It is interesting how much image quality gain is achieved by a single operation in gimp. The two images are just different by a single sharpen operation. In the process of decomposing images, it would seem that quality of presentation could be enhanced greatly by such simple techniques. In this case I am extracting the background as a very large quality texture that is not repeated for each frame, but is simply overlaid with the moving parts, which themselves are static 3D surfaces with rotation and translation.
The impetus to degrade visual quality is driven by bandwidth, but when information is transfered in a different way, the bandwidth is saved and the quality of product is also increased. When a great deal of processing is done once for wide distribution and not a distributed processing that is lossy, quality and speed are gained.
Existing tools can extract 3D from a scene and deliver it as a model with articulation and script. It isn't sensible to do such things for one copy sent to one person, but when information is available to millions, like YouTube, it would make sense to support a format that was more of a model system than a stream of lossy compressed data.
For my purposes it is to store a set of variables that are used by all video and simply modifying parts to store only "new" information and its association. I can use a text format that is just a set of 16 bit numbers that describe a sequence of words, as well as their context sense ( as well as pronunciation ), as opposed to their literal ascii.