CausalCine: Real-Time Autoregressive Generation for Multi-Shot Video Narratives

Researchers introduced CausalCine, a real-time autoregressive video framework for multi-shot narratives.

It learns native shot-to-shot transitions and uses attention-based memory routing to retrieve relevant past context.

The system is distilled into a few-step generator, making long video synthesis faster and more interactive.

CausalCine aims to reduce common failures in long-form generation, including motion stagnation and semantic drift.