S2PD: Serial-to-Parallel Diffusion for Physically and Logically Consistent Video Generation

Paper Overview

Field: Video Generation Authors: Jeffrey Hu, Daniel Olmeda Reino, Ayush Tewari Published: 2026-10-05 arXiv: 2610.06847

Summary

Bidirectional video diffusion models denoise an entire video in parallel, but they persistently violate physical laws and simple symbolic rules even when trained on near-unlimited in-distribution data from procedural generators.

The authors propose Serial-to-Parallel Diffusion (S2PD), which performs autoregressive diffusion at high noise levels before switching to parallel diffusion at low noise levels:

• The autoregressive phase provides the serial computation required to coordinate interdependent events and produce valid state transitions.
• The parallel phase jointly refines the entire video and reduces sampling time relative to fully serial generation.

Implementation

S2PD is realized through two architectures:

1. A pixel-space diffusion transformer trained from scratch. 2. A pretrained video model adapted via LoRA fine-tuning and causal attention.

Results

Evaluated on games, physics simulations, and real-world videos, S2PD:

• Follows rules more reliably than matched bidirectional baselines.
• Produces videos with stronger temporal stability and higher sampling efficiency than other serial approaches.

Original Abstract

> S2PD performs autoregressive diffusion at high noise before switching to parallel diffusion at low noise. The autoregressive phase coordinates interdependent events for valid state transitions, while the parallel phase jointly refines and reduces sampling time.