I’v found that there are a lot of interlaced content that has this problem in source:
When top or bottom or both fields already have “blended” result (as on left field in example).
Its a good idea to make more training (or new model) that will learn to deinterlace it somehow.
Available models deinterlace it good (considering that in a lot of fps its not seen by eye), but not excellent. Frames contain lines where blended parts were.