The consecutive numerical method is implemented for construction of a depth velocity model by full waveform inversion. The inverse dynamic seismic problem is reduced to finding the minimum point of the objective functional characterizing the mean square deviation of the recorded data from those calculated for the current model of the medium. A distinctive feature of the proposed approach is decomposition of the model space into two components: a smoothly varying propagator (macrovelocity model) and a rapidly spatially varying component called a reflector. The minimum point is calculated in sequence in these two subspaces. This paper reports the obtained data on numerical experiments related to reconstruction of the Marmoussi2 velocity model using a real frequency range and a source–receiver offset.