CausalCine

Real-time Autoregressive Generation for Multi-Shot Video Narratives

Yihao Meng*1,2 Zichen Liu*1,2 Hao Ouyang2 Qiuyu Wang2 Ka Leong Cheng2 Yue Yu1,2 Hanlin Wang1,2 Haobo Li3,2 Jiapeng Zhu2 Yanhong Zeng2 Xing Zhu2 Yujun Shen2 Qifeng Chen1 Huamin Qu1
1HKUST, 2Ant Group, 3SJTU
* Equal contribution

CausalCine enables real-time interactive autoregressive multi-shot video generation. It streams video causally, follows new shot-level prompts, and reuses content-aware KV memory to preserve long-range story context and cross-shot consistency.

Interactive Demo
Shot t-3
Shot t-2
Shot t-1
Shot t
generating...
prompt: neon alley, rain, close-up
Real-time directing 16 FPS streaming generation on 8 NVIDIA H200 GPUs.
Causal multi-shot Stable rollouts across shot boundaries with new content.
Content-aware memory Retrieve relevant earlier shots by semantic content.
Prompt anytime Append new directions without recomputing past shots.

Video Gallery

Comparison