g'day hi saintdev, I found some explainations which are technically wrong while in practice not so bad... just a second saintdev: mixed refs: it is not about mixing references within a frame (if there was no such mixing, it would still be only one ref) but for different partitions of a macroblock saintdev: direct prediction: temporal guesses back and forward motion from p-frame motion (of future p-frame); temporal guesses motion from motion of neighbouring blocks (of current b-frame) ugh again saintdev: direct prediction: temporal guesses back and forward motion from p-frame motion (of future p-frame); spatial guesses motion from motion of neighbouring blocks (of current b-frame) ^ better weightb: whenever both future and past motion compensation is used (bidi mode and direct mode), it blends the two images according to weights which depend on distance, rather than in 1:1 ratio subme: avc always ever uses quarterpixel motion compemnsation; subme is simply the search effort when refining motion to sub-pixel level (how many positions are searched) ok I think that's good enough, although sometimes I'd explain things differently ;)