S2PD Combines Autoregressive and Parallel Diffusion to Generate Physically Consistent Video

Researchers have introduced Serial-to-Parallel Diffusion (S2PD), a hybrid video diffusion architecture that addresses a persistent failure mode of bidirectional diffusion models: violations of physical laws and logical rules even when trained on large, well-structured datasets. S2PD runs autoregressive denoising at high noise levels — where coordinating interdependent events matters most — then switches to parallel denoising at low noise levels to jointly refine the full video and recoup sampling speed.

The approach is validated across discrete rule-based environments (chess, Tetris, Rubik's Cube), continuous physical simulations (double pendulum, colliding balls), and real video datasets. Two implementations are presented: a pixel-space diffusion transformer trained from scratch, and a pretrained video model adapted via LoRA fine-tuning with causal attention. S2PD outperforms bidirectional baselines on rule-following and surpasses other serial methods on temporal stability and sampling efficiency. The work was funded in part by Toyota Motor Europe and used computing resources from the Isambard-AI National AI Research Resource.