Adapting Text-to-Image Generation with Feature Difference Instruction for Generic Image Restoration

Low-light Enhancement Desnowing Deraining Dehazing Denoising Motion Deblurring Mixed Degradations Figure 1. Our image restoration results. The proposed DiffRes introduces cross-modal understanding by leveraging multimodal visionlanguage models through an efficient adapter, without modifying the pre-trained diffusion model. This enables DiffRes to effectively produce clean and sharp results over 10 image restoration tasks. For simplicity, we only show several representative tasks here.