Task-aware all-in-one guided image super-resolution

Guided image super-resolution (GISR) aims to enhance the resolution of a low-resolution (LR) target image by leveraging complementary information from a high-resolution (HR) guidance image. However, due to the substantial modality diversity among GISR subtasks, most existing methods are tailored to individual subtasks, which significantly limits their generalizability and practical scalability. To address this limitation, we propose MAG-Net, the first all-in-one GISR framework capable of handling multiple GISR subtasks within a single unified model. MAG-Net is built upon a shared encoder-decoder backbone and integrates two key modules to handle heterogeneous input modalities and diverse task objectives. The Multi-modal Prompt Generation Module (MPGM) dynamically generates task-aware prompts by jointly encoding the input image pair and pre-defined textual task descriptions. These prompts serve as soft instructions, effectively capturing both visual features and textural cues, thus enabling the model to adaptively distinguish and respond to different GISR subtasks. The Multi-Guidance Routing Module (MGRM) is then designed to mitigate task interference and enhance feature specialization. This module leverages a Mixture-of-Experts (MoE) strategy to adaptively route intermediate features through task-relevant expert branches, guided by the task-aware prompt and the characteristics of the guidance image. Extensive experiments across various GISR subtasks demonstrate that MAG-Net achieves state-of-the-art performance in both all-in-one and one-by-one training settings. Code and pre-trained models will be released upon paper acceptance.