Come to think of this, doesn't manga represent a perfect dataset? Why not train a model on original/translated page pairs? This way, you achieve a full E2E pip

Come to think of this, doesn't manga represent a perfect dataset? Why not train a model on original/translated page pairs? This way, you achieve a full E2E pipeline where you don't have to worry about cleaning, joining pages, translation, editing. Or is this too computationally prohibitive at the moment?