MR Fingerprinting is an emerging attractive candidate for multi-contrast imaging since it quickly generates reliable tissue parameter maps. However, contrast-weighted images generated from parameter maps often exhibit artifacts due to model and acquisition imperfections. Instead of direct modeling, we propose a supervised method to learn the mapping from MRF data directly to synthesized contrast-weighted images, i.e., direct contrast synthesis (DCS). In-vivo experiments on both volunteers and patients show substantial improvements of our proposed method over previous DCS method and methods that derive synthetic images from parameter maps.