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

TL;DR

CausalCine is a novel autoregressive framework for real-time, multi-shot video narrative generation that maintains coherence across shots and supports interactive editing.

Contribution

It introduces a causal autoregressive model with Content-Aware Memory Routing and distillation for real-time, multi-shot video synthesis with dynamic prompts.

Findings

Outperforms autoregressive baselines in coherence and quality.

Approaches bidirectional models in capability.

Enables streaming, interactive video generation.

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, we 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. To achieve this, we first train a causal base model on native multi-shot sequences to learn complex shot transitions prior to acceleration. We then propose Content-Aware Memory Routing (CAMR), which dynamically retrieves historical KV entries according to attention-based relevance scores rather than temporal proximity, preserving cross-shot coherence under bounded active memory. Finally, we distill the causal base model into a few-step generator for real-time interactive generation. Extensive experiments demonstrate that CausalCine significantly outperforms autoregressive baselines and approaches the capability of bidirectional models while unlocking the streaming interactivity of causal generation. Demo available at https://yihao-meng.github.io/CausalCine/