In the last decade in the petroleum industry, geostatistical filtering solutions based on Factorial Kriging technique have been developed and applied to seismic data sets in various operational contexts. These solutions commonly assume stationarity for the underlying random function, which limits their efficiency as soon as the target area becomes large or involves complex structural patterns.In this paper we introduce M-Factorial Kriging models, which allow to account for non-stationary effects that are encountered within seismic data sets. In the framework of noise attenuation issues, sources of non-stationarity relate for example to signal absorption, geological structuration, spatial variations of signal-to-noise ratio or varying geometrical features of noise.M-Factorial Kriging models ensure a better efficiency of the resulting geostatistical filtering process. As a consequence, signal and noise are better separated. This is illustrated by applying M-Factorial Kriging to a noisy PSTM amplitude section.