> The main drawback to this method is, like Gaussian splatting, they have to train an entire neural network for every puzzle.

They discuss how to avoid this in the section "Joint Compression via Weight Sharing Between Puzzles".