IRCDG: Adaptive all-in-one image restoration via component-divided guidance

Image restoration is a fundamental task that aims to recover high-quality, clean images from their degraded versions. Although deep learning-based approaches have achieved remarkable progress, they often struggle to generalize across diverse degradation types and levels, limiting their practical utility. Moreover, we observe that different regions within an image present varying levels of restoration difficulty, which can be more efficiently addressed using networks of different capacities. To tackle these challenges, we propose IRCDG, an adaptive all-in-one image restoration framework that integrates component-divided guidance. Our method first decomposes the image into distinct components and adaptively assigns appropriate network depths to process different regions based on their complexity. On top of this, we introduce frequency-based prompts to encode degradation-specific information, which dynamically guides the restoration process. This design not only enhances generalization across various degradation types and levels but also reduces computational overhead while improving the model's representational capacity across different depths. Extensive experiments across multiple image restoration tasks demonstrate that our method achieves state-of-the-art performance with superior efficiency.