Improved Cinematic-Guided Camera Language Transfer in 3D Scene

Yuan Wang, Zhuoling Jiang, Bailin Deng, Yipeng Qin

Keywords: camera language transfer, cinematic feature, camera language loss

Abstract: Directors and cinematographers often recreate iconic scenes by replicating the underlying camera language to evoke shared aesthetic and narrative meaning. In this work, we refer to this as the task of Cinematic-Guided Camera Language Transfer, where the goal is to reproduce the cinematic camera language of a reference video clip in a new 3D scene. The pioneer work, Jaws~\cite{wang2023jaws}, tackles this problem by adapting generic computer vision methods but fails to model the essential principles of cinematography, often leading to inaccurate framing, motion mismatches, and loss of expressive intent. To overcome these limitations, we systematically define the objectives of camera language transfer, grounding them in professional cinematography literature. Specifically, we conduct an in-depth review of cinematography literature to identify eight key cinematic features and encode them into five novel camera language losses. These losses not only guide optimization of camera parameters for effective transfer, but also serve as quantitative metrics for evaluating cinematographic fidelity. Extensive experiments demonstrate the superiority of our method.

Submission Number: 342