daclip-uir

Why we included this project

DA-CLIP tackles the messy reality that most image restoration tools handle one problem at a time. Instead of training separate networks for blur, haze, noise, rain, snow, low light, JPEG artifacts, and shadows, it uses a single model that first figures out what is wrong with an image, then runs a diffusion-based pipeline to fix it. The repo includes pretrained weights, a Gradio app for testing your own photos, and Colab and Hugging Face demos, so you can see how it behaves on real images before investing in training. The ICLR 2024 paper and the follow-up wild-IR model, aimed at real-world mixed degradations, make it a solid baseline and a useful starting point for extending restoration to new distortion types.