reenactment error when drive frames > target frames

[gowdygamble]

Incredible work and thank you all for the great repo! I'm running into a bug that I wanted to flag:

when I run reenactment with more driving frames than target frames, I break inside _create_flame_param_batch (nha_optimizer) when I try to get expression (and other) parameters from indices that don't exist in the target frame set (line 412 for me but I've probably added a bunch of debugging junk).

I've tried working it through - as best as I can tell, we do the following (lines 1865-1876 for me):

make flame_params_offsets based on the incoming driver batch
overwrite shape with the default neutral (base) target shape
then for all the other keys,
we get the delta between the current drive frame and the drive neutral
and then overwrite the corresponding flame_params_offset with the neutral target + the delta we just obtained

we then use the new flame_params_offset to make an image as we would normally.
So its not obvious to me why the original flame_params_offsets depend on the driving frame indices. They're all about to get overwritten anyways right?

I'm probably missing something obvious, just wanted to know if other people had hit this. I'm trying some klugey fixes, interested in hearing best practice.

Thanks again!