✦
Summary
CausalCine is a real-time autoregressive video generation system for multi-shot narrative creation. It is designed for interactive directing, where a user can append new shot-level prompts while the system streams video causally. Instead of generating a single isolated clip, CausalCine maintains story context across shots so a sequence can evolve with continuity and new creative direction.
Key Features
Generates multi-shot video narratives using autoregressive causal streaming.
Allows new shot-level prompts to be appended without recomputing prior shots.
Uses content-aware KV memory to preserve long-range story context.
Maintains cross-shot consistency across evolving video sequences.
Supports real-time directing workflows for interactive video generation.
Retrieves relevant earlier shots based on semantic content.
Demonstrates streaming generation performance on high-end multi-GPU hardware.
Targets AI filmmaking, storyboarding, and long-form generative video research.
About CausalCine
The system reuses content-aware KV memory to preserve long-range narrative context and cross-shot consistency. This allows earlier shots to influence later ones through semantically relevant memory retrieval, while new prompts can introduce scene changes, camera directions, or story beats without recomputing the past. The project reports real-time generation behavior, including streaming generation around 16 FPS on a multi-H200 setup.
CausalCine is useful for research into long-form video generation, AI filmmaking, interactive storyboarding, and generative editing tools. Its core value is causal control: users can direct the next shot as the narrative unfolds rather than waiting for a full offline render. This makes it a strong research prototype for future cinematic agents and real-time video narrative systems.