The User (has) Generated Content... so what now?
summary
Handles thousands of short UGC videos by adding a heavier analysis step, a normalization stage that fixes PCM audio, broken PTS, empty frames and HDR→SDR, and a manifest verification pass that fetches playlists and segments to catch playback-breaking files. Uses ffmpeg, exiftool and libzimg for conversions, measures quality with VMAF, and runs ~1,000 trial encodes per asset to tune per-title encoding (preferring VP9 now, planning AV1). Reports wild inputs like 0.1 → 880,000 fps spikes.
AI-written summary · about
chapters (15)
- 0:00Introduction to UGC challenges
- 1:49Silent video mystery
- 3:58Jumper video timing issue
- 6:27Single frame video loop
- 7:32Empty frame video artifacts
- 8:35Color and HDR to SDR issues
- 9:17Broken reference frames
- 9:38Variable frame rate problems
- 10:00Extreme metadata anomalies
- 11:24Pipeline changes for UGC
- 14:18Processing speed vs quality
- 16:24Per-title encoding strategy
- 18:54Optimizing beyond encoding
- 19:58Observability and error handling
- 20:40Q&A and closing remarks
related talks
27mSummarizing & analyzing videos
11mThings Developers Believe About Video Files (Proven Wrong by User Uploads)
50mDemuxed 2023 Lightning Talks
11mFree ABR Renditions for User Generated Content Platforms
16mInterview
10mAdventures with VP9
31mAV1 Update
23mLessons learned from streaming the World Cup in UHD with HDR