namin72/Drag-to-Live-Dataset

该数据集包含经过处理的云层延时视频及其通过CoTracker提取的运动轨迹，专为在边缘设备上训练潜在轨迹引导模型（如Wan-Move适配）而设计。数据集结构包括调整大小（256x256）并剪辑的视频片段（8-16帧）、包含CoTracker提取的密集点轨迹的.npz文件，以及链接视频与其对应轨迹文件的元数据index.jsonl。原始视频来源于Pexels（免费使用许可），并通过CoTracker进行了调整大小、剪辑和运动轨迹标注处理。此数据集用于Gods Finger (Drag-to-Live)项目，训练一个用于运动控制的LoRA适配器。

This dataset contains processed time-lapse videos of clouds and their motion trajectories extracted using CoTracker. It is designed for training Latent Trajectory Guidance models (like Wan-Move adaptation) on Edge Devices. The dataset structure includes resized (256x256) and clipped video segments (8-16 frames), .npz files containing dense point trajectories extracted by CoTracker, and index.jsonl metadata linking videos to their corresponding track files. Original videos are sourced from Pexels (Free to use license) and processed by resizing, clipping, and annotating with motion tracks using CoTracker. This dataset is used for the Gods Finger (Drag-to-Live) project to train a LoRA adapter for motion control.