One of the numbers every guide says to change in encoding settings is subq/subme, but there has been little objective study of just how beneficial this can be. So I put my 'ponymath' script to use and ran some tests myself.

Though subme is specified as a number, it isn't actually interpreted as one - instead it sets flags enabling additional subpixel motion estimation on an increasingly large set of frames. The default is 7, and the maximum 11, which enables the most precise for of subpixel refinement on all frames. There's a lot of debate on forums over how beneficial higher subq is, if any benefit at all, so here is some hard data. Four test clips are used: A high-motion animation, ten minutes (Actually a clip from the film 'Iron Man: The Rise of Technovore'), ten minutes live-action from a DVD of reasonable but not great video quality, ten minutes of live-action HDTV and the 'Noogenesis' music video that I happened to have to hand at the time.


Test one: Everything but subq default.

720p high-motion animation, ten minutes. Live-action video from DVD, ten minutes. Live-action 720p high-quality, ten minutes. 'Noogenesis' music video, 720p

Across a variety of inputs a consistant pattern can be observed: Moving from the default 7 up to 8 produces a small improvement in SSIM, all else remaining equal, but going up any higher produces negligable further improvement if any at all. There is a very good reason for this: According to the wiki, subq settings of 10 or above don't work without setting trellis=2 and aq-mode=1 or 2. Neither of which is a default. Trellis=2 is also unavailable at the more constraining profile levels. Of more interest is that there is little difference between subq=8 and 9.


Test two: Enabling trellis=2:aq-mode=2

720p high-motion animation. Live action video from DVD. Live-action 720p high-quality. 'Noogenesis' music video, 720p.

The jump from 8 to 9 isn't much, but the jump from 9 to 10 is. An increase of 0.0005 is certainly worthwhile. Going from 10 to 11, though, seems to provide minimal improvement - while the SSIM score is increased, it isn't increased significantly. On some videos the SSIM for subq=11 was actually lower than for subq=10, if only by less than 0.0001.


Test three: As above, but with me=tesa

720p high-motion animation. Live-action video from DVD. Live-action 720p high-quality. 'Noogenesis' music video.

There are many settings that could affect how changing subq performs - in video encodeing, all things are interdependant. The most obvious of these is the motion estimation used. If encoding for maximum quality and not careing about time, the 'tesa' setting is highly adviseable, so the effects of subq changes in conjunction with me=tesa should also be tested. There is actually very little change. Even the advantage of me=tesa isn't evident, as it only becomes apparent at higher merange settings.


In conclusion, for all of the four videos I tested representing animation, HDTV and DVD, the optimal value to use for subq would be subq=10. It may be reasonably assumed that this is the case for almost all videos.