I admit I don't know much about that field. But I don't see why video generation methods not follow similar rate of improvement as image generation. There's almost infinite amount of training data (e.g. YouTube), and the hardware keeps improving. I can easily imagine - in a couple of years - typing "generate a 30 seconds video in style X where characters are tossing fireballs at each other (optional description of the characters and the scene)" and getting professional looking results. And in another couple of years generating entire movies.

It’s a question of internal consistency. The video you linked skipped that by starting with actual actors, but “generate a 30 second video” means no shortcuts.

There’s a common issue in AI generated still art for dynamic shots like someone dunking a basketball to have extra arms. This doesn’t show up in pictures of people standing still because the training set is completely dominated by people standing with their arms to the sides. But in dynamic shots either seems fine and it doesn’t understand the underlying limitations the way a human does.

I don’t think it quite needs AGI to “make a 30 second clip of a fireball fight” but it’s much much harder than a still image of a “wizard holding a fireball.”