cINN	Output size
Learnable downsampling	$4\times 14\times 14$
Level 1 conditional section	$4\times 14\times 14$
Learnable downsampling	$16\times 7\times 7$
Level 2 conditional section	$16\times 7\times 7$
Flatten	784
Split: 656 to output	128
Level 3 dense-conditional section	128