It has been done before. Problem is that it's not a standard process with known rules. Neural network prediction results may vary a lot based on how closely your test data fits the training set, and how well the network's been trained. Gans are particularly difficult to train, since the target isn't well defined (in order to achieve innovation beyond the base set).
Furthermore, outside the academy almost nobody's releasing their weights & data, as it's more commercially viable to sell predictions as a service, plus not everyone has the means & knowledge to set up a prediction environment. This stalls technological improvement as a whole.
Long story short, yes. Look it up online, ymmv.
Example (haven't tested):
https :// hig . dev/apps/ai-photo-restore/