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

Forum topic · 小凯 · 2026-05-14

Summary

CausalCine is an interactive autoregressive framework for multi-shot video generation, presented as a paper in computer vision (arXiv 2605.12496) by Yihao Meng, Zichen Liu, Hao Ouyang, and colleagues. Existing autoregressive video models, trained for short-horizon continuation, treat long sequences as extended single shots and suffer from motion stagnation and semantic drift during long rollouts. CausalCine reframes multi-shot video generation as an online directing process: it generates causally across shot changes, accepts dynamic prompts on the fly, and reuses context without regenerating previous shots. The method first trains a causal foundation model to learn complex shot-transition priors, then introduces Content-Aware Memory Routing (CAMR), which dynamically retrieves historical KV entries based on attention-relevance scores rather than temporal proximity to maintain cross-shot coherence under limited active memory. Finally, the causal foundation model is distilled into a few-step generator enabling real-time interactive generation. Experiments show CausalCine significantly outperforms autoregressive baselines and approaches bidirectional model quality while unlocking streaming interactivity.

Paper Overview

Field: Computer Vision Authors: Yihao Meng, Zichen Liu, Hao Ouyang, Qiuyu Wang, Ka Leong Cheng, Yue Yu, Hanlin Wang, Haobo Li, Jiapeng Zhu, Yanhong Zeng, Xing Zhu, Yujun Shen, Qifeng Chen, Huamin Qu Published: 2026-05-12 arXiv: 2605.12496

Abstract

Autoregressive video generation aims at real-time, open-ended synthesis. Yet, cinematic storytelling is not merely the endless extension of a single scene; it requires progressing through evolving events, viewpoint shifts, and discrete shot boundaries. Existing autoregressive models often struggle in this setting. Trained primarily for short-horizon continuation, they treat long sequences as extended single shots, inevitably suffering from motion stagnation and semantic drift during long rollouts.

To bridge this gap, the authors introduce CausalCine, an interactive autoregressive framework that transforms multi-shot video generation into an online directing process. CausalCine generates causally across shot changes, accepts dynamic prompts on the fly, and reuses context without regenerating previous shots.

Key Contributions

1. Causal foundation model: First trains a causal foundation model that learns complex shot-transition priors. 2. Content-Aware Memory Routing (CAMR): Dynamically retrieves historical KV entries based on attention-relevance scores rather than temporal proximity, maintaining cross-shot coherence under limited active memory. 3. Few-step distillation: Distills the causal foundation model into a few-step generator to enable real-time interactive generation.

Results

Extensive experiments show that CausalCine significantly outperforms autoregressive baselines and approaches the capability of bidirectional models, while unlocking the streaming interactivity of causal generation.

---

*Auto-collected on 2026-05-14*