How do AI video generation models work?

Temporal consistency is the part that fascinates me most. Image diffusion models already struggle with spatial coherence in complex scenes, but video adds the time dimension where even small inconsistencies between frames become immediately obvious to human perception. The autoencoder approach for computational efficiency is clever - compressing video into a latent space before running diffusion saves massive compute, but it also means the quality ceiling is partly determined by how good your encoder-decoder pair is. Curious whether the next big leap comes from better architectures or from training on higher-quality curated datasets. Right now it feels like we're in the 'scaling the data' phase similar to where LLMs were two years ago.