ProRes is a novel approach to image restoration that leverages degradation-aware visual prompts to control the restoration process. The model is designed to handle a variety of image degradation types, such as noise and blur, and allows for weighted combinations for customized image restoration. This approach provides a universal solution to image restoration tasks, overcoming the limitations of task-specific methods that often suffer from uncontrollable and undesired predictions.
The ProRes model uses a vanilla Vision Transformer (ViT) without any task-specific designs, making it adaptable to a wide range of image restoration tasks. Furthermore, the model can easily adapt to new tasks through efficient prompt tuning with only a few images. This makes ProRes a versatile tool for image restoration, capable of achieving competitive performance compared to task-specific methods. The model demonstrates its ability for controllable restoration and adaptation for new tasks, making it a promising solution for universal image restoration.
Publication date: June 26, 2023
Project Page: https://github.com/leonmakise/ProRes
Paper: https://arxiv.org/pdf/2306.13653.pdf
