S2PD: Serial-to-Parallel Diffusion for Physically and Logically Consistent Video Generation
Jeffrey Hu, Daniel Olmeda Reino, Ayush Tewari
University of Cambridge · Toyota Motor Europe
Published October 5, 2026 · on the feed October 7, 2026 · ▲ 2 · arXiv 2610.06847
The abstract
Bidirectional video diffusion models denoise entire videos in parallel, yet when trained on effectively unlimited in-distribution data from procedural generators, continue to violate physical laws and simple symbolic rules. We introduce Serial-to-Parallel Diffusion (S2PD), which performs autoregressive diffusion at high noise before switching to parallel diffusion at low noise. The autoregressive phase provides the serial computation needed to coordinate interdependent events and produce valid state transitions while the parallel phase jointly refines the entire video and reduces sampling time relative to fully serial generation. We implement S2PD with two architectures: a pixel-space diffusion transformer trained from scratch and a pretrained video model adapted through LoRA fine-tuning with causal attention. Across games, physical simulations, and real video, S2PD follows rules more reliably than matched bidirectional baselines and generates videos with greater temporal stability and sampling efficiency than other serial methods.
The Brief
2:01A short podcast summary of the paper. Written and voiced by AI.
Transcript
S2PD. Serial-to-Parallel Diffusion for Physically and Logically Consistent Video Generation. By Jeffrey Hu and two others, from the University of Cambridge and Toyota Motor Europe. Published on the fifth of October. Its code repository is worked out: about one hundred and seventy source files, with a readme and a licence.
Bidirectional video diffusion models denoise a whole video in parallel. Train them on effectively unlimited in-distribution data from procedural generators, and they still violate physical laws and simple symbolic rules. More data of exactly the right kind does not fix it. That points away from the data and toward the way generation is organised. When every frame is denoised at the same time, nothing in the process works through events in the order they happen.
Serial-to-Parallel Diffusion, or S2PD, splits generation by noise level. At high noise it runs autoregressive diffusion. At low noise it switches to parallel diffusion. The serial phase supplies what the authors argue is missing: the serial computation needed to coordinate interdependent events and produce valid transitions from one state to the next.
Once that structure is in place, the parallel phase refines the entire video jointly, and it cuts sampling time compared with generating everything serially. S2PD is implemented on two architectures. One is a pixel-space diffusion transformer trained from scratch. The other is a pretrained video model, adapted through LoRA fine-tuning with causal attention.
The tests span games, physical simulations, and real video. Against matched bidirectional baselines, S2PD follows rules more reliably. Against other serial methods, it produces videos with greater temporal stability and better sampling efficiency. Each comparison targets one side of the hybrid: the rules are what the serial phase was added for, and the speed is what the parallel phase was kept for.
The design assigns each mode of computation to the stage that needs it. Getting the events right is treated as sequential work, done while the video is still coarse. Polishing is treated as parallel work, done all at once at the end.
More with an account
Listen to every day’s papers
Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.
Join free