Datasets:
metadata
license: mit
task_categories:
- video-generation
- image-to-video
language:
- en
tags:
- clouds
- time-lapse
- trajectory-control
- cotracker
size_categories:
- n<1K
Drag-to-Live Dataset ☁️
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.
Dataset Structure
- videos/: Resized (256x256) and clipped video segments (8-16 frames).
- tracks/:
.npzfiles containing dense point trajectories extracted by CoTracker. - index.jsonl: Metadata linking videos to their corresponding track files.
Source
- Original Videos: Sourced from Pexels (Free to use license).
- Processing: Resized, clipped, and annotated with motion tracks using CoTracker.
Usage
This dataset is used for the God's Finger (Drag-to-Live) project to train a LoRA adapter for motion control.