I have a dumb question about image GANs. It seems models with more parameters are able to better create readable text. Instead of adding more parameters, could you modify the cost function to try to convert the image to text and reward real characters, and reward even more real words?