Tiger200K: Manually Curated High Visual Quality Video Dataset from UGC Platform

TL;DR

This paper addresses the shortage of high-visual-quality open data for text-to-video models by introducing Tiger200K, a manually curated dataset sourced from UGC platforms to emphasize visual fidelity. It presents a practical data-construction pipeline combining manual curation, TransNetV2-based scene segmentation, safe-zone cropping via OCR and border analysis, motion filtering, and bilingual captions generated by a visual LLM. The authors report 85k scene segments and 170k video clips from 4151 videos, with substantial portions at 4K+ resolution and tight quality controls, including safe-zone retention and bilingual captioning. The work aims to enable more effective post-training and quality-tuning of video generation models and envisions ongoing expansion and open-source releases to accelerate research in open video generation.

Abstract

The recent surge in open-source text-to-video generation models has significantly energized the research community, yet their dependence on proprietary training datasets remains a key constraint. While existing open datasets like Koala-36M employ algorithmic filtering of web-scraped videos from early platforms, they still lack the quality required for fine-tuning advanced video generation models. We present Tiger200K, a manually curated high visual quality video dataset sourced from User-Generated Content (UGC) platforms. By prioritizing visual fidelity and aesthetic quality, Tiger200K underscores the critical role of human expertise in data curation, and providing high-quality, temporally consistent video-text pairs for fine-tuning and optimizing video generation architectures through a simple but effective pipeline including shot boundary detection, OCR, border detecting, motion filter and fine bilingual caption. The dataset will undergo ongoing expansion and be released as an open-source initiative to advance research and applications in video generative models. Project page: https://tinytigerpan.github.io/tiger200k/

Figures (7)

• Figure 1: Visualization of randomly sampled clips in Tiger200k dataset. These clips demonstrate enhanced visual and aesthetic quality thanks to professional human curation in the data selection and review process and quality filtering data pipeline.
• Figure 2: The pipeline of data construction. First, the selected video will segment by scene and further subdivided into cuts of fix length. Then, methods such as OCR, border detection, and optical flow estimation are used for quality filtering. Finally, manual review was carried out and bilingual fine-grained caption was performed using VLM.
• Figure 3: The inheritance relationship among Koala-36m and other datasets. Most are collected from the early Youtube content. The quantities in the figure are relative.
• Figure 4: The content differences between the video frames of dissolve transition are relatively small, resulting in the failure of the video scene boundary detecting algorithm based on statistical differences.
• Figure 5: Comparison of the shot detecting results of PySceneDetect with different parameters and TransNetv2 on synthetic test videos (variety of transition methods) and real-world UGC videos.