Cambridge's S2PD uses serial diffusion to make video generation physically consistent
kastnerkyle · x · 2026-10-08
Researchers from Cambridge and Toyota Motor Europe introduce S2PD (Serial-to-Parallel Diffusion), addressing why bidirectional video diffusion models violate physics and simple symbolic rules even when trained on unlimited procedural data.
Method: autoregressive diffusion at high noise provides the serial computation needed for valid state transitions and interdependent events; parallel diffusion at low noise jointly refines the video, cutting sampling time versus fully serial generation. Implemented both as a pixel-space diffusion transformer trained from scratch and via LoRA + causal attention on a pretrained video model.
Results: across games, simulations, and real video, serial methods beat matched bidirectional baselines, produce multiple plausible continuations from one start frame, and offer better temporal stability and efficiency than other serial approaches.
Related event: Cambridge's S2PD Makes Video Diffusion Physically Consistent(3 posts)→
More from Multimodal
- Runway launches MCP: generate video in ChatGPT, Claude, Grok and Cursor from one chat — runwayml · 2026-10-08
- Runway brings its video generation directly into ChatGPT: brief, generate, and give notes in one chat — runwayml · 2026-10-08
- 128,000 Micro-Dramas Launched in Q1 2026, Over 95% AI-Generated, Epoch AI Reports — Jsevillamol · 2026-10-08
- LiveAvatar Drops 30 New AI Avatars With 6 Emotions and Full-Body Streaming, Free Now — alifcoder · 2026-10-08
- First MPA-approved AI film 'Gods Don't Give Gifts' heads to theaters, R-rated, submitting for Best Animated Feature Oscar — emmanuelvivier · 2026-10-08
- Free browser tool extracts exact keyframes to fix AI animation character flicker — QuietPound6205 · 2026-10-08