⚡Delivering High Quality GenAI Video

Traditional video infrastructure was built around a "one-to-many" paradigm: a camera captures reality, a fixed encoder compresses it, and a CDN fans it out to passive viewers. Generative-AI video flips every assumption in that pipeline. Content is synthesized rather than filmed. Videos are shifting from being created once for millions of viewers to being generated on-demand and uniquely per user. Creators and consumers blur into the same role, and distribution increasingly happens at the edge.  

This talk dissects three fault-lines where today's natural-video stack breaks under generative workloads, and discusses potential solutions:  Creation → Encoding: Diffusion models produce frames whose temporal coherence is algorithmic, not physical. Standard codecs waste bits preserving pseudo-motion that could instead be regenerated from seeds or latent tokens. Encoding → Delivery: Generative streams are many-to-many: millions of personalized outputs, each short-lived. That breaks GOP-based CDN caching and overwhelms origin bandwidth.  Delivery → Experience: Interactivity expectations (prompt-to-pixel < 200 ms) demand tight co-design of model partitioning, adaptive bitrate in latent space, and real-time provenance watermarks.   

Attendees will leave with a mental framework for re-architecting encoding and delivery around the distinctive physics of generative video.