may be this?
not sure if we can speed up decoding in libavcoded - mplayer seems to seek much faster but I guess it can skip decoding whole frames for that and our pipeline not ready for this, because how you will deal with empty frames after such fast seeking? I mean you can hit any frame in arbitrary place, so decoding it due to b and p frames might take a while, so randomly-placed edit probably seeks/decodes a lot just for avoiding empty frm problem (while for playback there is toggle about playing every frame or not. But I think it only drops at output?)