迈向叙事动画：人体与相机运动的联合合成
Towards Storytelling Animations: Joint Synthesis of Human and Camera Motions
把相机与角色视为同等实体、在扩散过程中显式建模两两交互的思路，可直接迁移到需要联合控制相机与动态主体的视频生成或虚拟摄影任务中；其真实+合成混合数据训练策略也值得借鉴。
原文依据 we propose a novel diffusion-based generative model to jointly synthesize character and camera motions in 3D space
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论要有效地讲述一个故事，3D 动画往往需要精心规划角色与相机在三维场景中的行为，其中相机的摆放与运动决定了角色在屏幕上的呈现方式。因此，创作叙事动画颇具挑战。尽管角色运动合成与虚拟摄影领域已取得显著进展，但以往方法只关注角色运动生成或相机运动生成，难以同时处理这两项任务。本文提出一种新颖的基于扩散的生成模型，在三维空间中联合合成角色与相机运动，以创作叙事型 3D 动画。该模型将三维场景中的各个角色与相机视为独立且同等重要的实体，并在生成过程中显式建模它们之间的两两交互。通过在真实与合成角色-相机运动数据的混合数据集上训练，该模型能够生成高质量的多角色运动，并配以富有表现力的相机运动。我们表明，该模型在人体运动生成与相机运动任务上优于现有的专用方法。
英文原文摘要 1243 字符
To tell a story effectively, a 3D animation often necessitates carefully planned behaviors of both characters and the camera in the 3D scene, where the camera placement and movement determine how the characters are displayed on screen. Thus, creating storytelling animations can be challenging. While significant progress has been made in the fields of character motion synthesis and virtual cinematography, previous methods focus on either character motion generation or camera motion generation, falling short of handling the two tasks simultaneously. In this paper, we propose a novel diffusion-based generative model to jointly synthesize character and camera motions in 3D space for creating storytelling 3D animations. Our model treats individual characters and the camera in a 3D scene as independent, equally important entities, and explicitly models pairwise interactions among them in the generation process. By being trained on a mixture dataset of real and synthetic character-camera motion data, our model is capable of generating high-quality multi-character motions coupled with compelling camera motions. We show that our model outperforms existing specialized approaches on the human motion generation and camera motion tasks.