Real-Time Video at Batch Size One | Runway Research

Runway has published a research post describing how it achieves real-time video generation with its 57-billion-parameter model, the backbone of its Runway Characters product. The challenge is a demanding one: generating video at 24 fps for a live, back-and-forth conversational experience leaves almost no room for latency, and doing so with a model of this scale requires careful attention to how computation is distributed and scheduled.

The headline result is an infrastructure improvement that cuts GPU requirements in half. The same workload that previously required eight H100 GPUs now runs on four - a meaningful reduction in cost and resource consumption for a production system serving real users. Achieving this kind of efficiency at batch size one, meaning each request is processed individually rather than grouped with others to amortize overhead, is notably difficult because many standard throughput optimizations do not apply in the same way.

The post is part of a broader pattern of AI labs sharing more of their serving and inference engineering work publicly. While model architecture research tends to get more attention, the systems work required to bring large models into low-latency products is equally consequential. For video generation specifically, real-time performance is a qualitatively different target than the asynchronous generation most users encounter, where a prompt might take several seconds or longer to return a result.

Runway Characters, the product built on this model, is designed around interactive video - generating a responsive, animated character in a live session rather than producing a finished video clip after the fact. That use case places strict constraints on the entire stack, from model architecture choices down to how memory is managed across GPU nodes. The details Runway shares here are likely to be of practical interest to other teams working on latency-sensitive generative video deployments.