CausalCine: Real-Time Autoregressive Multi-Shot Video Generation with Content-Aware Memory Routing

Forum topic · 小凯 · 2026-05-14

Summary

CausalCine (arXiv:2605.12496) is an interactive autoregressive framework that transforms multi-shot video generation into an online directing process, addressing the motion stagnation and semantic drift that plague existing autoregressive models during long rollouts. Trained mostly for short-horizon continuation, prior models treat long sequences as extended single shots, whereas cinematic storytelling requires evolving events, viewpoint shifts, and discrete shot boundaries. CausalCine generates causally across shot changes, accepts dynamic prompts on the fly, and reuses context without regenerating previous shots. The method trains a causal foundation model to learn complex shot-transition priors, introduces Content-Aware Memory Routing (CAMR) that dynamically retrieves historical KV entries by attention relevance scores rather than temporal proximity to maintain cross-shot coherence under limited active memory, and distills the foundation model into a few-step generator enabling real-time interactive generation. Experiments show CausalCine significantly outperforms autoregressive baselines and approaches bidirectional model quality while retaining streaming interactivity.

Paper Overview

Field: Computer Vision arXiv: 2605.12496

Background

Autoregressive video generation aims at real-time, open-ended synthesis. However, 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 here: trained primarily for short-horizon continuation, they treat long sequences as extended single shots and inevitably suffer from motion stagnation and semantic drift during long rollouts.

CausalCine

CausalCine is an interactive autoregressive framework that transforms multi-shot video generation into an online directing process. It:

Generates causally across shot changes
Accepts dynamic prompts on the fly
Reuses context without regenerating previous shots

Method

1. Causal foundation model: first trained to learn 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: the causal foundation model is distilled into a few-step generator for real-time interactive generation.

Results

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

--- *Collected automatically on 2026-05-14.*