MAGNET: an all-in-one foundation model for cross-modal and cross-dimensional microscopic image restoration

The application of emerging Foundation models in current microscopy has remained limited due to task-specific model designs, strong modality dependence and insufficient robustness under complex degradations. Here, we introduce MAGNET (Microscopic All-in-one General fouNdation model for imagE resToration), the first end-to-end All-in-One foundation model designed for universal microscopic image restoration. MAGNET integrates multi-task, cross-modal and cross-dimensional (2D + 3D) image restoration capabilities within a single unified architecture, and it is specifically designed to handle composite degradations in practical imaging systems. The framework comprises three components: (i) a task-aware, prompt-guided feature enhancement module for adaptive learning across diverse degradations; (ii) an LIIF-based reconstruction module enabling continuous, resolution-adaptive restoration; and (iii) a dimension-compatible triple-plane projection module supporting both 2D and 3D data. Trained on large-scale datasets spanning 5 major microscopy modalities (structured illumination, confocal, light-sheet, wide-field and two-photon) and 8 representative restoration tasks, MAGNET achieves state-of-the-art performance across super-resolution, denoising, deblurring, background suppression, isotropic reconstruction, aberration correction, de-scattering and virtual staining. Beyond supervised training, MAGNET supports direct inference on unseen systems, efficient few-shot fine-tuning and self-supervised test-time optimization, enabling robust performance across diverse data and imaging conditions. MAGNET serves as a unified, generalizable, and scalable computational framework for microscopy, enabling more efficient and reliable downstream analyses in cell biology, neuroscience, and pathology.