[MPlayer-dev-eng] [RFC] preliminary x264 encoding help text

[D Richard Felker III]

On Sat, Apr 09, 2005 at 06:00:10PM -0500, Jeff Clagg wrote:
> And on a semi-related note, can anyone give a convincing justification for
> why the man page still states that increasing frameref is more effective
> on anime vs. live material? I fear this comment may have its origins in
> some poorly-considered and half-assed tests I did way back last fall.

I can explain the theoretical reason. Often anime, or at least bad
anime, uses the technique of flipping through a short sequence of
frames (often just 2 for a flashing effect, but sometimes 5-10 for
repeated clouds/waves/scenery going by, etc.) to save money by not
drawing as many frames. Actually it's more common in American kiddie
cartoons than anime, but oh well. Anyway, if you have the same frame
you're trying to encode as a backreference from several frames back,
you hardly use any space encoding it. Otherwise it may use a lot,
especially if it's something like flashing where almost all the change
is texture instead of motion.

There's also the issue of large objects quickly moving across a
totally still background, which happens quite often in animation, both
good and bad. Multiple backreferences will help you here too.

My personal feeling: the whole thing is stupid. It does save a lot of
space under certain strange circumstances, but keeping so many
reference frames destroys cache coherency for decoding and makes it so
you need a cpu with huge cache or insanely fast memory io to have any
chance of decoding the movie. IMO there are much better ways of
improving compression (like using snow instead of h264.. :)

Rich