Dense, Sparse, And Weak Labels For Visual Understanding And Generation