Any suggestion on designing network for one-hot condition?

[gitlabspy]

I want to implement generation with label on some other dataset which is more complex and have higher resolution than mnist. So, if using convolutional layers in subnet for coupling layers, should I simply transform condition into a data-shape-like tensor (same height and width with data) then concatenate with the splitted data for transforming the other half, or some other ways?
Due to the limitation on computational resource, I don't have much time for exploring this. So, I want to consult you guys here. Thx for the great work! 😄