they're doing some shenanigans that amount to training a neural network except they're not really trying to create a network they're just using the numerical optimization tools to perform automatic level of detail cleanup. the process seems to involve taking random render samples of lighting and the model and then moving parts of the smaller mesh around with the goal of being visually identical to the original input and perceptively enforces this.

also seems that it will consider the normal mapping during this process meaning it will evolve a detail map that pairs with the smoller mesh to make it look right.

kinda looks like it would make that Instant Meshes project very very useful.

cc @freemo @purple @shadowferret