Ensuring reproducibility of Magnetic Resonance (MR) images from different scanners is crucial in multicenter studies, as scanner-induced variability is known to impact the results significantly. To address this problem, we introduce a novel unsupervised deep learning approach aimed at achieving 3 primary objectives/advantages: (1) create a scanner-free space that enables the uniform transfer of images across different scanners in a denoised setting, (2) impart the appearance of a specific training scanner to images, transferring its unique characteristics, (3) avoid time-consuming preprocessing of MR images. The proposed methodology is based on disentangling image information into two distinct spaces: one encoding the scanner-specific information and one capturing the anatomical/biological structure of the image. We trained our model on two open-source datasets (ADNI and PPMI) of T1-weighted brain MR images of normal control patients. We tested it on a real-world clinical dataset from the Italian Neuroimaging Network, comparing its performance with a state-of-the-art model. The results show the superiority of the proposed model in harmonizing images for clinical research, demonstrating its effectiveness in achieving consistent and reproducible harmonization of the MR images across (unseen) scanning environments. Code is available at https://github.com/luca2245/DISARM_Harmonization .