Cambridge's S2PD mixes serial and parallel diffusion for physically consistent video generation
CambUni · hf · 2026-10-07
Cambridge introduces Serial-to-Parallel Diffusion (S2PD), fixing bidirectional video diffusion models' tendency to violate physics and symbolic rules even with unlimited procedural training data. S2PD runs autoregressive diffusion at high noise (providing the serial computation to coordinate interdependent events) then switches to parallel diffusion at low noise for joint refinement and faster sampling. Implemented both as a from-scratch pixel-space DiT and via LoRA + causal attention on a pretrained video model, S2PD follows rules more reliably than matched bidirectional baselines and beats other serial methods on stability and efficiency.
Related event: Cambridge's S2PD Makes Video Diffusion Physically Consistent(3 posts)→
More from Multimodal
- Envato launches Burst mode: 6 creative image directions per credit, up to 10x faster — aakashgupta · 2026-10-08
- Envato's Burst mode turns one idea into six image directions for a single AI credit — eptwts · 2026-10-08
- Kandinsky 6.0 launches: 29B Pro and 3B Lite video models with ComfyUI support — KokaOP · 2026-10-08
- LoCoSplat ditches learned 3D networks: 4.2x faster, 6.7x less memory feed-forward 3DGS — zhenjun_zhao · 2026-10-08
- 15s MH3 video takes 90 min on a 4080 Super while others claim 8 min on laptops — redpandafire · 2026-10-08
- Nano Banana 2.1 wows users in latest community image generation tests — miilesus · 2026-10-08