There is something with the AI needing to have a ‘sense’ for the world the scene exists in so longer videos can be created that are coherent. Currently we’ve only seen long videos that have no consistency and jump around a lot like an acid trip.
A good path forward is to fuse these image-element compositing tools with some of the 3d scene inference ones. So you start out with 'giant fish riding in a golf cart, using its tail to steer', then give that as ground truth to a modeling tool that figures out a fish and a wheeled vehicle well enough to reference some canonical examples with detailed shape and structure, the idea of weight etc. Then you build a new model with those and do some physics simulation (while maintaining a localized constraint of unreality that allows the fish to somehow stay in the seat of the golf cart).
