Choice of channel size progression

[theodore-evans]

Regarding the progression of channel size through the synthesising network, you have chosen a non-geometric progression of 256 > 512 > 256 > 128 > 64 > 32 > 3 channels.

Aside from the practical implications of outputting to 7 x 7 x 1024 channels on the second dense layer, is there any other rationale to this choice? Or conversely, is there any downside to this decision?