Light-Syn
收藏资源简介:
Light-Syn是由北京人工智能研究院等机构联合构建的合成数据集,专为视频重光照与相机轨迹控制任务设计。该数据集通过逆向映射的降级流程,从真实世界单目视频中生成包含静态、动态及AI生成场景的多样化训练样本,涵盖几何与光照解耦的配对数据。其核心创新在于通过几何一致性投影,将原始视频的几何信息与合成光照线索映射至降级视图,解决了多视角多光照真实数据稀缺的难题。该数据集主要应用于生成式视频渲染领域,支持联合相机-光照控制模型的训练,旨在突破现有方法在光照保真度与时序一致性之间的权衡瓶颈。
Light-Syn is a synthetic dataset jointly constructed by institutions including the Beijing Academy of Artificial Intelligence, specifically tailored for video relighting and camera trajectory control tasks. This dataset generates diverse training samples encompassing static, dynamic and AI-generated scenes from real-world monocular videos via a reverse-mapping downgrading pipeline, covering paired data with decoupled geometry and illumination. Its core innovation lies in mapping the geometric information of the original video and synthetic illumination cues to the downgraded views through geometric-consistent projection, thereby resolving the scarcity issue of real-world multi-view and multi-illumination data. This dataset is primarily applied in the field of generative video rendering, enabling the training of joint camera-and-illumination control models, and aims to break through the trade-off bottleneck between illumination fidelity and temporal consistency in existing approaches.

- 1Light-X: Generative 4D Video Rendering with Camera and Illumination Control新加坡南洋理工大学S实验室, 北京人工智能研究院, 华中科技大学, 清华大学智能产业研究院, 香港中文大学(深圳)未来网络研究院, 上海交通大学, 宁波诺丁汉大学 · 2025年



