In the left-hand column, we show the ratio cs/cs+fg for the 0.1 and 1 per cent LOS varying foreground models (top and bottom, respectively). In the right-hand column, we show the same ratio but with the foreground fitting errors as a result of performing GMCA on the new foreground cubes with LOFAR noise. We see that foreground avoidance is affected badly by introducing a 1 per cent LOS variation, while GMCA is still able to make a good recovery, under the assumption of common resolution channels. Linestyles are as in Fig. 4.