CausalCine enables real-time multi-shot video consistency
CausalCine is a research framework from HKUST and Ant Group that generates multi-shot video narratives in real-time while maintaining character and scene consistency. It uses Content-Aware Memory Routing to prevent semantic drift across cinematic shot boundaries.
CausalCine's approach to "online directing" marks a shift from passive video generation to interactive narrative control.
- –Content-Aware Memory Routing (CAMR) dynamically retrieves visual context based on attention-based relevance rather than temporal proximity, ensuring cross-shot coherence.
- –The framework achieves 16 FPS on H200 GPUs through model distillation, enabling real-time steering with dynamic prompts during the generation process.
- –Training on native multi-shot data allows the system to learn complex camera transitions and viewpoint shifts that single-shot models typically fail to maintain.
- –The ability to prompt "on the fly" for the next shot while the current one is still streaming opens new possibilities for interactive AI storytelling and game design.
DISCOVERED
1h ago
2026-05-17
PUBLISHED
1h ago
2026-05-17
RELEVANCE
AUTHOR
AI Search