Diffraction images can be used for modeling reservoir heterogeneities at or below the seismic wavelength scale. However, the extraction of diffractions is challenging because their amplitude is weaker than that of overlapping reflections. Recently, deep learning (DL) approaches have been used as a powerful tool for diffraction extraction. Most DL approaches use a classification algorithm that classifies pixels in the seismic data as diffraction, reflection, noise, or diffraction with reflection, and takes whole values for the classified diffraction pixels. Thus, these DL methods cannot extract diffraction energy from pixels for which diffractions are masked by reflections. We proposed a DL-based diffraction extraction method that preserves the amplitude and phase characteristics of diffractions. Through the systematic generation of a training dataset using synthetic modeling based on t-distributed stochastic neighbor embedding (t-SNE) analysis, this technique extracts not only faint diffractions, but also diffraction tails overlapped by strong reflection events. We also demonstrated that the DL model pre-trained with basic synthetic dataset can be applied to seismic field data through transfer learning. Because the diffractions extracted by our method preserve the amplitude and phase, they can be used for velocity model building and high-resolution diffraction imaging.