Demonstration of AI-based fancam production for the Kohaku Uta Gassen using 8K cameras and VVERTIGO post-production pipeline

Abstract

KBS and NHK jointly validated an AI-assisted production workflow that extracts multiple HD fan-cams from a single 8K source, replacing multi-rig setups. VVERTIGO detects performers, links them with a graph-based clustering algorithm retrained on the 4 M-image KBS6700 face corpus, cutting identity error by 0.9 pp and yielding 15% more valid trajectories than a VGGFace2 baseline. Four RTX 4000 Ada GPUs export 8K footage three times faster than leading NLEs, while 40Gb/s IP-over-Thunderbolt networking enables simultaneous playback, AI analysis, and rendering across Mac workstations. Camera trials with Canon R5C and Sony UHC-8300 confirmed full pipeline compatibility; the Sony chain powered the live 75th Kohaku Uta Gassen, where three editors—trained in four hours—delivered performer-centric videos 8–10 min after each act and completed all edits before sign-off. The study demonstrates that data-centric AI and commodity 8K acquisition can meet prime-time deadlines while reducing on-site crews and reclaiming audience seating.