Learning Skeletal Articulations with Neural Blend Shapes

So here the idea is to predict which points will be more affected by a deformation, given a 3D mesh model and the corresponding skeleton, right?

The output is probably something similar to a normal map. Can the deformation be handled efficiently on the GPU using a geometry shader?

Hopefully in a few years we won't have aberrations like parts of a body colliding with another. This is also a frequent issue with boats in game.

(haven't read the article/paper yet but...) I wonder how this works compared to the games industry standard using inverse kinematics?

edit: looks like this paper is dedicated to the motion capture side of animations, rather than comparing/blending with procedural generation techniques.