and what's the size of the neural network you have to ship for this to work? has anyone done the math on the break even point compared to other compression tool

and what's the size of the neural network you have to ship for this to work? has anyone done the math on the break even point compared to other compression tools? e: actually a better metric would be how much does it compress compared to doing the resolution increase with just lanczos in place of the neural net and keeping the Delta part intact