AI Watchdog: HD-VG-130M

Editor’s note: This search tool is part of The Atlantic’s investigation into how YouTube videos are taken to train AI tools. This work is part of AI Watchdog, The Atlantic’s ongoing investigation into the generative-AI industry.

---

HD-VG-130M is a collection of 1,549,408 YouTube videos, according to our analysis of the data set. It was compiled by Microsoft, and published in 2023 for the purpose of training video-generating AI models. The data set’s compilers sought out high-definition videos with a high “aesthetic score,” as judged by an AI model. They rejected videos that contained a creator’s name or watermark to prevent text from accidentally appearing in videos generated by AI models. The data set has been used by Nvidia, and its developers claim that “🎉 Up to January 2024, our data set has been downloaded by more than 50 universities and research institutes!”