A frame deadline changes the serving problem
Runway has published an engineering account of the system behind its live Characters product. The research report describes a roughly 57-billion-parameter video diffusion transformer generating four-frame chunks at 24 frames per second.
Each chunk has a 166.7-millisecond deadline. Missing it can stall both the face and the synchronized voice, making sustained response timing more important than a fast average result on a short batch job.
Where the reported speedup came from
Runway says its launch software needed 240.9 milliseconds per diffusion step on the four-H100 configuration. Cache changes and other optimizations reduced that to 159.3 milliseconds; an additional fused communication kernel reached 147.4 milliseconds.
A major change was to apply repeated cache transformations when new frames arrive, rather than reprocessing mostly unchanged history on every attention call. The report also describes overlapping decoding with generation and reducing communication overhead. These are system-level results for the reported implementation, not a promise that any 57-billion-parameter video model will achieve the same timing.
Production failures are part of the evidence
The account includes a rollback after a software upgrade exposed growing kernel-buffer allocations. Fixed-prompt benchmarks had missed the variability encountered in real sessions. Runway changed the fused-kernel path and expanded testing before re-enabling it.
The report also separates video generation from full conversational delay: the measured median response across 51 turns was 2.13 seconds, including the voice agent and delivery. Franklin’s analysis is that a real-time frame rate and an instant conversational response are different claims. Teams evaluating live avatars should measure both sustained playback and end-to-end turn latency under varied prompts and long sessions.