Wild-IR not working as expected

[cyy2427]

Hello. I just checked out the updated implementation of Wild-IR and test the output by running app.py in universal-image-restoration/config/wild-ir, and I have made sure that I placed the pretrained models in right paths and no errors were thrown during the gradio app was running. However, the output images are all in the same size as the input images and the quality was remained unchanged. The related opt file inference.yaml and code in models directory only supports denoising model rather than SR model.

According to all clues mentioned above, I wonder if the Wild-IR implementation in this repo is still not working as intended (not the same as the impressive results shown in the paper). I would appreciate more helpful information from you.

[Algolzw]

Hi! The current Wild-IR model only supports normal image restoration (denoising, deblurring, dejpeg...) where the LQ and HQ images have the same size. If you want to perform super-resolution you can try to resize the LQ image before feeding it to the model.