Using a neural network the method seems to be that you end up with a probability for each possible outcome.
To predict the next frame in a monochrome movie of size 400x400 with 8 shades of gray, it seems like there seems to be: 8^(160000) possibilities.
On the other had if you just predicted the probability for each pixel individually you would end up with some kind of image which gets progressively blurred.
Perhaps what you want is to generate a few possibilities that are none-the-less quite sharp. In a similar way to weather prediction(?)
So how would you go about designing a neural network that takes reads a movie and tries to predict the next frame?