World-R1纯文本数据集
收藏资源简介:
World-R1纯文本数据集由微软研究院与浙江大学联合构建,专为增强视频生成模型的3D一致性而设计。该数据集包含多类别、多层级的相机运动控制文本指令,旨在通过强化学习优化模型对几何约束的隐式理解。数据通过合成生成,避免了真实3D数据的依赖,并采用去耦训练策略平衡刚体几何与动态场景的灵活性。其核心应用于文本到视频生成领域,解决现有模型在长序列和大视角运动中的几何失真问题,推动视频生成向可扩展的世界模拟演进。
The World-R1 plain-text dataset was jointly developed by Microsoft Research and Zhejiang University, specifically designed to improve the 3D consistency of video generation models. This dataset contains multi-category and multi-level text instructions for camera motion control, aiming to optimize the model's implicit comprehension of geometric constraints through reinforcement learning. All data in this dataset is synthetically generated, eliminating reliance on real 3D data, and a decoupled training strategy is employed to balance the flexibility between rigid-body geometry and dynamic scenes. Its core application lies in the text-to-video generation domain, where it addresses the geometric distortion issues of existing models during long-sequence and wide-angle camera motion, thereby advancing the development of video generation toward scalable world simulation.




