OK this might be an edge case but I feel like one of the most difficult thing with video restoration is dealing with blended frames. Some AviSynth filters like SRestore do a great job but not perfect, and cannot do anything if there are too many blended frames. Training a model for this purpose shouldn’t be too difficult, it would mean training a blend-deinterlaced footage against the same correctly deinterlaced footage. It would be a tremendous help in restorating older VHS content, and something that has not been done yet.
It is so difficult to get rid of blended frames. This suggestion needs more attention.