionICLab/SwinFuSR ? reverse-engineered prompt

Build me a project that does RGB guided thermal image super resolution like the SwinFuSR paper. I want to be able to train a model on paired thermal low resolution images, RGB guide images, and ground truth thermal images, then test it on a validation or test set and save the results and metrics.

Please include simple command line scripts for training and testing, with the main settings coming from json config files, like dataset paths, scale factor, batch size, learning rate, and whether ground truth is available. It should support x8 and x16 upscaling, loading pretrained weights for fine tuning, and a mode where the RGB guide image is sometimes missing during training.

Make the setup and running instructions easy to follow, and keep the code organized so I can switch between small and larger model settings from the config files. If anything is unclear, look up the current docs online if you need to.