Legacy animation optimization

So we are fairly close to where we want to be but any little extra would help, looking for techniques I might have missed.

Other approaches entirely like gpu instanced or mesh per frame I'm aware of but those are not suitable here.

The biggest gains I've gotten is from simplifying the animation clips themselves. Curve sampling is the largest consumer of main thread cpu time with average sized character models. Ie it uses more cpu then the skinning calculations. Mesh LOD has some impact and it of course impacts rendering which animation clip eval doesn't. But purely based on main thread usage having LOD's for animation clips seems to be the biggest winner.

Any other tricks that might help?