The Seriality Gap in Video Diffusion Models
This presentation examines a fundamental computational limitation in bidirectional video diffusion models: their inability to simulate deterministic physical processes that require serial propagation of dependent events. Using hard-sphere collision dynamics as a controlled diagnostic task, the paper reveals a persistent performance gap—the seriality gap—between bidirectional diffusion and more explicitly serial inference paradigms like autoregressive generation, and demonstrates that this gap cannot be closed by simply adding denoising steps or architectural width.Script
Video diffusion models excel at generating realistic footage, but can they predict what happens when one billiard ball strikes another, which strikes a third, which strikes a fourth? The answer reveals a fundamental computational gap.
The researchers designed a minimal test using hard-sphere dynamics in a box. With multiple balls, each collision changes the velocities that govern the next collision, creating a causally linked chain that must be resolved step by step.
Bidirectional diffusion models degrade sharply as the chain of dependent events lengthens, even though they spend more compute on longer videos. Autoregressive models, which generate frame by frame, maintain nearly constant accuracy. This performance gap persists across all tested configurations.
Adding more denoising steps or making the model wider does nothing to close the seriality gap. The critical factor is backbone depth, which provides the serial computational structure needed to propagate state through collision chains, mirroring the physical causality itself.
The failures are not subtle. Bidirectional generations exhibit state-consistency breakdowns like balls splitting, merging, swapping colors, or tunneling through walls. These violations occur precisely where serial propagation is essential, and human evaluators overwhelmingly prefer the physically plausible autoregressive outputs.
The theoretical result is stark: bidirectional diffusion, with fixed-depth backbones, operates in a parallel complexity class fundamentally mismatched to problems requiring inherently serial computation. For video models to simulate chains of dependent events at scale, they must provision serial depth in either architecture or inference structure. Visit EmergentMind.com to explore the full paper and create your own research video.
