Method
Models aren’t short on visual capability. Given a detailed enough prompt, they render individual principles with striking fidelity, but the failure shows up specifically when several principles must unfold together, along two distinct axes.
-
1
Principle Omission.
Secondary physics gets dropped entirely: the apple falls realistically, but the water it lands in stays undisturbed. -
2
Temporal Shortcutting.
Unfolding dynamics collapse into a single visual snapshot: a splash rendered as a static spraying effect instead of progressing through impact → crown formation → submersion.
HiPhy’s dual-level design addresses both directly: per-principle supervision targets Principle Omission, while fine-grained sub-stage reward design addresses Temporal Shortcutting. At the foundational level, HiPhy explicitly optimizes the chronological progression of each physical event by breaking it into smaller stages; at the same time, it aligns these stages with global video coherence and semantic alignment objectives, so multi-principle interactions compose naturally.
HiPhy explicitly decomposes complex physical events (e.g., “an apple falling into water”) into temporally ordered sub-stages. This lets the reward recursively score the chronological progression, completeness, and alignment of the underlying physics.