Reliable multi-view forecasting is hard because cameras must agree on the same evolving 3D scene. A new arXiv paper proposes Video-Point Model (VPM), a world model that takes multi-view video histories as input and predicts future frames by learning point trajectories in 4D space.
The approach frames forecasting as a trajectory prediction problem: instead of generating each view independently, VPM tracks points through space and time, which helps maintain consistency across cameras. The abstract notes that the model conditions on past multi-view video and then forecasts the next states.
Because the abstract is truncated, the full architecture and evaluation details are not available in this source. Still, the core idea is clear: representing the world as moving points rather than separate 2D frames may offer a stronger basis for multi-view prediction.