Histogram of the jet offsets for a deep learning model trained with noiseless (orange), one trained with noisy input data (blue) and clean images generated with wsclean (green). The offset range is capped from −25° to 25° for visibility reasons. Only small differences are present between the two distributions obtained from the deep learning models. Reconstructed jet orientations fit the true values well, which is supported by the means and standard deviations. Small pixel offsets in the cleaning process can already cause large offsets of the jet angle. This is a possible explanation for the larger deviations occurring in the case of wsclean.