CineWeaver: Training-Free Reference-Controllable Multi-Shot Long Video Generation for Cinematic Storytelling

Cinematic storytelling requires generating long-form videos that feature multiple distinct shots, consistent character identities, and precise control over scenes. Existing text-to-video diffusion models are typically optimized for single-shot generation and struggle to handle the structural requirements of multi-shot narratives without extensive retraining or specialized architectural modifications. This paper asks whether it is possible to achieve these cinematic capabilities using a unified, training-free framework.

The authors identify that the primary obstacle to multi-shot generation is the structural bias toward temporal continuity inherent in pretrained diffusion models. To overcome this, they propose CineWeaver, which modifies the inference process rather than the model weights. The framework employs three main strategies:

1. Shot-aware RoPE and Masked Self-Attention: By introducing "gap frames" and applying specific attention masks, the model disrupts the default temporal coupling between frames, allowing for distinct shot boundaries.
2. Shot-wise Reference Routing: Instead of applying a global reference image to the entire video, the framework routes specific reference tokens to designated shots, enabling per-shot control over characters and objects.
3. Anchor Memory Mechanism: To ensure long-term consistency across extended sequences, the authors implement an anchor memory that caches global appearance cues, allowing the model to maintain visual coherence even as the narrative progresses through multiple shots.

CineWeaver demonstrates that complex cinematic video generation can be achieved without the need for retraining or fine-tuning. By decoupling shot composition from the model's learned priors, the framework allows users to generate long-form, multi-shot videos with clear transitions and consistent identities. This approach provides a flexible, unified solution for film pre-visualization and digital storytelling, effectively bridging the gap between short-form video synthesis and professional narrative requirements.