The Seriality Gap in Video Diffusion Models

CatAstro_Piyush · x · 2026-07-17

This thread discusses whether video diffusion models can truly predict long-chain causal events, focusing on a phenomenon known as the seriality gap: standard video diffusion models degrade significantly as dependency chains grow longer, and the author argues that simply adding more compute won't fix this.

The thread also proposes a stronger theoretical claim: diffusion models might not be "true sequence models" like RNNs, functioning more like a fixed-depth network. If true, they shouldn't be expected to reliably handle video tasks requiring long sequences of step-by-step reasoning.

Related event: Video Diffusion Models Expose Seriality Gap(2 posts)→

Original post →

More from Research

Research channel →