Uploaded July 2026 | Updated November 2026, 5 days ago
VB Transform 2026 Highlights
Distillation decoded: Inside Runway's real-time AI video generation architecture and optimization strategy.
The industry benchmark for video generation has shifted from what can you generate to how fast you can render it. Runway is at the forefront of this shift, moving beyond pre-rendered content to Runway Characters - an autoregressive system that generates video live, frame-by-frame, as the conversation happens. This creates an extreme inference constraint: every user session is a real-time, high-velocity job where too much latency kills a product.
In this deep-dive session, Ryan Philips, head of enterprise product at Runway, shares the inference optimization strategies required to power real-time visual dynamics at scale.
Attendees will learn:
How distillation approaches like adversarial post-training (APT) and distribution matching distillation (DMD) determine your latency floor, output quality ceiling and training cost;
Critical architectural decisions necessary to shift from batch-processed video to live, frame-by-frame agentic generation; and
Practical capacity planning and GPU utilization strategies for managing real-time inference jobs at enterprise scale.
VB Transform 2026 Highlights
Distillation decoded: Inside Runway's real-time AI video generation architecture and optimization strategy.
The industry benchmark for video generation has shifted from what can you generate to how fast you can render it. Runway is at the forefront of this shift, moving beyond pre-rendered content to Runway Characters - an autoregressive system that generates video live, frame-by-frame, as the conversation happens. This creates an extreme inference constraint: every user session is a real-time, high-velocity job where too much latency kills a product.
In this deep-dive session, Ryan Philips, head of enterprise product at Runway, shares the inference optimization strategies required to power real-time visual dynamics at scale.
Attendees will learn:
How distillation approaches like adversarial post-training (APT) and distribution matching distillation (DMD) determine your latency floor, output quality ceiling and training cost;
Critical architectural decisions necessary to shift from batch-processed video to live, frame-by-frame agentic generation; and
Practical capacity planning and GPU utilization strategies for managing real-time inference jobs at enterprise scale.










