This is fairly specific to my use, the 10 batch may not be really well optimized.
First run is still running so this may not actually work. :) I'll update with changes if needed.
Update: 5 frames seemed a bit much. Also trying in a 100 batch (issue with doing all at once is file descriptors)
Update2: 100 was still a bit slow (loads all 100 images in ram, then does the math, then dumps images) Running at 25 images appears to have less churn than 10, but doesn't kill the box when doing the math. Also moving to a 2 tween, initial test render looks promising.
End result: http://www.youtube.com/watch?v=qkfAyRyeln4#!