In materials science, microstructure and the associated extrinsic properties are essential for engineering advanced structural and functional materials, but its robust reconstruction and generation are still challenging. In this work, utilizing the stable diffusion (SD) models and ControlNet, we implement and apply an algorithm to reconstruct and generate microstructures including both the phase and grain orientation information, encompassing additionally microstructure interpolation and inpainting. It is demonstrated that our image-based approach can be applied to replicate and analyze complex microstructure features with exceptional statistical and morphological fidelity, backed by a robust training on an extensive dataset of 576, 000 synthetic microstructures. Therefore, with the ensured superior accuracy, efficiency, and versatility over conventional methodologies, our approach highlights its generative capabilities in exploring previously unexplored microstructures, paving the way for the future data-driven development of advanced materials with tailored properties.