1) The star trails can be any length - a function of the comet's arc second per minute drift and the length of the exposure
2) The direction of the star trail can be any angle - a function of the camera orientation and the comet orbit
None of these should be problematic.
The first one should not be a problem at all given that you have reasonable number of different star trail length in your training data.
The second one is solved using image augmentation - you just randomly rotate input images during training.
I would be more worried about 'structure' of star trails since there could be different sequences of filters, lengths of exposures, various gaps, etc. etc. However if all the images are like the example you attached everything should be fine.
Edited by nekitmm, 19 September 2019 - 01:42 PM.