Right it's just creating a 2D basis space and placing the various labels inside that space. (And the previous paper was about compositing elements of multiple

Right it's just creating a 2D basis space and placing the various labels inside that space. (And the previous paper was about compositing elements of multiple dataset images into a single training example - likely overfitting).