Prediction example for an instrument not used in training: ZTF (Bellm et al. 2019). Left: a science image exposure. Top right: mask from the ZTF pipeline. Bottom right: flagging by MAXIMASK; the trail is correctly recovered. Also, MAXIMASK CNN is able to correctly flag pixels where the trail overlaps sources whereas in the ZTF pipeline, all pixels (i.e., pixels belonging only to the trail, pixels belonging only to sources, and pixels belonging to both the trail and sources) are flagged as both trail and source.