A Survey on Autoregressive Video Generation

songhan_mit · x · 2026-07-15

This blog post surveys the evolution of autoregressive video generation, focusing on how video diffusion models are gradually shifting toward causal modeling pathways better suited for real-time, interactive, and long-duration generation.

It emphasizes that video generation is moving from "simply generating longer videos" to "endowing the generation process with controllable temporal causal structures," bringing it closer to a deployable foundational capability. The article also maps out the evolution of related research and discusses why these methods could become a crucial foundation for the next phase of video generation.

Original post →

More from Multimodal

Multimodal channel →