S2PD: Serial-to-Parallel Diffusion for Physically and Logically Consistent Video Generation
S2PD splits video diffusion into a serial planning phase and a parallel refinement phase to reduce rule-breaking in generated clips.
The paper argues that fully parallel denoising still produces physical and logical violations, even with unlimited procedural training data. S2PD starts autoregressively at high noise so events can be coordinated in sequence, then switches to parallel diffusion at low noise for faster whole-video refinement. The authors test it with a pixel-space diffusion transformer and a LoRA-adapted pretrained video model. They report stronger rule-following than bidirectional baselines, with better temporal stability and sampling efficiency than fully serial methods. HF Daily Papers' note
The paper argues that fully parallel denoising still produces physical and logical violations, even with unlimited procedural training data. S2PD starts autoregressively at high noise so events can be coordinated in sequence, then switches to parallel diffusion at low noise for faster whole-video refinement. The authors test it with a pixel-space diffusion transformer and a LoRA-adapted pretrained video model. They report stronger rule-following than bidirectional baselines, with better temporal stability and sampling efficiency than fully serial methods. HF Daily Papers' note
score 5