Unlocking Temporal Generalization in Hamiltonian Video Dynamics Models
Hamiltonian Generative Networks (HGN) aim to predict video dynamics using a continuous-time energy function for better temporal generalization. However, these models fail when predicting at step sizes outside their training range due to issues like energy magnitude growth and integrator errors. The authors fix this by applying spectral normalization and substepping the integrator during inference, allowing stable predictions at finer or coarser time resolutions.