Researchers have developed the MIRA-Net model, a U-Net-based multimodal instruction-driven recovery architecture that reduces ionizing radiation dose in PET/CT, CT, PET, and MRI without loss of image quality. The adaptive guidance module estimates modality and degradation from the input, generates instructions to modulate processing, and selects appropriate paths in a single model. The model was tested on CT denoising, PET synthesis and super-resolution MRI tasks. When training for individual tasks, it achieved performance comparable to or better than specialized models. As a unified model across modalities, it maintained the same performance without decline. It demonstrated robust generalization with consistent metrics on a local clinical dataset. In a double-blind study with radiologists, MIRA-Net images were more often rated as diagnostic, with higher scores for anatomical clarity, lesion conspicuity, and noise control. MIRA-Net thus enables multimodal image recovery without sacrificing diagnostic quality.