Runway Characters serves 57B video transformer at 24fps

Researchers describe the system behind Runway Characters, which serves a 57-billion-parameter autoregressive video diffusion transformer for real-time conversation. Operating at 704 by 1088 resolution and 24 frames per second on four H100 GPUs, the pipeline utilizes optimizations like a ring cache and FP8 quantization to achieve sustained real-time performance. Open article →