Context. The far-side activity of the Sun can be inferred by interpreting the near-side wave field using local helioseismic techniques. However, detections are limited to strongly active regions because signal-to-noise ratio of the data is low. Recently, we developed the FarNet and FarNet-II neural networks to improve the identification of active regions on far-side seismic maps. Aims. We aim to use FarNet-II to leverage seismic data to infer far-side magnetograms, including the magnetic field strength and polarity. Methods. We used FarNet-II to produce sequences of 11 consecutive binned magnetograms with a 12-hour cadence of a central section of the far side, where each pixel was assigned to one of nine possible classes that define its magnetic field and polarity. The inputs to the network are sequences of phase-shift maps of the same regions, computed using helioseismic holography. We trained the network using a cross-validation approach to estimate its reliability. The targets for the training and the cross-validation were obtained from near-side Helioseismic and Magnetic Imager magnetograms, taken half a rotation later than the seismic data. The metric we used for the evaluation is the volumetric Dice, a newly defined metric that measures the overlap between the outputs and the targets. The results were compared with Solar Orbiter data from a period with far-side coverage between May 2022 and September 2022. Results. FarNet-II achieves an average volumetric Dice of 0.249, showing a good visual superposition between the targets and outputs of the network. The comparisons of the outputs and the Solar Orbiter magnetograms are also similar. Conclusions. FarNet-II can correctly predict the level of activity and the polarity of far-side regions using near-side seismic data. This capability can be leveraged in space-weather forecasting.